instruction
stringlengths 13
150
| input
stringlengths 36
29.1k
| output
stringlengths 31
29.5k
| source
stringlengths 45
45
|
|---|---|---|---|
Extracting the top-k value-indices from a 1-D Tensor | Given a 1-D tensor in Torch (torch.Tensor), containing values which can be compared (say floating point), how can we extract the indices of the top-k values in that tensor?
Apart from the brute-force method, I am looking for some API call, that Torch/lua provides, which can perform this task efficiently.
| As of pull request #496 Torch now includes a built-in API named torch.topk. Example:
> t = torch.Tensor{9, 1, 8, 2, 7, 3, 6, 4, 5}
-- obtain the 3 smallest elements
> res = t:topk(3)
> print(res)
1
2
3
[torch.DoubleTensor of size 3]
-- you can also get the indices in addition
> res, ind = t:topk(3)
> print(ind)
2
4
6
[torch.LongTensor of size 3]
-- alternatively you can obtain the k largest elements as follow
-- (see the API documentation for more details)
> res = t:topk(3, true)
> print(res)
9
8
7
[torch.DoubleTensor of size 3]
At the time of writing the CPU implementation follows a sort and narrow approach (there are plans to improve it in the future). That being said an optimized GPU implementation for cutorch is currently being reviewed.
| https://stackoverflow.com/questions/34750268/ |
How to Display Custom Images in Tensorboard (e.g. Matplotlib Plots)? | The Image Dashboard section of the Tensorboard ReadMe says:
Since the image dashboard supports arbitrary pngs, you can use this to embed custom visualizations (e.g. matplotlib scatterplots) into TensorBoard.
I see how a pyplot image could be written to file, read back in as a tensor, and then used with tf.image_summary() to write it to TensorBoard, but this statement from the readme suggests there is a more direct way. Is there? If so, is there any further documentation and/or examples of how to do this efficiently?
| It is quite easy to do if you have the image in a memory buffer. Below, I show an example, where a pyplot is saved to a buffer and then converted to a TF image representation which is then sent to an image summary.
import io
import matplotlib.pyplot as plt
import tensorflow as tf
def gen_plot():
"""Create a pyplot plot and save to buffer."""
plt.figure()
plt.plot([1, 2])
plt.title("test")
buf = io.BytesIO()
plt.savefig(buf, format='png')
buf.seek(0)
return buf
# Prepare the plot
plot_buf = gen_plot()
# Convert PNG buffer to TF image
image = tf.image.decode_png(plot_buf.getvalue(), channels=4)
# Add the batch dimension
image = tf.expand_dims(image, 0)
# Add image summary
summary_op = tf.summary.image("plot", image)
# Session
with tf.Session() as sess:
# Run
summary = sess.run(summary_op)
# Write summary
writer = tf.train.SummaryWriter('./logs')
writer.add_summary(summary)
writer.close()
This gives the following TensorBoard visualization:
| https://stackoverflow.com/questions/38543850/ |
Python wheels: cp27mu not supported | I'm trying to install pytorch (http://pytorch.org/) on Linux, and according to my machine configuration, the command I should run is:
pip install https://s3.amazonaws.com/pytorch/whl/torch-0.1.6.post17-cp27-cp27mu-linux_x86_64.whl
On one machine (Linux distribution Slackware 14.1) the installation fails with error: torch-0.1.6.post17-cp27-cp27mu-linux_x86_64.whl is not a supported wheel on this platform., while on another one (Ubuntu 15.10) it succeeds. From what I understood, the problem seems to be the cp27mu in the wheel name.
Using the command import pip; print(pip.pep425tags.get_supported()) from the Python shell, I get this from the Slackware machine:
[('cp27', 'cp27m', 'manylinux1_x86_64'), ('cp27', 'cp27m', 'linux_x86_64'), ('cp27', 'none', 'manylinux1_x86_64'), ('cp27', 'none', 'linux_x86_64'), ('py2', 'none', 'manylinux1_x86_64'), ('py2', 'none', 'linux_x86_64'), ('cp27', 'none', 'any'), ('cp2', 'none', 'any'), ('py27', 'none', 'any'), ('py2', 'none', 'any'), ('py26', 'none', 'any'), ('py25', 'none', 'any'), ('py24', 'none', 'any'), ('py23', 'none', 'any'), ('py22', 'none', 'any'), ('py21', 'none', 'any'), ('py20', 'none', 'any')]
and this from the Ubuntu machine:
[('cp27', 'cp27mu', 'manylinux1_x86_64'), ('cp27', 'cp27mu', 'linux_x86_64'), ('cp27', 'none', 'manylinux1_x86_64'), ('cp27', 'none', 'linux_x86_64'), ('py2', 'none', 'manylinux1_x86_64'), ('py2', 'none', 'linux_x86_64'), ('cp27', 'none', 'any'), ('cp2', 'none', 'any'), ('py27', 'none', 'any'), ('py2', 'none', 'any'), ('py26', 'none', 'any'), ('py25', 'none', 'any'), ('py24', 'none', 'any'), ('py23', 'none', 'any'), ('py22', 'none', 'any'), ('py21', 'none', 'any'), ('py20', 'none', 'any')]
From https://www.python.org/dev/peps/pep-0513/, it seems to me that supporting cp27m or cp27mu depends on an option passed at compile time, --enable-unicode.
Now, maybe at this point I shouldn't even be asking the question, but just to be sure, does this mean that I have to compile Python with the --enable-unicode=ucs4 on the Slackware machine in order to install that wheel?
| This is exactly that.
Recompile python under slack with --enable-unicode=ucs4 and you can then install the whl.
| https://stackoverflow.com/questions/41767005/ |
Loading Torch7 trained models (.t7) in PyTorch | I am using Torch7 library for implementing neural networks. Mostly, I rely on pre-trained models. In Lua I use torch.load function to load a model saved as torch .t7 file. I am curious about switching to PyTorch( http://pytorch.org) and I read the documents. I couldn't find any information regarding the mechanisms to load a pre-trained model. The only relevant information I was able to find is this page:http://pytorch.org/docs/torch.html
But the function torch.load described in the page seems to load a file saved with pickle. If someone has additional information on loading .t7 models in PyTorch, please share it here.
| As of PyTorch 1.0 torch.utils.serialization is completely removed. Hence no one can import models from Lua Torch into PyTorch anymore. Instead, I would suggest installing PyTorch 0.4.1 through pip in a conda environment (so that you can remove it after this) and use this repo to convert your Lua Torch model to PyTorch model, not just the torch.nn.legacy model that you cannot use for training. Then use PyTorch 1.xx to do whatever with it. You can also train your converted Lua Torch models in PyTorch this way :)
| https://stackoverflow.com/questions/41861354/ |
PyTorch: How to use DataLoaders for custom Datasets | How to make use of the torch.utils.data.Dataset and torch.utils.data.DataLoader on your own data (not just the torchvision.datasets)?
Is there a way to use the inbuilt DataLoaders which they use on TorchVisionDatasets to be used on any dataset?
| Yes, that is possible. Just create the objects by yourself, e.g.
import torch.utils.data as data_utils
train = data_utils.TensorDataset(features, targets)
train_loader = data_utils.DataLoader(train, batch_size=50, shuffle=True)
where features and targets are tensors. features has to be 2-D, i.e. a matrix where each line represents one training sample, and targets may be 1-D or 2-D, depending on whether you are trying to predict a scalar or a vector.
Hope that helps!
EDIT: response to @sarthak's question
Basically yes. If you create an object of type TensorData, then the constructor investigates whether the first dimensions of the feature tensor (which is actually called data_tensor) and the target tensor (called target_tensor) have the same length:
assert data_tensor.size(0) == target_tensor.size(0)
However, if you want to feed these data into a neural network subsequently, then you need to be careful. While convolution layers work on data like yours, (I think) all of the other types of layers expect the data to be given in matrix form. So, if you run into an issue like this, then an easy solution would be to convert your 4D-dataset (given as some kind of tensor, e.g. FloatTensor) into a matrix by using the method view. For your 5000xnxnx3 dataset, this would look like this:
2d_dataset = 4d_dataset.view(5000, -1)
(The value -1 tells PyTorch to figure out the length of the second dimension automatically.)
| https://stackoverflow.com/questions/41924453/ |
What does .view() do in PyTorch? | What does .view() do to a tensor x? What do negative values mean?
x = x.view(-1, 16 * 5 * 5)
| view() reshapes the tensor without copying memory, similar to numpy's reshape().
Given a tensor a with 16 elements:
import torch
a = torch.range(1, 16)
To reshape this tensor to make it a 4 x 4 tensor, use:
a = a.view(4, 4)
Now a will be a 4 x 4 tensor. Note that after the reshape the total number of elements need to remain the same. Reshaping the tensor a to a 3 x 5 tensor would not be appropriate.
What is the meaning of parameter -1?
If there is any situation that you don't know how many rows you want but are sure of the number of columns, then you can specify this with a -1. (Note that you can extend this to tensors with more dimensions. Only one of the axis value can be -1). This is a way of telling the library: "give me a tensor that has these many columns and you compute the appropriate number of rows that is necessary to make this happen".
This can be seen in this model definition code. After the line x = self.pool(F.relu(self.conv2(x))) in the forward function, you will have a 16 depth feature map. You have to flatten this to give it to the fully connected layer. So you tell PyTorch to reshape the tensor you obtained to have specific number of columns and tell it to decide the number of rows by itself.
| https://stackoverflow.com/questions/42479902/ |
How do I print the model summary in PyTorch? | How do I print the summary of a model in PyTorch like what model.summary() does in Keras:
Model Summary:
____________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
====================================================================================================
input_1 (InputLayer) (None, 1, 15, 27) 0
____________________________________________________________________________________________________
convolution2d_1 (Convolution2D) (None, 8, 15, 27) 872 input_1[0][0]
____________________________________________________________________________________________________
maxpooling2d_1 (MaxPooling2D) (None, 8, 7, 27) 0 convolution2d_1[0][0]
____________________________________________________________________________________________________
flatten_1 (Flatten) (None, 1512) 0 maxpooling2d_1[0][0]
____________________________________________________________________________________________________
dense_1 (Dense) (None, 1) 1513 flatten_1[0][0]
====================================================================================================
Total params: 2,385
Trainable params: 2,385
Non-trainable params: 0
| While you will not get as detailed information about the model as in Keras' model.summary, simply printing the model will give you some idea about the different layers involved and their specifications.
For instance:
from torchvision import models
model = models.vgg16()
print(model)
The output in this case would be something as follows:
VGG (
(features): Sequential (
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU (inplace)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU (inplace)
(4): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))
(5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(6): ReLU (inplace)
(7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(8): ReLU (inplace)
(9): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))
(10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(11): ReLU (inplace)
(12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(13): ReLU (inplace)
(14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(15): ReLU (inplace)
(16): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))
(17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(18): ReLU (inplace)
(19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(20): ReLU (inplace)
(21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(22): ReLU (inplace)
(23): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))
(24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(25): ReLU (inplace)
(26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(27): ReLU (inplace)
(28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(29): ReLU (inplace)
(30): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))
)
(classifier): Sequential (
(0): Dropout (p = 0.5)
(1): Linear (25088 -> 4096)
(2): ReLU (inplace)
(3): Dropout (p = 0.5)
(4): Linear (4096 -> 4096)
(5): ReLU (inplace)
(6): Linear (4096 -> 1000)
)
)
Now you could, as mentioned by Kashyap, use the state_dict method to get the weights of the different layers. But using this listing of the layers would perhaps provide more direction is creating a helper function to get that Keras like model summary! Hope this helps!
| https://stackoverflow.com/questions/42480111/ |
How do I save a trained model in PyTorch? | How do I save a trained model in PyTorch? I have read that:
torch.save()/torch.load() is for saving/loading a serializable object.
model.state_dict()/model.load_state_dict() is for saving/loading model state.
| Found this page on their github repo:
Recommended approach for saving a model
There are two main approaches for serializing and restoring a model.
The first (recommended) saves and loads only the model parameters:
torch.save(the_model.state_dict(), PATH)
Then later:
the_model = TheModelClass(*args, **kwargs)
the_model.load_state_dict(torch.load(PATH))
The second saves and loads the entire model:
torch.save(the_model, PATH)
Then later:
the_model = torch.load(PATH)
However in this case, the serialized data is bound to the specific classes and the exact directory structure used, so it can break in various ways when used in other projects, or after some serious refactors.
See also: Save and Load the Model section from the official PyTorch tutorials.
| https://stackoverflow.com/questions/42703500/ |
L1/L2 regularization in PyTorch | How do I add L1/L2 regularization in PyTorch without manually computing it?
| See the documentation. Add a weight_decay parameter to the optimizer for L2 regularization.
| https://stackoverflow.com/questions/42704283/ |
How can I install torchtext? | I have PyTorch installed in my machine but whenever I try to do the following-
from torchtext import data
from torchtext import datasets
I get the following error.
ImportError: No module named 'torchtext'
How can I install torchtext?
| The package was released with setuptools support. You can clone the repository and run python setup.py install. Unfortunately, I don't think that they have released it on pip.
| https://stackoverflow.com/questions/42711144/ |
What is the difference between view() and unsqueeze() in Torch? | Use of unsqueeze():
input = torch.Tensor(2, 4, 3) # input: 2 x 4 x 3
print(input.unsqueeze(0).size()) # prints - torch.size([1, 2, 4, 3])
Use of view():
input = torch.Tensor(2, 4, 3) # input: 2 x 4 x 3
print(input.view(1, -1, -1, -1).size()) # prints - torch.size([1, 2, 4, 3])
According to documentation, unsqueeze() inserts singleton dim at position given as parameter and view() creates a view with different dimensions of the storage associated with tensor.
What view() does is clear to me, but I am unable to distinguish it from unsqueeze(). Moreover, I don't understand when to use view() and when to use unsqueeze()?
Any help with good explanation would be appreciated!
| view() can only take a single -1 argument.
So, if you want to add a singleton dimension, you would need to provide all the dimensions as arguments. For e.g., if A is a 2x3x4 tensor, to add a singleton dimension, you would need to do A:view(2, 1, 3, 4).
However, sometimes, the dimensionality of the input is unknown when the operation is being used. Thus, we dont know that A is 2x3x4, but we would still like to insert a singleton dimension. This happens a lot when using minibatches of tensors, where the last dimension is usually unknown. In these cases, the nn.Unsqueeze is useful and lets us insert the dimension without explicitly being aware of the other dimensions when writing the code.
| https://stackoverflow.com/questions/42866654/ |
Why tensor.view() is not working in pytorch? | I have the following piece of code.
embedded = self.embedding(input).view(1, 1, -1)
embedded = self.drop(embedded)
print(embedded[0].size(), hidden[0].size())
concatenated_output = torch.cat((embedded[0], hidden[0]), 1)
The last line of the code is giving me the following error.
RuntimeError: inconsistent tensor sizes at /data/users/soumith/miniconda2/conda-bld/pytorch-0.1.9_1487344852722/work/torch/lib/THC/generic/THCTensorMath.cu:141
Please note, when I am printing the tensor shapes at line no. 3, I am getting the following output.
torch.size([1, 300]) torch.size([1, 1, 300])
Why I am getting [1, 300] shape for embedded tensor even though I have used the view method as view(1, 1, -1)?
Any help would be appreciated!
| embedded was a 3d-tensor and hidden was a tuple of two elements (hidden states and cell states) where each element is a 3d-tensor. hidden was the output from LSTM layer. In PyTorch, LSTM returns hidden states [h] and cell states [c] as a tuple which made me confused about the error.
So, I updated the last line of the code as follows and it solved the problem.
concatenated_output = torch.cat((embedded, hidden[0]), 1)
| https://stackoverflow.com/questions/42866752/ |
How to convert a list or numpy array to a 1d torch tensor? | I have a list (or, a numpy array) of float values. I want to create a 1d torch tensor that will contain all those values. I can create the torch tensor and run a loop to store the values.
But I want to know is there any way, I can create a torch tensor with initial values from a list or array? Also suggest me if there is any pythonic way to achieve this as I am working in pytorch.
| These are general operations in pytorch and available in the documentation. PyTorch allows easy interfacing with numpy. There is a method called from_numpy and the documentation is available here
import numpy as np
import torch
array = np.arange(1, 11)
tensor = torch.from_numpy(array)
| https://stackoverflow.com/questions/42894882/ |
AttributeError: module 'torch' has no attribute 'cmul' | I was trying to do element-wise multiplication of two tensors using the example provided here.
My code:
import torch
x = torch.Tensor([2, 3])
y = torch.Tensor([2, 1])
z = torch.cmul(x, y)
print(z)
It is giving me the following error.
AttributeError: module 'torch' has no attribute 'cmul'
Can anyone tell me why I am getting this error?
| I got the solution. Instead of using cmul, I need to use mul. The following code worked for me!
import torch
x = torch.Tensor([2, 3])
y = torch.Tensor([2, 1])
z = torch.mul(x, y)
print(z)
PS: I was using pytorch, not lua.
| https://stackoverflow.com/questions/42939708/ |
Performing Convolution (NOT cross-correlation) in pytorch | I have a network that I am trying to implement in pytorch, and I cannot seem to figure out how to implement "pure" convolution. In tensorflow it could be accomplished like this:
def conv2d_flipkernel(x, k, name=None):
return tf.nn.conv2d(x, flipkernel(k), name=name,
strides=(1, 1, 1, 1), padding='SAME')
With the flipkernel function being:
def flipkernel(kern):
return kern[(slice(None, None, -1),) * 2 + (slice(None), slice(None))]
How can something similar be done in pytorch?
| TLDR Use the convolution from the functional toolbox, torch.nn.fuctional.conv2d, not torch.nn.Conv2d, and flip your filter around the vertical and horizontal axis.
torch.nn.Conv2d is a convolutional layer for a network. Because weights are learned, it does not matter if it is implemented using cross-correlation, because the network will simply learn a mirrored version of the kernel (Thanks @etarion for this clarification).
torch.nn.fuctional.conv2d performs convolution with the inputs and weights provided as arguments, similar to the tensorflow function in your example. I wrote a simple test to determine whether, like the tensorflow function, it is actually performing cross-correlation and it is necessary to flip the filter for correct convolutional results.
import torch
import torch.nn.functional as F
import torch.autograd as autograd
import numpy as np
#A vertical edge detection filter.
#Because this filter is not symmetric, for correct convolution the filter must be flipped before element-wise multiplication
filters = autograd.Variable(torch.FloatTensor([[[[-1, 1]]]]))
#A test image of a square
inputs = autograd.Variable(torch.FloatTensor([[[[0,0,0,0,0,0,0], [0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0], [0, 0, 1, 1, 1, 0, 0],
[0,0,0,0,0,0,0]]]]))
print(F.conv2d(inputs, filters))
This outputs
Variable containing:
(0 ,0 ,.,.) =
0 0 0 0 0 0
0 1 0 0 -1 0
0 1 0 0 -1 0
0 1 0 0 -1 0
0 0 0 0 0 0
[torch.FloatTensor of size 1x1x5x6]
This output is the result for cross-correlation. Therefore, we need to flip the filter
def flip_tensor(t):
flipped = t.numpy().copy()
for i in range(len(filters.size())):
flipped = np.flip(flipped,i) #Reverse given tensor on dimention i
return torch.from_numpy(flipped.copy())
print(F.conv2d(inputs, autograd.Variable(flip_tensor(filters.data))))
The new output is the correct result for convolution.
Variable containing:
(0 ,0 ,.,.) =
0 0 0 0 0 0
0 -1 0 0 1 0
0 -1 0 0 1 0
0 -1 0 0 1 0
0 0 0 0 0 0
[torch.FloatTensor of size 1x1x5x6]
| https://stackoverflow.com/questions/42970009/ |
AttributeError: cannot assign module before Module.__init__() call | I am getting the following error.
Traceback (most recent call last):
File "main.py", line 63, in <module>
question_classifier = QuestionClassifier(corpus.dictionary, embeddings_index, corpus.max_sent_length, args)
File "/net/if5/wua4nw/wasi/academic/research_with_prof_chang/projects/question_answering/duplicate_question_detection/source/question_classifier.py", line 26, in __init__
self.embedding = EmbeddingLayer(len(dictionary), args.emsize, args.dropout)
File "/if5/wua4nw/anaconda3/lib/python3.5/site-packages/torch/nn/modules/module.py", line 255, in __setattr__
"cannot assign module before Module.__init__() call")
AttributeError: cannot assign module before Module.__init__() call
I have a class as follows.
class QuestionClassifier(nn.Module):
def __init__(self, dictionary, embeddings_index, max_seq_length, args):
self.embedding = EmbeddingLayer(len(dictionary), args.emsize, args.dropout)
self.encoder = EncoderRNN(args.emsize, args.nhid, args.model, args.bidirection, args.nlayers, args.dropout)
self.drop = nn.Dropout(args.dropout)
So, when I run the following line:
question_classifier = QuestionClassifier(corpus.dictionary, embeddings_index, corpus.max_sent_length, args)
I get the above mentioned error. Here, EmbeddingLayer and EncoderRNN is a class written by me which inherits nn.module like the QuestionClassifier class.
What I am doing wrong here?
| Looking at the pytorch source code for Module, we see in the docstring an example of deriving from Module includes:
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
self.conv1 = nn.Conv2d(1, 20, 5)
self.conv2 = nn.Conv2d(20, 20, 5)
So you probably want to call Module's init the same way in your derived class:
super(QuestionClassifier, self).__init__()
| https://stackoverflow.com/questions/43080583/ |
How to count the amount of layers in a CNN? | The Pytorch implementation of ResNet-18.
has the following structure, which appears to be 54 layers, not 18.
So why is it called "18"? How many layers does it actually have?
ResNet (
(conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
(relu): ReLU (inplace)
(maxpool): MaxPool2d (size=(3, 3), stride=(2, 2), padding=(1, 1), dilation=(1, 1))
(layer1): Sequential (
(0): BasicBlock (
(conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
(relu): ReLU (inplace)
(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
)
(1): BasicBlock (
(conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
(relu): ReLU (inplace)
(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
)
)
(layer2): Sequential (
(0): BasicBlock (
(conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True)
(relu): ReLU (inplace)
(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True)
(downsample): Sequential (
(0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True)
)
)
(1): BasicBlock (
(conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True)
(relu): ReLU (inplace)
(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True)
)
)
(layer3): Sequential (
(0): BasicBlock (
(conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
(relu): ReLU (inplace)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
(downsample): Sequential (
(0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
)
)
(1): BasicBlock (
(conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
(relu): ReLU (inplace)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
)
)
(layer4): Sequential (
(0): BasicBlock (
(conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
(relu): ReLU (inplace)
(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
(downsample): Sequential (
(0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
)
)
(1): BasicBlock (
(conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
(relu): ReLU (inplace)
(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
)
)
(avgpool): AvgPool2d (
)
(fc): Linear (512 -> 1000)
)
| From your output, we can know that there are 20 convolution layers (one 7x7 conv, 16 3x3 conv, and plus 3 1x1 conv for downsample). Basically, if you ignore the 1x1 conv, and counting the FC (linear) layer, the number of layers are 18.
And I've also made an example on how to visualize your architecture in pytorch via graphviz, hope it will help you understand your architecture.
| https://stackoverflow.com/questions/43175778/ |
IDE autocomplete for pytorch | I'm using Visual Studio Code. Recently tried out Kite. Both of which doesn't seem to have autocomplete for pytorch.
Is is possible with these tools ? If not, can someone suggest an editor that does ?
Thank you!
| Use Pycharm https://www.jetbrains.com/pycharm/
Get Community version, because it is free with debugger and autocomplete. (more than enough for student)
To get autocomplete and run/debug the code on Pycharm IDE, you have to set correct project interpreter path to your environment (or virtual environment) which you install pytorch
The screenshot shows that I set Project Interpreter path to conda environment named 'pytorch', which I install Pytorch.
| https://stackoverflow.com/questions/43287476/ |
PyTorch reshape tensor dimension | I want to reshape a vector of shape (5,) into a matrix of shape (1, 5).
With numpy, I can do:
>>> import numpy as np
>>> a = np.array([1, 2, 3, 4, 5])
>>> a.shape
(5,)
>>> a = np.reshape(a, (1, 5))
>>> a.shape
(1, 5)
>>> a
array([[1, 2, 3, 4, 5]])
But how do I do this with PyTorch?
| Use torch.unsqueeze(input, dim, out=None):
>>> import torch
>>> a = torch.Tensor([1, 2, 3, 4, 5])
>>> a
1
2
3
4
5
[torch.FloatTensor of size 5]
>>> a = a.unsqueeze(0)
>>> a
1 2 3 4 5
[torch.FloatTensor of size 1x5]
| https://stackoverflow.com/questions/43328632/ |
Pytorch, what are the gradient arguments | I am reading through the documentation of PyTorch and found an example where they write
gradients = torch.FloatTensor([0.1, 1.0, 0.0001])
y.backward(gradients)
print(x.grad)
where x was an initial variable, from which y was constructed (a 3-vector). The question is, what are the 0.1, 1.0 and 0.0001 arguments of the gradients tensor ? The documentation is not very clear on that.
|
The original code I haven't found on PyTorch website anymore.
gradients = torch.FloatTensor([0.1, 1.0, 0.0001])
y.backward(gradients)
print(x.grad)
The problem with the code above is there is no function based on how to calculate the gradients. This means we don't know how many parameters (arguments the function takes) and the dimension of parameters.
To fully understand this I created an example close to the original:
Example 1:
a = torch.tensor([1.0, 2.0, 3.0], requires_grad = True)
b = torch.tensor([3.0, 4.0, 5.0], requires_grad = True)
c = torch.tensor([6.0, 7.0, 8.0], requires_grad = True)
y=3*a + 2*b*b + torch.log(c)
gradients = torch.FloatTensor([0.1, 1.0, 0.0001])
y.backward(gradients,retain_graph=True)
print(a.grad) # tensor([3.0000e-01, 3.0000e+00, 3.0000e-04])
print(b.grad) # tensor([1.2000e+00, 1.6000e+01, 2.0000e-03])
print(c.grad) # tensor([1.6667e-02, 1.4286e-01, 1.2500e-05])
I assumed our function is y=3*a + 2*b*b + torch.log(c) and the parameters are tensors with three elements inside.
You can think of the gradients = torch.FloatTensor([0.1, 1.0, 0.0001]) like this is the accumulator.
As you may hear, PyTorch autograd system calculation is equivalent to Jacobian product.
In case you have a function, like we did:
y=3*a + 2*b*b + torch.log(c)
Jacobian would be [3, 4*b, 1/c]. However, this Jacobian is not how PyTorch is doing things to calculate the gradients at a certain point.
PyTorch uses forward pass and backward mode automatic differentiation (AD) in tandem.
There is no symbolic math involved and no numerical differentiation.
Numerical differentiation would be to calculate δy/δb, for b=1 and b=1+ε where ε is small.
If you don't use gradients in y.backward():
Example 2
a = torch.tensor(0.1, requires_grad = True)
b = torch.tensor(1.0, requires_grad = True)
c = torch.tensor(0.1, requires_grad = True)
y=3*a + 2*b*b + torch.log(c)
y.backward()
print(a.grad) # tensor(3.)
print(b.grad) # tensor(4.)
print(c.grad) # tensor(10.)
You will simply get the result at a point, based on how you set your a, b, c tensors initially.
Be careful how you initialize your a, b, c:
Example 3:
a = torch.empty(1, requires_grad = True, pin_memory=True)
b = torch.empty(1, requires_grad = True, pin_memory=True)
c = torch.empty(1, requires_grad = True, pin_memory=True)
y=3*a + 2*b*b + torch.log(c)
gradients = torch.FloatTensor([0.1, 1.0, 0.0001])
y.backward(gradients)
print(a.grad) # tensor([3.3003])
print(b.grad) # tensor([0.])
print(c.grad) # tensor([inf])
If you use torch.empty() and don't use pin_memory=True you may have different results each time.
Also, note gradients are like accumulators so zero them when needed.
Example 4:
a = torch.tensor(1.0, requires_grad = True)
b = torch.tensor(1.0, requires_grad = True)
c = torch.tensor(1.0, requires_grad = True)
y=3*a + 2*b*b + torch.log(c)
y.backward(retain_graph=True)
y.backward()
print(a.grad) # tensor(6.)
print(b.grad) # tensor(8.)
print(c.grad) # tensor(2.)
Lastly few tips on terms PyTorch uses:
PyTorch creates a dynamic computational graph when calculating the gradients in forward pass. This looks much like a tree.
So you will often hear the leaves of this tree are input tensors and the root is output tensor.
Gradients are calculated by tracing the graph from the root to the leaf and multiplying every gradient in the way using the chain rule. This multiplying occurs in the backward pass.
Back some time I created PyTorch Automatic Differentiation tutorial that you may check interesting explaining all the tiny details about AD.
| https://stackoverflow.com/questions/43451125/ |
Error when compiling pytorch: 'cstdint' file not found | I was attempting to compile pytorch using NO_CUDA=1 python setup.py install on Mac OS X, but I got these errors:
In file included from /Users/ezyang/Dev/pytorch-tmp/torch/lib/tmp_install/include/THPP/Tensor.hpp:3:
/Users/ezyang/Dev/pytorch-tmp/torch/lib/tmp_install/include/THPP/Storage.hpp:6:10: fatal error:
'cstdint' file not found
#include <cstdint>
^
1 error generated.
In file included from torch/csrc/autograd/functions/init.cpp:2:
In file included from torch/csrc/autograd/functions/batch_normalization.h:4:
In file included from /Users/ezyang/Dev/pytorch-tmp/torch/lib/tmp_install/include/THPP/THPP.h:4:
/Users/ezyang/Dev/pytorch-tmp/torch/lib/tmp_install/include/THPP/Storage.hpp:6:10: fatal error:
'cstdint' file not found
#include <cstdint>
^
1 error generated.
1 error generated.
1 error generated.
1 error generated.
In file included from torch/csrc/autograd/python_hook.cpp:5:
In file included from /Users/ezyang/Dev/pytorch-tmp/torch/csrc/THP.h:32:
/Users/ezyang/Dev/pytorch-tmp/torch/csrc/utils.h:6:10: fatal error:
'type_traits' file not found
#include <type_traits>
^
13 errors generated.
1 error generated.
1 error generated.
error: command 'gcc' failed with exit status 1
| You need to setup environment variables to have Python use the correct C compiler on OS X. You should do this instead:
NO_CUDA=1 MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install
| https://stackoverflow.com/questions/43640454/ |
How to use the BCELoss in PyTorch? | I want to write a simple autoencoder in PyTorch and use BCELoss, however, I get NaN out, since it expects the targets to be between 0 and 1. Could someone post a simple use case of BCELoss?
| Update
The BCELoss function did not use to be numerically stable. See this issue https://github.com/pytorch/pytorch/issues/751. However, this issue has been resolved with Pull #1792, so that BCELoss is numerically stable now!
Old answer
If you build PyTorch from source, you can use the numerically stable function BCEWithLogitsLoss(contributed in https://github.com/pytorch/pytorch/pull/1792), which takes logits as input.
Otherwise, you can use the following function (contributed by yzgao in the above issue):
class StableBCELoss(nn.modules.Module):
def __init__(self):
super(StableBCELoss, self).__init__()
def forward(self, input, target):
neg_abs = - input.abs()
loss = input.clamp(min=0) - input * target + (1 + neg_abs.exp()).log()
return loss.mean()
| https://stackoverflow.com/questions/43708693/ |
pytorch Network.parameters() missing 1 required positional argument: 'self' | There's a problem when I call Network.parameters() in pytorch in this line in my main function:
optimizer = optim.SGD(Network.parameters(), lr=0.001, momentum=0.9)
I get the error code:
TypeError: parameters() missing 1 required positional argument: 'self'
My network is defined in this class
class Network(nn.Module):
def __init__(self):
super(Network, self).__init__()
self.conv1 = nn.Conv2d(1, 32, 5)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(32, 64, 5)
self.pool2 = nn.MaxPool2d(2, 2)
self.conv3 = nn.Conv2d(64, 64, 5)
self.pool2 = nn.MaxPool2d(2, 2)
self.fc1 = nn.Linear(64 * 5 * 5, 512)
self.fc2 = nn.Linear(512, 640)
self.fc3 = nn.Linear(640, 3756)
def forward(self, x):
x = self.pool(F.relu(self.conv(x)))
x = self.pool(F.relu(self.conv2(x)))
x = self.pool(F.relu(self.conv3(x)))
x = x.view(-1, 64 * 5 * 5)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
Pretty sure that I imported all torch modules correctly.
Any ideas of what I'm doing wrong here?
Thank you!
| When doing Network.parameters() you are calling the static method parameters.
But, parameters is an instance method.
So you have to instansiate Network before calling parameters.
network = Network()
optimizer = optim.SGD(network.parameters(), lr=0.001, momentum=0.9)
Or, if you only needs Network first this particular line:
optimizer = optim.SGD(Network().parameters(), lr=0.001, momentum=0.9)
| https://stackoverflow.com/questions/43779500/ |
Pytorch, TypeError: object() takes no parameters | This is probably a beginner question, but nevertheless: When running an image classifier build with pytorch, I get this error:
Traceback (most recent call last):
File "/pytorch/kanji_torch.py", line 47, in <module>
network = Network()
File "/pytorch/kanji_torch.py", line 113, in __init__
self.conv1 = nn.Conv2d(1, 32, 5)
File "/python3.5/site-packages/torch/nn/modules/conv.py", line 233, in __init__
False, _pair(0), groups, bias)
File "/python3.5/site-packages/torch/nn/modules/conv.py", line 32, in __init__
out_channels, in_channels // groups, *kernel_size))
TypeError: object() takes no parameters
I define the network class like this:
class Network(torch.nn.Module):
def __init__(self):
super(Network, self).__init__()
self.conv1 = nn.Conv2d(1, 32, 5)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(32, 64, 5)
self.pool2 = nn.MaxPool2d(2, 2)
self.conv3 = nn.Conv2d(64, 64, 5)
self.pool2 = nn.MaxPool2d(2, 2)
self.fc1 = nn.Linear(64 * 5 * 5, 512)
self.fc2 = nn.Linear(512, 640)
self.fc3 = nn.Linear(640, 3756)
Pretty sure that I imported all the relevant pytorch library modules correctly.
(import torch.nn as nn and
import torch)
Any ideas of what I'm doing wrong?
Thank you!
| You might have a problem with your pytorch version, when I run the code:
class Network(torch.nn.Module):
def __init__(self):
super(Network, self).__init__()
self.conv1 = nn.Conv2d(1, 32, 5)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(32, 64, 5)
self.pool2 = nn.MaxPool2d(2, 2)
self.conv3 = nn.Conv2d(64, 64, 5)
self.pool2 = nn.MaxPool2d(2, 2)
self.fc1 = nn.Linear(64 * 5 * 5, 512)
self.fc2 = nn.Linear(512, 640)
self.fc3 = nn.Linear(640, 3756)
print(network)
The output is:
Network (
(conv1): Conv2d(1, 32, kernel_size=(5, 5), stride=(1, 1))
(pool): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))
(conv2): Conv2d(32, 64, kernel_size=(5, 5), stride=(1, 1))
(pool2): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))
(conv3): Conv2d(64, 64, kernel_size=(5, 5), stride=(1, 1))
(fc1): Linear (1600 -> 512)
(fc2): Linear (512 -> 640)
(fc3): Linear (640 -> 3756)
)
I would suggest to update/reinstall pytorch.
| https://stackoverflow.com/questions/43781526/ |
How to run PyTorch on GPU by default? | I want to run PyTorch using cuda. I set model.cuda() and torch.cuda.LongTensor() for all tensors.
Do I have to create tensors using .cuda explicitly if I have used model.cuda()?
Is there a way to make all computations run on GPU by default?
| I do not think you can specify that you want to use cuda tensors by default.
However you should have a look to the pytorch offical examples.
In the imagenet training/testing script, they use a wrapper over the model called DataParallel.
This wrapper has two advantages:
it handles the data parallelism over multiple GPUs
it handles the casting of cpu tensors to cuda tensors
As you can see in L164, you don't have to cast manually your inputs/targets to cuda.
Note that, if you have multiple GPUs and you want to use a single one, launch any python/pytorch scripts with the CUDA_VISIBLE_DEVICES prefix. For instance CUDA_VISIBLE_DEVICES=0 python main.py.
| https://stackoverflow.com/questions/43806326/ |
Fixing a subset of weights in Neural network during training | I am considering creating a customized neural network. The basic structure is the same as usual, but I want to truncate the connections between layers. For example, if I construct a network with two hidden layers, I would like to delete some weights and keep the others, like so:
This is not conventional dropout (to avoid overfitting), since the remaining weights (connections) should be specified and fixed.
Are there any ways in python to do it? Tensorflow, pytorch, theano or any other modules?
| Yes you can do this in tensorflow.
You would have some layer in your tensorflow code something like so:
m = tf.Variable( [width,height] , dtype=tf.float32 ))
b = tf.Variable( [height] , dtype=tf.float32 ))
h = tf.sigmoid( tf.matmul( x,m ) + b )
What you want is some new matrix, let's call it k for kill. It is going to kill specific neural connections. The neural connections are defined in m. This would be your new configuration
k = tf.Constant( kill_matrix , dtype=tf.float32 )
m = tf.Variable( [width,height] , dtype=tf.float32 )
b = tf.Variable( [height] , dtype=tf.float32 )
h = tf.sigmoid( tf.matmul( x, tf.multiply(m,k) ) + b )
Your kill_matrix is a matrix of 1's and 0's. Insert a 1 for every neural connection you want to keep and a 0 for every one you want to kill.
| https://stackoverflow.com/questions/43851657/ |
Accuracy score in pyTorch LSTM | I have been running this LSTM tutorial on the wikigold.conll NER data set
training_data contains a list of tuples of sequences and tags, for example:
training_data = [
("They also have a song called \" wake up \"".split(), ["O", "O", "O", "O", "O", "O", "I-MISC", "I-MISC", "I-MISC", "I-MISC"]),
("Major General John C. Scheidt Jr.".split(), ["O", "O", "I-PER", "I-PER", "I-PER"])
]
And I wrote down this function
def predict(indices):
"""Gets a list of indices of training_data, and returns a list of predicted lists of tags"""
for index in indicies:
inputs = prepare_sequence(training_data[index][0], word_to_ix)
tag_scores = model(inputs)
values, target = torch.max(tag_scores, 1)
yield target
This way I can get the predicted labels for specific indices in the training data.
However, how do I evaluate the accuracy score across all training data.
Accuracy being, the amount of words correctly classified across all sentences divided by the word count.
This is what I came up with, which is extremely slow and ugly:
y_pred = list(predict([s for s, t in training_data]))
y_true = [t for s, t in training_data]
c=0
s=0
for i in range(len(training_data)):
n = len(y_true[i])
#super ugly and ineffiicient
s+=(sum(sum(list(y_true[i].view(-1, n) == y_pred[i].view(-1, n).data))))
c+=n
print ('Training accuracy:{a}'.format(a=float(s)/c))
How can this be done efficiently in pytorch ?
P.S:
I've been trying to use sklearn's accuracy_score unsuccessfully
| I would use numpy in order to not iterate the list in pure python.
The results are the same, but it runs much faster
def accuracy_score(y_true, y_pred):
y_pred = np.concatenate(tuple(y_pred))
y_true = np.concatenate(tuple([[t for t in y] for y in y_true])).reshape(y_pred.shape)
return (y_true == y_pred).sum() / float(len(y_true))
And this is how to use it:
#original code:
y_pred = list(predict([s for s, t in training_data]))
y_true = [t for s, t in training_data]
#numpy accuracy score
print(accuracy_score(y_true, y_pred))
| https://stackoverflow.com/questions/43962599/ |
Can I slice tensors with logical indexing or lists of indices? | I'm trying to slice a PyTorch tensor using a logical index on the columns. I want the columns that correspond to a 1 value in the index vector. Both slicing and logical indexing are possible, but are they possible together? If so, how? My attempt keeps throwing the unhelpful error
TypeError: indexing a tensor with an object of type ByteTensor. The
only supported types are integers, slices, numpy scalars and
torch.LongTensor or torch.ByteTensor as the only argument.
MCVE
Desired Output
import torch
C = torch.LongTensor([[1, 3], [4, 6]])
# 1 3
# 4 6
Logical indexing on the columns only:
A_log = torch.ByteTensor([1, 0, 1]) # the logical index
B = torch.LongTensor([[1, 2, 3], [4, 5, 6]])
C = B[:, A_log] # Throws error
If the vectors are the same size, logical indexing works:
B_truncated = torch.LongTensor([1, 2, 3])
C = B_truncated[A_log]
And I can get the desired result by repeating the logical index so that it has the same size as the tensor I am indexing, but then I also have to reshape the output.
C = B[A_log.repeat(2, 1)] # [torch.LongTensor of size 4]
C = C.resize_(2, 2)
I also tried using a list of indices:
A_idx = torch.LongTensor([0, 2]) # the index vector
C = B[:, A_idx] # Throws error
If I want contiguous ranges of indices, slicing works:
C = B[:, 1:2]
| I think this is implemented as the index_select function, you can try
import torch
A_idx = torch.LongTensor([0, 2]) # the index vector
B = torch.LongTensor([[1, 2, 3], [4, 5, 6]])
C = B.index_select(1, A_idx)
# 1 3
# 4 6
| https://stackoverflow.com/questions/43989310/ |
PyTorch: How to convert pretrained FC layers in a CNN to Conv layers | I want to convert a pre-trained CNN (like VGG-16) to a fully convolutional network in Pytorch. How can I do so?
| You can do that as follows (see comments for description):
import torch
import torch.nn as nn
from torchvision import models
# 1. LOAD PRE-TRAINED VGG16
model = models.vgg16(pretrained=True)
# 2. GET CONV LAYERS
features = model.features
# 3. GET FULLY CONNECTED LAYERS
fcLayers = nn.Sequential(
# stop at last layer
*list(model.classifier.children())[:-1]
)
# 4. CONVERT FULLY CONNECTED LAYERS TO CONVOLUTIONAL LAYERS
### convert first fc layer to conv layer with 512x7x7 kernel
fc = fcLayers[0].state_dict()
in_ch = 512
out_ch = fc["weight"].size(0)
firstConv = nn.Conv2d(in_ch, out_ch, 7, 7)
### get the weights from the fc layer
firstConv.load_state_dict({"weight":fc["weight"].view(out_ch, in_ch, 7, 7),
"bias":fc["bias"]})
# CREATE A LIST OF CONVS
convList = [firstConv]
# Similarly convert the remaining linear layers to conv layers
for layer in enumerate(fcLayers[1:]):
if isinstance(module, nn.Linear):
# Convert the nn.Linear to nn.Conv
fc = module.state_dict()
in_ch = fc["weight"].size(1)
out_ch = fc["weight"].size(0)
conv = nn.Conv2d(in_ch, out_ch, 1, 1)
conv.load_state_dict({"weight":fc["weight"].view(out_ch, in_ch, 1, 1),
"bias":fc["bias"]})
convList += [conv]
else:
# Append other layers such as ReLU and Dropout
convList += [layer]
# Set the conv layers as a nn.Sequential module
convLayers = nn.Sequential(*convList)
| https://stackoverflow.com/questions/44146655/ |
how does the pytorch autograd work? | I submitted this as an issue to cycleGAN pytorch implementation, but since nobody replied me there, i will ask again here.
I'm mainly puzzled by the fact that multiple forward passes was called before one single backward pass, see the following in code cycle_gan_model
# GAN loss
# D_A(G_A(A))
self.fake_B = self.netG_A.forward(self.real_A)
pred_fake = self.netD_A.forward(self.fake_B)
self.loss_G_A = self.criterionGAN(pred_fake, True)
# D_B(G_B(B))
self.fake_A = self.netG_B.forward(self.real_B)
pred_fake = self.netD_B.forward(self.fake_A)
self.loss_G_B = self.criterionGAN(pred_fake, True)
# Forward cycle loss G_B(G_A(A))
self.rec_A = self.netG_B.forward(self.fake_B)
self.loss_cycle_A = self.criterionCycle(self.rec_A, self.real_A) * lambda_A
# Backward cycle loss G_A(G_B(B))
self.rec_B = self.netG_A.forward(self.fake_A)
self.loss_cycle_B = self.criterionCycle(self.rec_B, self.real_B) * lambda_B
# combined loss
self.loss_G = self.loss_G_A + self.loss_G_B + self.loss_cycle_A + self.loss_cycle_B + self.loss_idt_A + self.loss_idt_B
self.loss_G.backward()
The way I see it, G_A and G_B each has three forward passes, twice accepting the real data (real_A or real_B) and twice the fake data (fake_B or fake_A).
In tensorflow (I think) the backward pass is always computed w.r.t the last input data. In this case, the backpropagation of loss_G would be wrong. One should instead do backward pass thrice, each immediately following their involving forward pass.
Specifically, netG_A's gradients from loss_G_A is w.r.t real_A but its gradients from loss_cycle_B is w.r.t fake_A.
I assume this is somehow taken care of in pytorch. But how does the model know w.r.t which input data it should compute the gradients?
| Pytorch uses a tape based system for automatic differentiation. that means that it will backpropagate from the last operation it did. I think that the best way to understand is make a diagram from the process. I attach one that I did by hand
Now you will see that some modules are "repeated". The way I think about them is the same way I think about RNNs; in that way, the gradients will just be added.
| https://stackoverflow.com/questions/44198736/ |
Convolutional NN for text input in PyTorch | I am trying to implement a text classification model using a CNN. As far as I know, for text data, we should use 1d Convolutions. I saw an example in pytorch using Conv2d but I want to know how can I apply Conv1d for text? Or, it is actually not possible?
Here is my model scenario:
Number of in-channels: 1, Number of out-channels: 128
Kernel size : 3 (only want to consider trigrams)
Batch size : 16
So, I will provide tensors of shape, <16, 1, 28, 300> where 28 is the length of a sentence. I want to use Conv1d which will give me 128 feature maps of length 26 (as I am considering trigrams).
I am not sure, how to define nn.Conv1d() for this setting. I can use Conv2d but want to know is it possible to achieve the same using Conv1d?
| This example of Conv1d and Pool1d layers into an RNN resolved my issue.
So, I need to consider the embedding dimension as the number of in-channels while using nn.Conv1d as follows.
m = nn.Conv1d(200, 10, 2) # in-channels = 200, out-channels = 10
input = Variable(torch.randn(10, 200, 5)) # 200 = embedding dim, 5 = seq length
feature_maps = m(input)
print(feature_maps.size()) # feature_maps size = 10,10,4
| https://stackoverflow.com/questions/44212831/ |
OverflowError: (34, 'Numerical result out of range') in PyTorch | I am getting the following error (see the stacktrace) when I ran my code in a different GPU (Tesla K-20, cuda 7.5 installed, 6GB memory). Code works fine if I run in GeForce 1080 or Titan X GPU.
Stacktrace:
File "code/source/main.py", line 68, in <module>
train.train_epochs(train_batches, dev_batches, args.epochs)
File "/gpfs/home/g/e/geniiexe/BigRed2/code/source/train.py", line 34, in train_epochs
losses = self.train(train_batches, dev_batches, (epoch + 1))
File "/gpfs/home/g/e/geniiexe/BigRed2/code/source/train.py", line 76, in train
self.optimizer.step()
File "/gpfs/home/g/e/geniiexe/BigRed2/anaconda3/lib/python3.5/site-packages/torch/optim/adam.py", line 70, in step
bias_correction1 = 1 - beta1 ** state['step']
OverflowError: (34, 'Numerical result out of range')
So, what can be the reason to get such error in a different GPU (Tesla K-20) while it works fine in GeForce or Titan X GPU? Moreover what the error means? Is it related to memory overflow which I don't think so.
| One workaround that is suggested in discuss.pytorch.org is as follows.
Replacing the following lines in adam.py:-
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']
BY
bias_correction1 = 1 - beta1 ** min(state['step'], 1022)
bias_correction2 = 1 - beta2 ** min(state['step'], 1022)
| https://stackoverflow.com/questions/44230829/ |
KeyError: 'unexpected key "module.encoder.embedding.weight" in state_dict' | I am getting the following error while trying to load a saved model.
KeyError: 'unexpected key "module.encoder.embedding.weight" in state_dict'
This is the function I am using to load a saved model.
def load_model_states(model, tag):
"""Load a previously saved model states."""
filename = os.path.join(args.save_path, tag)
with open(filename, 'rb') as f:
model.load_state_dict(torch.load(f))
The model is a sequence-to-sequence network whose init function (constructor) is given below.
def __init__(self, dictionary, embedding_index, max_sent_length, args):
""""Constructor of the class."""
super(Sequence2Sequence, self).__init__()
self.dictionary = dictionary
self.embedding_index = embedding_index
self.config = args
self.encoder = Encoder(len(self.dictionary), self.config)
self.decoder = AttentionDecoder(len(self.dictionary), max_sent_length, self.config)
self.criterion = nn.NLLLoss() # Negative log-likelihood loss
# Initializing the weight parameters for the embedding layer in the encoder.
self.encoder.init_embedding_weights(self.dictionary, self.embedding_index, self.config.emsize)
When I print the model (sequence-to-sequence network), I get the following.
Sequence2Sequence (
(encoder): Encoder (
(drop): Dropout (p = 0.25)
(embedding): Embedding(43723, 300)
(rnn): LSTM(300, 300, batch_first=True, dropout=0.25)
)
(decoder): AttentionDecoder (
(embedding): Embedding(43723, 300)
(attn): Linear (600 -> 12)
(attn_combine): Linear (600 -> 300)
(drop): Dropout (p = 0.25)
(out): Linear (300 -> 43723)
(rnn): LSTM(300, 300, batch_first=True, dropout=0.25)
)
(criterion): NLLLoss (
)
)
So, module.encoder.embedding is an embedding layer, and module.encoder.embedding.weight represents the associated weight matrix. So, why it says- unexpected key "module.encoder.embedding.weight" in state_dict?
| I solved the problem. Actually I was saving the model using nn.DataParallel, which stores the model in module, and then I was trying to load it without DataParallel. So, either I need to add a nn.DataParallel temporarily in my network for loading purposes, or I can load the weights file, create a new ordered dict without the module prefix, and load it back.
The second workaround looks like the following.
# original saved file with DataParallel
state_dict = torch.load('myfile.pth.tar')
# create new OrderedDict that does not contain `module.`
from collections import OrderedDict
new_state_dict = OrderedDict()
for k, v in state_dict.items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
# load params
model.load_state_dict(new_state_dict)
Reference: https://discuss.pytorch.org/t/solved-keyerror-unexpected-key-module-encoder-embedding-weight-in-state-dict/1686
| https://stackoverflow.com/questions/44230907/ |
backward, grad function in pytorch | I'm trying to implement backward, grad function in pytorch.
But, I don't know why this value is returned.
Here is my code.
x = Variable(torch.FloatTensor([[1,2],[3,4]]), requires_grad=True)
y = x + 2
z = y * y
gradient = torch.ones(2, 2)
z.backward(gradient)
print(x.grad)
I think that result value should be [[6,8],[10,12]]
Because of dz/dx= 2*(x+2) and x=1,2,3,4
But returned value is [[7,9],[11,13]]
Why this is happened.. I want to know how gradient, grad function is doing.
Help me please..
| The below piece of code on pytorch v0.12.1
import torch
from torch.autograd import Variable
x = Variable(torch.FloatTensor([[1,2],[3,4]]), requires_grad=True)
y = x + 2
z = y * y
gradient = torch.ones(2, 2)
z.backward(gradient)
print(x.grad)
returns
Variable containing:
6 8
10 12
[torch.FloatTensor of size 2x2]
Update your pytorch installation. This explains the working of autograd, which handles gradient computation for pytorch.
| https://stackoverflow.com/questions/44264018/ |
Why should be the function backward be called only on 1 element tensor or with gradients w.r.t to Variable? | I am new to pytorch. I want to understand as to why we can't call the backward function on a variable containing a tensor of say size say [2,2].
And if we do want to call it on a variable containing tensor of say size say [2,2], we have to do that by first defining a gradient tensor and then calling the backward function on the variable containing the tensor w.r.t the defined gradients.
| from the tutorial on autograd
If you want to compute the derivatives, you can call .backward() on a
Variable. If Variable is a scalar (i.e. it holds a one element data),
you don’t need to specify any arguments to backward(), however if it
has more elements, you need to specify a grad_output argument that is
a tensor of matching shape.
Basically to start the chain rule you need a gradient AT THE OUTPUT, to get it going. In the event the output is a scalar loss function ( which it usually is - normally you are beginning the backward pass at the loss variable ) , its an implied value of 1.0
from tutorial :
let's backprop now out.backward() is equivalent to doing
out.backward(torch.Tensor([1.0]))
but maybe you only want to update a subgraph ( somewhere deep in the network) ... and the value of a Variable is a matrix of weights. Then you have to tell it where the begin. From one of their chief devs ( somewhere in the links )
Yes, that's correct. We only support differentiation of scalar
functions, so if you want to start backward form a non-scalar value
you need to provide dout / dy
The gradients argument
https://discuss.pytorch.org/t/how-the-backward-works-for-torch-variable/907/8 ok explanation
Pytorch, what are the gradient arguments good explanation
http://pytorch.org/tutorials/beginner/blitz/autograd_tutorial.html tutorial
| https://stackoverflow.com/questions/44264443/ |
PyTorch: access weights of a specific module in nn.Sequential() | When I use a pre-defined module in PyTorch, I can typically access its weights fairly easily. However, how do I access them if I wrapped the module in nn.Sequential() first? r.g:
class My_Model_1(nn.Module):
def __init__(self,D_in,D_out):
super(My_Model_1, self).__init__()
self.layer = nn.Linear(D_in,D_out)
def forward(self,x):
out = self.layer(x)
return out
class My_Model_2(nn.Module):
def __init__(self,D_in,D_out):
super(My_Model_2, self).__init__()
self.layer = nn.Sequential(nn.Linear(D_in,D_out))
def forward(self,x):
out = self.layer(x)
return out
model_1 = My_Model_1(10,10)
print(model_1.layer.weight)
model_2 = My_Model_2(10,10)
How do I print the weights now?
model_2.layer.0.weight doesn't work.
| From the PyTorch forum, this is the recommended way:
model_2.layer[0].weight
| https://stackoverflow.com/questions/44284506/ |
Deep learning: save and load a universal machine model through different libraries | My questions can be divided into two parts.
Is there a format of machine learning model file that can be used through different libraries? For example, I saved a model by pytorch, then load it using tensorflow?
If not, is there a library that can help transfer the formats so that a pytorch machine learning model can be used directly in keras?
The reason why I ask this question is that recently I need to adjust some of my previous trained models in tensorflow to pytorch.
An update for this question:
Facebook and Microsoft are going to launch a model standard called ONNX, which is used for transferring models between different framworks, for example between Pytorch to Caffe2. Link in the following:
https://research.fb.com/facebook-and-microsoft-introduce-new-open-ecosystem-for-interchangeable-ai-frameworks/
An further update for this question:
Tensorflow itself use Protocol Buffer format to store model file, which can be used for transfer between different models. Link in the following:
https://www.tensorflow.org/extend/tool_developers/
| Very interesting question. A neural network is a mathematical abstraction consisting of a network of layers (convolution, recurrent, ...), operations (dot product, non-linearity, ...) and their respective parameters (weights, biases).
AFAIK, there's not an universal model file. Nonetheless, different libraries allow users to save their models in a binary format.
There's no library for conversion but there's effort on github repo that addresses this question.
| https://stackoverflow.com/questions/44318616/ |
How to get a uniform distribution in a range [r1,r2] in PyTorch? | I want to get a 2-D torch.Tensor with size [a,b] filled with values from a uniform distribution (in range [r1,r2]) in PyTorch.
| If U is a random variable uniformly distributed on [0, 1], then (r1 - r2) * U + r2 is uniformly distributed on [r1, r2].
Thus, you just need:
(r1 - r2) * torch.rand(a, b) + r2
Alternatively, you can simply use:
torch.FloatTensor(a, b).uniform_(r1, r2)
To fully explain this formulation, let's look at some concrete numbers:
r1 = 2 # Create uniform random numbers in half-open interval [2.0, 5.0)
r2 = 5
a = 1 # Create tensor shape 1 x 7
b = 7
We can break down the expression (r1 - r2) * torch.rand(a, b) + r2 as follows:
torch.rand(a, b) produces an a x b (1x7) tensor with numbers uniformly distributed in the range [0.0, 1.0).
x = torch.rand(a, b)
print(x)
# tensor([[0.5671, 0.9814, 0.8324, 0.0241, 0.2072, 0.6192, 0.4704]])
(r1 - r2) * torch.rand(a, b) produces numbers distributed in the uniform range [0.0, -3.0)
print((r1 - r2) * x)
tensor([[-1.7014, -2.9441, -2.4972, -0.0722, -0.6216, -1.8577, -1.4112]])
(r1 - r2) * torch.rand(a, b) + r2 produces numbers in the uniform range [5.0, 2.0)
print((r1 - r2) * x + r2)
tensor([[3.2986, 2.0559, 2.5028, 4.9278, 4.3784, 3.1423, 3.5888]])
Now, let's break down the answer suggested by @Jonasson: (r2 - r1) * torch.rand(a, b) + r1
Again, torch.rand(a, b) produces (1x7) numbers uniformly distributed in the range [0.0, 1.0).
x = torch.rand(a, b)
print(x)
# tensor([[0.5671, 0.9814, 0.8324, 0.0241, 0.2072, 0.6192, 0.4704]])
(r2 - r1) * torch.rand(a, b) produces numbers uniformly distributed in the range [0.0, 3.0).
print((r2 - r1) * x)
# tensor([[1.7014, 2.9441, 2.4972, 0.0722, 0.6216, 1.8577, 1.4112]])
(r2 - r1) * torch.rand(a, b) + r1 produces numbers uniformly distributed in the range [2.0, 5.0)
print((r2 - r1) * x + r1)
tensor([[3.7014, 4.9441, 4.4972, 2.0722, 2.6216, 3.8577, 3.4112]])
In summary, (r1 - r2) * torch.rand(a, b) + r2 produces numbers in the range [r2, r1), while (r2 - r1) * torch.rand(a, b) + r1 produces numbers in the range [r1, r2).
| https://stackoverflow.com/questions/44328530/ |
How can I install python modules in a docker image? | I have an image called: Image and a running container called: container. I want to install pytorch and anacoda. what's the easiest way to do this? Do I have to change the dockerfile and build a new image? Thanks a lot.
| Yes, the best thing is to build your image in such a way it has the python modules are in there.
Here is an example. I build an image with the build dependencies:
$ docker build -t oz123/alpine-test-mycoolapp:0.5 - < Image
Sending build context to Docker daemon 2.56 kB
Step 1 : FROM alpine:3.5
---> 88e169ea8f46
Step 2 : ENV MB_VERSION 3.1.4
---> Running in 4587d36fa4ae
---> b7c55df49803
Removing intermediate container 4587d36fa4ae
Step 3 : ENV CFLAGS -O2
---> Running in 19fe06dcc314
---> 31f6a4f27d4b
Removing intermediate container 19fe06dcc314
Step 4 : RUN apk add --no-cache python3 py3-pip gcc python3-dev py3-cffi file git curl autoconf automake py3-cryptography linux-headers musl-dev libffi-dev openssl-dev build-base
---> Running in f01b60b1b5b9
fetch http://dl-cdn.alpinelinux.org/alpine/v3.5/main/x86_64/APKINDEX.tar.gz
fetch http://dl-cdn.alpinelinux.org/alpine/v3.5/community/x86_64/APKINDEX.tar.gz
(1/57) Upgrading musl (1.1.15-r5 -> 1.1.15-r6)
(2/57) Upgrading zlib (1.2.8-r2 -> 1.2.11-r0)
(3/57) Installing m4 (1.4.17-r1)
(4/57) Installing perl (5.24.0-r0)
(5/57) Installing autoconf (2.69-r0)
(6/57) Installing automake (1.15-r0)
(7/57) Installing binutils-libs (2.27-r1)
...
Note, I am installing Python's pip inside the image, so later I can download packages from pypi. Packages like numpy might require a C compiler and tool chain, so I am installing these too.
After building the packages which require the build tools chain I remove the tool chain packages:
RUN apk del file pkgconf autoconf m4 automake perl g++ libstdc++
After you have your base image, you can run your application code in
an image building on top of it:
$ cat Dockerfile
FROM oz123/alpine-test-mycoolapp
ADD . /code
WORKDIR /code
RUN pip3 install -r requirements.txt -r requirements_dev.txt
RUN pip3 install -e .
RUN make clean
CMD ["pytest", "-vv", "-s"]
I simply run this with docker.
| https://stackoverflow.com/questions/44339375/ |
How to convert Pytorch autograd.Variable to Numpy? | The title says it all. I want to convert a PyTorch autograd.Variable to its equivalent numpy array. In their official documentation they advocated using a.numpy() to get the equivalent numpy array (for PyTorch tensor). But this gives me the following error:
Traceback (most recent call last): File "stdin", line 1, in module
File
"/home/bishwajit/anaconda3/lib/python3.6/site-packages/torch/autograd/variable.py",
line 63, in getattr raise AttributeError(name) AttributeError:
numpy
Is there any way I can circumvent this?
| Two possible case
Using GPU: If you try to convert a cuda float-tensor directly to numpy like shown below,it will throw an error.
x.data.numpy()
RuntimeError: numpy conversion for FloatTensor is not supported
So, you cant covert a cuda float-tensor directly to numpy, instead you have to convert it into a cpu float-tensor first, and try converting into numpy, like shown below.
x.data.cpu().numpy()
Using CPU: Converting a CPU tensor is straight forward.
x.data.numpy()
| https://stackoverflow.com/questions/44340848/ |
PyTorch Linear layer input dimension mismatch | Im getting this error when passing the input data to the Linear (Fully Connected Layer) in PyTorch:
matrices expected, got 4D, 2D tensors
I fully understand the problem since the input data has a shape (N,C,H,W) (from a Convolutional+MaxPool layer) where:
N: Data Samples
C: Channels of the data
H,W: Height and Width
Nevertheless I was expecting PyTorch to do the "reshaping" of the data form:
[ N , D1,...Dn] --> [ N, D] where D = D1*D2*....Dn
I try to reshape the Variable.data, but I've read that this approach is not recommended since the gradients will conserve the previous shape, and that in general you should not mutate a Variable.data shape.
I am pretty sure there is a simple solution that goes along with the framework, but i haven't find it.
Is there a good solution for this?
PD: The Fully connected layer has as input size the value C * H * W
| After reading some Examples I found the solution. here is how you do it without messing up the forward/backward pass flow:
(_, C, H, W) = x.data.size()
x = x.view( -1 , C * H * W)
| https://stackoverflow.com/questions/44357055/ |
What is the relationship between PyTorch and Torch? | There are two PyTorch repositories :
https://github.com/hughperkins/pytorch
https://github.com/pytorch/pytorch
The first clearly requires Torch and lua and is a wrapper, but the second doesn't make any reference to the Torch project except with its name.
How is it related to the Lua Torch?
| Here a short comparison on pytorch and torch.
Torch:
A Tensor library like numpy, unlike numpy it has strong GPU support.
Lua is a wrapper for Torch (Yes! you need to have a good understanding of Lua), and for that you will need LuaRocks package manager.
PyTorch:
No need for the LuaRocks package manager, no need to write code in Lua. And because we are using Python, we can develop Deep Learning models with utmost flexibility. We can also exploit major Python packages likes scipy, numpy, matplotlib and Cython with PyTorch's own autograd.
There is a detailed discussion on this on pytorch forum. Adding to that both PyTorch and Torch use THNN. Torch provides lua wrappers to the THNN library while Pytorch provides Python wrappers for the same.
PyTorch's recurrent nets, weight sharing and memory usage with the flexibility of interfacing with C, and the current speed of Torch.
For more insights, have a look at this discussion session here.
| https://stackoverflow.com/questions/44371560/ |
Handling C++ arrays in Cython (with numpy and pytorch) | I am trying to use cython to wrap a C++ library (fastText, if its relevant). The C++ library classes load a very large array from disk. My wrapper instantiates a class from the C++ library to load the array, then uses cython memory views and numpy.asarray to turn the array into a numpy array, then calls torch.from_numpy to create a tensor.
The problem arising is how to handle deallocating the memory for the array.
Right now, I get pointer being freed was not allocated when the program exits. This is, I expect, because both the C++ code and numpy/pytorch are trying to manage the same chunk of RAM.
I could simply comment out the destructor in the C++ library, but that feels like its going to cause me a different problem down the road.
How should I approach the issue? Is there any kind of best practices documentation somewhere on how to handle memory sharing with C++ and cython?
If I modify the C++ library to wrap the array in a shared_ptr, will cython (and numpy, pytorch, etc.) share the shared_ptr properly?
I apologize if the question is naive; Python garbage collection is very mysterious to me.
Any advice is appreciated.
| I can think of three sensible ways of doing it. I'll outline them below (i.e. none of the code will be complete but hopefully it will be clear how to finish it).
1. C++ owns the memory; Cython/Python holds a shared pointer to the C++ class
(This is looks to be the lines you're already thinking along).
Start by creating a Cython class that holds a shared pointer
from libcpp.memory cimport shared_ptr
cdef class Holder:
cdef shared_ptr[cpp_class] ptr
@staticmethod
cdef make_holder(shared_ptr[cpp_class] ptr):
cdef holder = Holder() # empty class
holder.ptr = ptr
return holder
You then need to define the buffer protocol for Holder. This allows direct access to the memory allocated by cpp_class in a way that both numpy arrays and Cython memoryviews can understand. Thus they hold a reference to a Holder instance, which in turn keeps a cpp_class alive. (Use np.asarray(holder_instance) to create a numpy array that uses the instance's memory)
The buffer protocol is a little involved but Cython has fairly extensive documentation and you should largely be able to copy and paste their examples. The two methods you need to add to Holder are __getbuffer__ and __releasebuffer__.
2. Python owns the memory; Your C++ class holds a pointer to the Python object
In this version you allocate the memory as a numpy array (using the Python C API interface). When your C++ class is destructed in decrements the reference count of the array, however if Python holds references to that array then the array can outlive the C++ class.
#include <numpy/arrayobject.h>
#include <Python.h>
class cpp_class {
private:
PyObject* arr;
double* data;
public:
cpp_class() {
arr = PyArray_SimpleNew(...); // details left to be filled in
data = PyArray_DATA(reinterpret_cast<PyArrayObject*>(arr));
# fill in the data
}
~cpp_class() {
Py_DECREF(arr); // release our reference to it
}
PyObject* get_np_array() {
Py_INCREF(arr); // Cython expects this to be done before it receives a PyObject
return arr;
}
};
See the numpy documentation for details of the how to allocate numpy arrays from C/C++. Be careful of reference counting if you define copy/move constructors.
The Cython wrapper then looks like:
cdef extern from "some_header.hpp":
cdef cppclass cpp_class:
# whatever constructors you want to allow
object get_np_array()
3. C++ transfers ownership of the data to Python/Cython
In this scheme C++ allocates the array, but Cython/Python is responsible for deallocating it. Once ownership is transferred C++ no longer has access to the data.
class cpp_class {
public:
double* data; // for simplicity this is public - you may want to use accessors
cpp_class() :
data(new double[50])
{/* fill the array as needed */}
~cpp_class() {
delete [] data;
}
};
// helper function for Cython
inline void del_cpp_array(double* a) {
delete [] a;
}
You then use the cython.view.array class to capture the allocated memory. This has a callback function which is used on destruction:
from cython cimport view
cdef extern from "some_header.hpp":
cdef cppclass cpp_class:
double* data
# whatever constructors and other functions
void del_cpp_array(double*)
# later
cdef cpp_class cpp_instance # create this however you like
# ...
# modify line below to match your data
arr = view.array(shape=(10, 2), itemsize=sizeof(double), format="d",
mode="C", allocate_buffer=False)
arr.data = <char*>cpp_instance.data
cpp_instance.data = None # reset to NULL pointer
arr.callback_free_data = del_cpp_array
arr can then be used with a memoryview or a numpy array.
You may have to mess about a bit with casting from void* or char* with del_cpp_array - I'm not sure exactly what types the Cython interface requires.
The first option is probably most work to implement but requires few changes to the C++ code. The second option may require changes to your C++ code that you don't want to make. The third option is simple but means that C++ no longer has access to the data, which might be a disadvantage.
| https://stackoverflow.com/questions/44396749/ |
Implementing Adagrad in Python | I'm trying to implement Adagrad in Python. For learning purposes, I am using matrix factorisation as an example. I'd be using Autograd for computing the gradients.
My main question is if the implementation is fine.
Problem description
Given a matrix A (M x N) having some missing entries, decompose into W and H having sizes (M x k) and (k X N) respectively. Goal would to learn W and H using Adagrad. I'd be following this guide for the Autograd implementation.
NB: I very well know that ALS based implementation are well-suited. I'm using Adagrad only for learning purposes
Customary imports
import autograd.numpy as np
import pandas as pd
Creating the matrix to be decomposed
A = np.array([[3, 4, 5, 2],
[4, 4, 3, 3],
[5, 5, 4, 3]], dtype=np.float32).T
Masking one entry
A[0, 0] = np.NAN
Defining the cost function
def cost(W, H):
pred = np.dot(W, H)
mask = ~np.isnan(A)
return np.sqrt(((pred - A)[mask].flatten() ** 2).mean(axis=None))
Decomposition params
rank = 2
learning_rate=0.01
n_steps = 10000
Gradient of cost wrt params W and H
from autograd import grad, multigrad
grad_cost= multigrad(cost, argnums=[0,1])
Main Adagrad routine (this needs to be checked)
shape = A.shape
# Initialising W and H
H = np.abs(np.random.randn(rank, shape[1]))
W = np.abs(np.random.randn(shape[0], rank))
# gt_w and gt_h contain accumulation of sum of gradients
gt_w = np.zeros_like(W)
gt_h = np.zeros_like(H)
# stability factor
eps = 1e-8
print "Iteration, Cost"
for i in range(n_steps):
if i%1000==0:
print "*"*20
print i,",", cost(W, H)
# computing grad. wrt W and H
del_W, del_H = grad_cost(W, H)
# Adding square of gradient
gt_w+= np.square(del_W)
gt_h+= np.square(del_H)
# modified learning rate
mod_learning_rate_W = np.divide(learning_rate, np.sqrt(gt_w+eps))
mod_learning_rate_H = np.divide(learning_rate, np.sqrt(gt_h+eps))
W = W-del_W*mod_learning_rate_W
H = H-del_H*mod_learning_rate_H
While the problem converges and I get a reasonable solution, I was wondering if the implementation is correct. Specifically, if the understanding of sum of gradients and then computing the adaptive learning rate is correct or not?
| At a cursory glance, your code closely matches that at https://github.com/benbo/adagrad/blob/master/adagrad.py
del_W, del_H = grad_cost(W, H)
matches
grad=f_grad(w,sd,*args)
gt_w+= np.square(del_W)
gt_h+= np.square(del_H)
matches
gti+=grad**2
mod_learning_rate_W = np.divide(learning_rate, np.sqrt(gt_w+eps))
mod_learning_rate_H = np.divide(learning_rate, np.sqrt(gt_h+eps))
matches
adjusted_grad = grad / (fudge_factor + np.sqrt(gti))
W = W-del_W*mod_learning_rate_W
H = H-del_H*mod_learning_rate_H
matches
w = w - stepsize*adjusted_grad
So, assuming that adagrad.py is correct and the translation is correct, would make your code correct. (consensus does not prove your code to be right, but it could be a hint)
| https://stackoverflow.com/questions/44405297/ |
pytorch custom layer "is not a Module subclass" | I am new to PyTorch, trying it out after using a different toolkit for a while.
I would like understand how to program custom layers and functions. And as a simple test, I wrote this:
class Testme(nn.Module): ## it _is_ a sublcass of module ##
def __init__(self):
super(Testme, self).__init__()
def forward(self, x):
return x / t_.max(x)
which is intended to cause the data passing through it to sum to 1. Not actually useful, just at test.
Then I plug it to the example code from the PyTorch Playground:
def make_layers(cfg, batch_norm=False):
layers = []
in_channels = 3
for i, v in enumerate(cfg):
if v == 'M':
layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
else:
padding = v[1] if isinstance(v, tuple) else 1
out_channels = v[0] if isinstance(v, tuple) else v
conv2d = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=padding)
if batch_norm:
layers += [conv2d, nn.BatchNorm2d(out_channels, affine=False), nn.ReLU()]
else:
layers += [conv2d, nn.ReLU()]
layers += [Testme] # here <------------------
in_channels = out_channels
return nn.Sequential(*layers)
The result is an error!
TypeError: model.Testme is not a Module subclass
Maybe this needs to be a Function rather than a Module? Also not clear what the difference is between Function, Module.
For example, why does a Function need a backward(), even if it is constructed entirely from standard pytorch primitive, whereas a Module does not need this?
| That's a simple one. You almost got it, but you forgot to actually create an instance of your new class Testme. You need to do this, even if the creation of an instance of a particular class doesn't take any parameters (as for Testme). But it's easier to forget than for a convolutional layer, to which you typically pass a lot of arguments.
Change the line you have indicated to the following and your problem is resolved.
layers += [Testme()]
| https://stackoverflow.com/questions/44406819/ |
when is a pytorch custom function needed (rather than only a module)? | Pytorch beginner here! Consider the following custom Module:
class Testme(nn.Module):
def __init__(self):
super(Testme, self).__init__()
def forward(self, x):
return x / t_.max(x).expand_as(x)
As far as I understand the documentation:
I believe this could also be implemented as a custom Function. A subclass of Function requires a backward() method, but the
Module does not. As well, in the doc example of a Linear Module, it depends on a Linear Function:
class Linear(nn.Module):
def __init__(self, input_features, output_features, bias=True):
...
def forward(self, input):
return Linear()(input, self.weight, self.bias)
The question: I do not understand the relation between Module and Function. In the first listing above (module Testme), should it have an associated function? If not, then it is possible to implement this without a backward method by subclassing Module, so why does a Function always require a backward method?
Perhaps Functions are intended only for functions that are not composed out of existing torch functions? Say differently: maybe modules do not need the associated Function if their forward method is composed entirely from previously defined torch functions?
| This information is gathered and summarised from the official PyTorch Documentaion.
torch.autograd.Functionreally lies at the heart of the autograd package in PyTorch. Any graph you build in PyTorch and any operation you conduct on Variables in PyTorch is based on a Function. Any function requires an __init__(), forward() and backward() method (see more here: http://pytorch.org/docs/notes/extending.html) . This enables PyTorch to compute results and compute gradients for Variables.
nn.Module()in contrast is really just a convenience for organising your model, your different layers, etc. For example, it organises all the trainable parameters in your model in .parameters()and allows you to add another layer to a model easily, etc. etc. It is not the place where you define a backward method, because in the forward() method, you're supposed to use subclasses of Function(), for which you have already defined backward(). Hence, if you have specified the order of operations in forward(), PyTorch already knows how to back-propagate gradients.
Now, when should you use what?
If you have an operation that is just a composition of existing implemented functions in PyTorch (like your thing above), there's really no point adding any subclass to Function() yourself. Because you can just stack operations up and build a dynamic graph. It's however a sensible idea to bunch these operations together. If any operation involves trainable parameters (for example a linear layer of a neural network), you should subclass nn.Module() and bunch your operations together in the forward method. This allows you to easily access parameters (as outlined above) for use of torch.optim, etc. If you don't have any trainable parameters, I would probably still bunch them together, but a standard Python function, where you take care of the instantination of each operation you use would be sufficient.
If you have a new custom operation (e.g. a new stochastic layer with some complicated sampling procedure), you should subclass Function() and define __init__(), forward() and backward() to tell PyTorch how to compute results and how to compute gradients, when you use this operation. Afterwards, you should either create a functional version to take care of instantinating the function and use your operation or create a module, if your operation has trainable parameters. Again, you can read more about this in the link above.
| https://stackoverflow.com/questions/44428784/ |
How to load a list of numpy arrays to pytorch dataset loader? | I have a huge list of numpy arrays, where each array represents an image and I want to load it using torch.utils.data.Dataloader object. But the documentation of torch.utils.data.Dataloader mentions that it loads data directly from a folder. How do I modify it for my cause? I am new to pytorch and any help would be greatly appreciated.
my numpy array for a single image looks something like this. The image is RBG image.
[[[ 70 82 94]
[ 67 81 93]
[ 66 82 94]
...,
[182 182 188]
[183 183 189]
[188 186 192]]
[[ 66 80 92]
[ 62 78 91]
[ 64 79 95]
...,
[176 176 182]
[178 178 184]
[180 180 186]]
[[ 62 82 93]
[ 62 81 96]
[ 65 80 99]
...,
[169 172 177]
[173 173 179]
[172 172 178]]
...,
| I think what DataLoader actually requires is an input that subclasses Dataset. You can either write your own dataset class that subclasses Datasetor use TensorDataset as I have done below:
import torch
import numpy as np
from torch.utils.data import TensorDataset, DataLoader
my_x = [np.array([[1.0,2],[3,4]]),np.array([[5.,6],[7,8]])] # a list of numpy arrays
my_y = [np.array([4.]), np.array([2.])] # another list of numpy arrays (targets)
tensor_x = torch.Tensor(my_x) # transform to torch tensor
tensor_y = torch.Tensor(my_y)
my_dataset = TensorDataset(tensor_x,tensor_y) # create your datset
my_dataloader = DataLoader(my_dataset) # create your dataloader
Works for me. Hope it helps you.
| https://stackoverflow.com/questions/44429199/ |
Creating one hot vector from indices given as a tensor | I have a tensor of size 4 x 6 where 4 is batch size and 6 is sequence length. Every element of the sequence vectors are some index (0 to n). I want to create a 4 x 6 x n tensor where the vectors in 3rd dimension will be one hot encoding of the index which means I want to put 1 in the specified index and rest of the values will be zero.
For example, I have the following tensor:
[[5, 3, 2, 11, 15, 15],
[1, 4, 6, 7, 3, 3],
[2, 4, 7, 8, 9, 10],
[11, 12, 15, 2, 5, 7]]
Here, all the values are in between (0 to n) where n = 15. So, I want to convert the tensor to a 4 X 6 X 16 tensor where the third dimension will represent one hot encoding vector.
How can I do that using PyTorch functionalities? Right now, I am doing this with loop but I want to avoid looping!
| NEW ANSWER
As of PyTorch 1.1, there is a one_hot function in torch.nn.functional. Given any tensor of indices indices and a maximal index n, you can create a one_hot version as follows:
n = 5
indices = torch.randint(0,n, size=(4,7))
one_hot = torch.nn.functional.one_hot(indices, n) # size=(4,7,n)
Very old Answer
At the moment, slicing and indexing can be a bit of a pain in PyTorch from my experience. I assume you don't want to convert your tensors to numpy arrays. The most elegant way I can think of at the moment is to use sparse tensors and then convert to a dense tensor. That would work as follows:
from torch.sparse import FloatTensor as STensor
batch_size = 4
seq_length = 6
feat_dim = 16
batch_idx = torch.LongTensor([i for i in range(batch_size) for s in range(seq_length)])
seq_idx = torch.LongTensor(list(range(seq_length))*batch_size)
feat_idx = torch.LongTensor([[5, 3, 2, 11, 15, 15], [1, 4, 6, 7, 3, 3],
[2, 4, 7, 8, 9, 10], [11, 12, 15, 2, 5, 7]]).view(24,)
my_stack = torch.stack([batch_idx, seq_idx, feat_idx]) # indices must be nDim * nEntries
my_final_array = STensor(my_stack, torch.ones(batch_size * seq_length),
torch.Size([batch_size, seq_length, feat_dim])).to_dense()
print(my_final_array)
Note: PyTorch is undergoing some work currently, that will add numpy style broadcasting and other functionalities within the next two or three weeks and other functionalities. So it's possible, there'll be better solutions available in the near future.
Hope this helps you a bit.
| https://stackoverflow.com/questions/44461772/ |
tensor division in pytorch. Assertion error | This is kind of a pytorch beginner question. In pytorch I'm trying to do element wise division with two tensors of size [5,5,3]. In numpy it works fine using np.divide(), but somehow I get an error here. I'm using PyTorch version 0.1.12 for Python 3.5.
c = [torch.DoubleTensor of size 5x5x3]
input_patch = [torch.FloatTensor of size 5x5x3]
input_patch is a slice of a torch.autograd Variable, and c is made by doing c = torch.from_numpy(self.patch_filt[:, :, :, 0]).float()
When doing:
torch.div(input_patch, c)
I get this error that I don't understand.
line 317, in div
assert not torch.is_tensor(other)
AssertionError
Does it mean that variable c should not be a torch_tensor? After casting c to also be a FloatTensor gives the same error still.
Thank you!
| Input_patch is a slice of a torch.autograd Variable, and c is made by doing
c = torch.from_numpy(self.patch_filt[:, :, :, 0]).float()
Anyway, mexmex, thanks to your comment I've solved it by defining c as
Variable(torch.from_numpy(self.patch_filt[:, :, :, 0])).float()
| https://stackoverflow.com/questions/44518677/ |
How do I multiply matrices in PyTorch? | With numpy, I can do a simple matrix multiplication like this:
a = numpy.ones((3, 2))
b = numpy.ones((2, 1))
result = a.dot(b)
However, this does not work with PyTorch:
a = torch.ones((3, 2))
b = torch.ones((2, 1))
result = torch.dot(a, b)
This code throws the following error:
RuntimeError: 1D tensors expected, but got 2D and 2D tensors
How do I perform matrix multiplication in PyTorch?
| Use torch.mm:
torch.mm(a, b)
torch.dot() behaves differently to np.dot(). There's been some discussion about what would be desirable here. Specifically, torch.dot() treats both a and b as 1D vectors (irrespective of their original shape) and computes their inner product. The error is thrown because this behaviour makes your a a vector of length 6 and your b a vector of length 2; hence their inner product can't be computed. For matrix multiplication in PyTorch, use torch.mm(). Numpy's np.dot() in contrast is more flexible; it computes the inner product for 1D arrays and performs matrix multiplication for 2D arrays.
torch.matmul performs matrix multiplications if both arguments are 2D and computes their dot product if both arguments are 1D. For inputs of such dimensions, its behaviour is the same as np.dot. It also lets you do broadcasting or matrix x matrix, matrix x vector and vector x vector operations in batches.
# 1D inputs, same as torch.dot
a = torch.rand(n)
b = torch.rand(n)
torch.matmul(a, b) # torch.Size([])
# 2D inputs, same as torch.mm
a = torch.rand(m, k)
b = torch.rand(k, j)
torch.matmul(a, b) # torch.Size([m, j])
| https://stackoverflow.com/questions/44524901/ |
python child process exit unexpectedly with exit code -9 | I have a PyTorch script with 16 processes. Following is a code snippet from the main process:
procs = [mp.Process(target=self.worker_wrapper, args=(i, )) for i in range(self.n_workers)]
for p in procs: p.start()
while True:
time.sleep(60)
for i, p in enumerate(procs):
self.logger.info('Check: id %d, exitcode %s, alive %s' % (
i, str(p.exitcode), str(p.is_alive())))
the worker_wrapper is this:
def worker_wrapper(self, id):
try:
self.worker(id)
except Exception as e:
self.logger.info(e)
self.logger.error(traceback.format_exc())
The worker will do some stuff and should never stop. It works well at beginning, but usually after 1 days running, some child processes (not all) exit unexpectedly, I get the output like this:
[2017-06-15 08:45:21,540] Check: id 13, exitcode -9, alive False
And it didn't catch any exception. I tried to search exitcode -9 but didn't get anything useful. Does anyone meet this before? Or do you have any idea about how to debug this? I'm sorry I can't provide a minimal reproducible code snippet, because it usually takes one day to reproduce. And I'm using python2.7
| Exit code -9 means the process was killed via SIGKILL.
It's probably due to your machine running out of memory and the OS kernel was triggering the OOM killer.
To verify that, check the kernel logs via dmesg you should notice the OOM killer targeting your worker processes.
You need to reduce the memory footprint of your application (or increase the memory of your server/workstation). Also verify your application is not leaking memory.
| https://stackoverflow.com/questions/44576035/ |
CUDA vs. DataParallel: Why the difference? | I have a simple neural network model and I apply either cuda() or DataParallel() on the model like following.
model = torch.nn.DataParallel(model).cuda()
OR,
model = model.cuda()
When I don't use DataParallel, rather simply transform my model to cuda(), I need to explicitly convert the batch inputs to cuda() and then give it to the model, otherwise it returns the following error.
torch.index_select received an invalid combination of arguments - got (torch.cuda.FloatTensor, int, torch.LongTensor)
But with DataParallel, the code works fine. Rest of the other things are same. Why this happens? Why when I use DataParallel, I don't need to transform the batch inputs explicitly to cuda()?
| Because, DataParallel allows CPU inputs, as it's first step is to transfer inputs to appropriate GPUs.
Info source: https://discuss.pytorch.org/t/cuda-vs-dataparallel-why-the-difference/4062/3
| https://stackoverflow.com/questions/44580450/ |
Custom loss function in PyTorch | I have three simple questions.
What will happen if my custom loss function is not differentiable? Will pytorch through error or do something else?
If I declare a loss variable in my custom function which will represent the final loss of the model, should I put requires_grad = True for that variable? or it doesn't matter? If it doesn't matter, then why?
I have seen people sometimes write a separate layer and compute the loss in the forward function. Which approach is preferable, writing a function or a layer? Why?
I need a clear and nice explanation to these questions to resolve my confusions. Please help.
| Let me have a go.
This depends on what you mean by "non-differentiable". The first definition that makes sense here is that PyTorch doesn't know how to compute gradients. If you try to compute gradients nevertheless, this will raise an error. The two possible scenarios are:
a) You're using a custom PyTorch operation for which gradients have not been implemented, e.g. torch.svd(). In that case you will get a TypeError:
import torch
from torch.autograd import Function
from torch.autograd import Variable
A = Variable(torch.randn(10,10), requires_grad=True)
u, s, v = torch.svd(A) # raises TypeError
b) You have implemented your own operation, but did not define backward(). In this case, you will get a NotImplementedError:
class my_function(Function): # forgot to define backward()
def forward(self, x):
return 2 * x
A = Variable(torch.randn(10,10))
B = my_function()(A)
C = torch.sum(B)
C.backward() # will raise NotImplementedError
The second definition that makes sense is "mathematically non-differentiable". Clearly, an operation which is mathematically not differentiable should either not have a backward() method implemented or a sensible sub-gradient. Consider for example torch.abs() whose backward() method returns the subgradient 0 at 0:
A = Variable(torch.Tensor([-1,0,1]),requires_grad=True)
B = torch.abs(A)
B.backward(torch.Tensor([1,1,1]))
A.grad.data
For these cases, you should refer to the PyTorch documentation directly and dig out the backward() method of the respective operation directly.
It doesn't matter. The use of requires_gradis to avoid unnecessary computations of gradients for subgraphs. If there’s a single input to an operation that requires gradient, its output will also require gradient. Conversely, only if all inputs don’t require gradient, the output also won’t require it. Backward computation is never performed in the subgraphs, where all Variables didn’t require gradients.
Since, there are most likely some Variables (for example parameters of a subclass of nn.Module()), your loss Variable will also require gradients automatically. However, you should notice that exactly for how requires_grad works (see above again), you can only change requires_grad for leaf variables of your graph anyway.
All the custom PyTorch loss functions, are subclasses of _Loss which is a subclass of nn.Module. See here. If you'd like to stick to this convention, you should subclass _Loss when defining your custom loss function. Apart from consistency, one advantage is that your subclass will raise an AssertionError, if you haven't marked your target variables as volatile or requires_grad = False. Another advantage is that you can nest your loss function in nn.Sequential(), because its a nn.Module I would recommend this approach for these reasons.
| https://stackoverflow.com/questions/44597523/ |
Is Torch7 defined-by-run like Pytorch? | Pytorch have Dynamic Neural Networks (defined-by-run) as opposed to Tensorflow which have to compile the computation graph before run.
I see that both Torch7 and PyTorch depend on TH, THC, THNN, THCUNN (C library). Does Torch7 have Dynamic Neural Networks (defined-by-run) feature ?
| No, Torch7 use static computational graphs, as in Tensorflow. It is one of the major differences between PyTorch and Torch7.
| https://stackoverflow.com/questions/44614977/ |
subtraction of scalar from tensor yields 'inconsistent tensor size' in pytorch | I'm using pytorch and my variables are
x = [torch.FloatTensor of size 1x3x32x32]
mean = Variable containing:
1.00000e-02 *
2.0518
[torch.FloatTensor of size 1]
what I want to do is subtract the scalar mean from x by doing
x = x - mean
However, I'm getting this error:
RuntimeError: inconsistent tensor size at /py/conda-
bld/pytorch_1493670682084/work/torch/lib/TH/generic/THTensorMath.c:831
What am I doing wrong?
Thanks a lot
| what you are trying only works if mean is truly a scalar, i.e. a float() (in this case) and not a torch.FloatTensor of size 1. You can either extract a true scalar from mean or expand mean to the size of x in order to perform the subtraction.
To extract the float from mean, do:
x = x - mean[0]
To expand mean to the size of x, do:
x = x - mean.expand_as(x)
Note that both of these methods subtract the mean from each element in your tensor.
| https://stackoverflow.com/questions/44631520/ |
Convert 'int' to pytorch 'Variable' makes problems | First project with pytorch and I got stuck trying to convert an MNIST label 'int' into a torch 'Variable'. Debugger says it has no dimension?!
# numpy mnist data
X_train, Y_train = read_data("training")
X_test , Y_test = read_data("testing")
arr = np.zeros(5)
for i in range(5):
# in your training loop:
costs_ = 0
for k in range(10000):
optimizer.zero_grad() # zero the gradient buffers
a = torch.from_numpy(np.expand_dims(X_train[k].flatten(), axis=0)).float()
b = torch.from_numpy(np.array(Y_train[k], dtype=np.float)).float()
input = Variable(a)
output = net(input)
target = Variable(b) # PROBLEM!!
loss = criterion(output, target)
loss.backward()
optimizer.step() # Does the update
costs_ += loss.data.numpy()
arr[i] = costs_
print(i)
Thrown error is: "RuntimeError: input and target have different number of elements: input[1 x 1] has 1 elements, while target[] has 0 elements at /b/wheel/pytorch-src/torch/lib/THNN/generic/MSECriterion.c:12"
| The error is telling you exactly what is happening. Your target variable is empty.
Edit (after the comment below):
if Y_train[k] = 5, then np.array(Y_train[k], dtype=np.float).shape = (), and in turn Variable(b) becomes a tensor with no dimension.
In order to fix this you will need to pass a list to np.array() and not a integer or a float.
Like this:
b = torch.from_numpy(np.array([Y_train[k]], dtype=np.float)).float()
| https://stackoverflow.com/questions/44631628/ |
Pytorch: how to add L1 regularizer to activations? | I would like to add the L1 regularizer to the activations output from a ReLU.
More generally, how does one add a regularizer only to a particular layer in the network?
Related material:
This similar post refers to adding L2 regularization, but it appears to add the regularization penalty to all layers of the network.
nn.modules.loss.L1Loss() seems relevant, but I do not yet understand how to use this.
The legacy module L1Penalty seems relevant also, but why has it been deprecated?
| Here is how you do this:
In your Module's forward return final output and layers' output for which you want to apply L1 regularization
loss variable will be sum of cross entropy loss of output w.r.t. targets and L1 penalties.
Here's an example code
import torch
from torch.autograd import Variable
from torch.nn import functional as F
class MLP(torch.nn.Module):
def __init__(self):
super(MLP, self).__init__()
self.linear1 = torch.nn.Linear(128, 32)
self.linear2 = torch.nn.Linear(32, 16)
self.linear3 = torch.nn.Linear(16, 2)
def forward(self, x):
layer1_out = F.relu(self.linear1(x))
layer2_out = F.relu(self.linear2(layer1_out))
out = self.linear3(layer2_out)
return out, layer1_out, layer2_out
batchsize = 4
lambda1, lambda2 = 0.5, 0.01
model = MLP()
optimizer = torch.optim.SGD(model.parameters(), lr=1e-4)
# usually following code is looped over all batches
# but let's just do a dummy batch for brevity
inputs = Variable(torch.rand(batchsize, 128))
targets = Variable(torch.ones(batchsize).long())
optimizer.zero_grad()
outputs, layer1_out, layer2_out = model(inputs)
cross_entropy_loss = F.cross_entropy(outputs, targets)
all_linear1_params = torch.cat([x.view(-1) for x in model.linear1.parameters()])
all_linear2_params = torch.cat([x.view(-1) for x in model.linear2.parameters()])
l1_regularization = lambda1 * torch.norm(all_linear1_params, 1)
l2_regularization = lambda2 * torch.norm(all_linear2_params, 2)
loss = cross_entropy_loss + l1_regularization + l2_regularization
loss.backward()
optimizer.step()
| https://stackoverflow.com/questions/44641976/ |
How do you implement variable-length recurrent neural networks? | What is a full working example (not snippets) of variable-length sequence inputs into recurrent neural networks (RNNs)?
For example PyTorch supposedly can implement variable-length sequences as input into RNNs, but there do not seem to be examples of full working code.
Relevant:
https://github.com/pytorch/pytorch/releases/tag/v0.1.10
https://discuss.pytorch.org/t/about-the-variable-length-input-in-rnn-scenario/345
| Sadly, there is no such thing as 'variable length' neural networks. This is because there is no way a network can 'know' which weights to use for extra input nodes that it wasn't trained for.
However, the reason you are seeing a 'variable length' on that page, is because they process:
a b c d e
a b c d e f g h
a b c d
a b
as
a b c d e 0 0 0
a b c d e f g h
a b c d 0 0 0 0
a b 0 0 0 0 0 0
They convert all 'empty' variables to 0. Which makes sense, as 0 does not add anything tot he networks hidden layers regardless of the weights, as anything*0 = 0.
So basically, you can have 'variable length' inputs, but you have to define some kind of maximum size; all inputs that are smaller than that size should be padded with zeros.
If you are classifying sentences on the other hand, you could use LSTM/GRU networks to handle the input sequentially.
| https://stackoverflow.com/questions/44642939/ |
How do you use PyTorch PackedSequence in code? | Can someone give a full working code (not a snippet, but something that runs on a variable-length recurrent neural network) on how would you use the PackedSequence method in PyTorch?
There do not seem to be any examples of this in the documentation, github, or the internet.
https://github.com/pytorch/pytorch/releases/tag/v0.1.10
| Not the most beautiful piece of code, but this is what I gathered for my personal use after going through PyTorch forums and docs. There can be certainly better ways to handle the sorting - restoring part, but I chose it to be in the network itself
EDIT: See answer from @tusonggao which makes torch utils take care of sorting parts
class Encoder(nn.Module):
def __init__(self, vocab_size, embedding_size, embedding_vectors=None, tune_embeddings=True, use_gru=True,
hidden_size=128, num_layers=1, bidrectional=True, dropout=0.6):
super(Encoder, self).__init__()
self.embed = nn.Embedding(vocab_size, embedding_size, padding_idx=0)
self.embed.weight.requires_grad = tune_embeddings
if embedding_vectors is not None:
assert embedding_vectors.shape[0] == vocab_size and embedding_vectors.shape[1] == embedding_size
self.embed.weight = nn.Parameter(torch.FloatTensor(embedding_vectors))
cell = nn.GRU if use_gru else nn.LSTM
self.rnn = cell(input_size=embedding_size, hidden_size=hidden_size, num_layers=num_layers,
batch_first=True, bidirectional=True, dropout=dropout)
def forward(self, x, x_lengths):
sorted_seq_lens, original_ordering = torch.sort(torch.LongTensor(x_lengths), dim=0, descending=True)
ex = self.embed(x[original_ordering])
pack = torch.nn.utils.rnn.pack_padded_sequence(ex, sorted_seq_lens.tolist(), batch_first=True)
out, _ = self.rnn(pack)
unpacked, unpacked_len = torch.nn.utils.rnn.pad_packed_sequence(out, batch_first=True)
indices = Variable(torch.LongTensor(np.array(unpacked_len) - 1).view(-1, 1)
.expand(unpacked.size(0), unpacked.size(2))
.unsqueeze(1))
last_encoded_states = unpacked.gather(dim=1, index=indices).squeeze(dim=1)
scatter_indices = Variable(original_ordering.view(-1, 1).expand_as(last_encoded_states))
encoded_reordered = last_encoded_states.clone().scatter_(dim=0, index=scatter_indices, src=last_encoded_states)
return encoded_reordered
| https://stackoverflow.com/questions/44643137/ |
Pytorch: Convert FloatTensor into DoubleTensor | I have 2 numpy arrays, which I convert into tensors to use the TensorDataset object.
import torch.utils.data as data_utils
X = np.zeros((100,30))
Y = np.zeros((100,30))
train = data_utils.TensorDataset(torch.from_numpy(X).double(), torch.from_numpy(Y))
train_loader = data_utils.DataLoader(train, batch_size=50, shuffle=True)
when I do:
for batch_idx, (data, target) in enumerate(train_loader):
data, target = Variable(data), Variable(target)
optimizer.zero_grad()
output = model(data) # error occurs here
I get the fallowing error:
TypeError: addmm_ received an invalid combination of arguments - got (int, int, torch.DoubleTensor, torch.FloatTensor), but expected one of:
[...]
* (float beta, float alpha, torch.DoubleTensor mat1, torch.DoubleTensor mat2)
didn't match because some of the arguments have invalid types: (int, int, torch.DoubleTensor, torch.FloatTensor)
* (float beta, float alpha, torch.SparseDoubleTensor mat1, torch.DoubleTensor mat2)
didn't match because some of the arguments have invalid types: (int, int, torch.DoubleTensor, torch.FloatTensor)
The last error comes from:
output.addmm_(0, 1, input, weight.t())
As you see in my code I tried converting the tensor by using .double() - but this did not work. Why is he casting one array into a FloatTensor object and the other into a DoubleTensor?
Any ideas?
| Your numpy arrays are 64-bit floating point and will be converted to torch.DoubleTensor standardly. Now, if you use them with your model, you'll need to make sure that your model parameters are also Double. Or you need to make sure, that your numpy arrays are cast as Float, because model parameters are standardly cast as float.
Hence, do either of the following:
data_utils.TensorDataset(torch.from_numpy(X).float(), torch.from_numpy(Y).float())
or do:
model.double()
Depeding, if you want to cast your model parameters, inputs and targets as Float or as Double.
| https://stackoverflow.com/questions/44717100/ |
Why do we need to explicitly call zero_grad()? | Why do we need to explicitly zero the gradients in PyTorch? Why can't gradients be zeroed when loss.backward() is called? What scenario is served by keeping the gradients on the graph and asking the user to explicitly zero the gradients?
| We explicitly need to call zero_grad() because, after loss.backward() (when gradients are computed), we need to use optimizer.step() to proceed gradient descent. More specifically, the gradients are not automatically zeroed because these two operations, loss.backward() and optimizer.step(), are separated, and optimizer.step() requires the just computed gradients.
In addition, sometimes, we need to accumulate gradient among some batches; to do that, we can simply call backward multiple times and optimize once.
| https://stackoverflow.com/questions/44732217/ |
PyTorch: Extract learned weights correctly | I am trying to extract the weights from a linear layer, but they do not appear to change, although error is dropping monotonously (i.e. training is happening). Printing the weights' sum, nothing happens because it stays constant:
np.sum(model.fc2.weight.data.numpy())
Here are the code snippets:
def train(epochs):
model.train()
for epoch in range(1, epochs+1):
# Train on train set
print(np.sum(model.fc2.weight.data.numpy()))
for batch_idx, (data, target) in enumerate(train_loader):
data, target = Variable(data), Variable(data)
optimizer.zero_grad()
output = model(data)
loss = criterion(output, target)
loss.backward()
optimizer.step()
and
# Define model
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
# an affine operation: y = Wx + b
self.fc1 = nn.Linear(100, 80, bias=False)
init.normal(self.fc1.weight, mean=0, std=1)
self.fc2 = nn.Linear(80, 87)
self.fc3 = nn.Linear(87, 94)
self.fc4 = nn.Linear(94, 100)
def forward(self, x):
x = self.fc1(x)
x = F.relu(self.fc2(x))
x = F.relu(self.fc3(x))
x = F.relu(self.fc4(x))
return x
Maybe I am looking on the wrong parameters, although I checked the docs. Thanks for your help!
| Use model.parameters() to get trainable weight for any model or layer. Remember to put it inside list(), or you cannot print it out.
The following code snip worked
>>> import torch
>>> import torch.nn as nn
>>> l = nn.Linear(3,5)
>>> w = list(l.parameters())
>>> w
| https://stackoverflow.com/questions/44734327/ |
Re-weight the input to a neural network | For example, I feed a set of images into a CNN. And the default weight of these images is 1. How can I re-weight some of these images so that they have different weights? Can 'DataLoader' achieve this goal in pytorch?
I learned two other possibilities:
Defining a custom loss function, providing weights for each sample as I require.
Repeating samples in the training set, which will result in more frequent samples having a higher weight in the final loss.
Is there any other way, we can achieve that? Any suggestion would be appreciated.
| I can think of two ways to achieve this.
Pass on the weight explicitly, when you backpropagate the gradients.
After you computed loss, and when you're about to backpropagate, you can pass a Tensor to backward() and all the subsequent gradients will be scaled by the corresponding element, i.e. do something like
loss = torch.pow(out-target,2)
loss.backward(my_weights) # is a Tensor of same dimension as loss
However, if you use want to assign individual weights in a batch, you can't use the custom loss functions from the nn.module which aggregates the loss over all samples in a batch.
Use torch.utils.data.sampler.WeightedRandomSampler
If you use PyTorch's data.utils anyway, this is simpler than multiplying your training set. However it doesn't assign exact weights, since it's stochastic. But if you're iterating over your training set a sufficient number of times, it's probably close enough.
| https://stackoverflow.com/questions/44743305/ |
Torch sum a tensor along an axis | How do I sum over the columns of a tensor?
torch.Size([10, 100]) ---> torch.Size([10])
| The simplest and best solution is to use torch.sum().
To sum all elements of a tensor:
torch.sum(x) # gives back a scalar
To sum over all rows (i.e. for each column):
torch.sum(x, dim=0) # size = [ncol]
To sum over all columns (i.e. for each row):
torch.sum(x, dim=1) # size = [nrow]
It should be noted that the dimension summed over is eliminated from the resulting tensor.
| https://stackoverflow.com/questions/44790670/ |
Apply json config file using parse_args() in pycharm | I'm running sequence to sequence code in git, but I got error about parse_args().
my code is like this:
parser = argparse.ArgumentParser()
parser.add_argument(
"--config",
help="path to json config",
required=True)
args = parser.parse_args()
config_file_path = args.config
config = read_config(config_file_path)
experiment_name = hyperparam_string(config)
my config file is like this :
{
"training": {
"optimizer": "adam",
"clip_c": 1,
"lrate": 0.0002,
},
"management": {
"monitor_loss": 1000,
"print_samples": 20000
}
When I run
args = parser.parse_args()
Pycharm raises error
pydevconsole.py: error: argument --config is required
I'd like to know that run this code through pycharm applying json config file.
I have searched in google since yesterday, but I can't find it.
Please help..
| args = parser.parse_args() parses the sys.argv[1:] list, which is provided to the interpreter from the operating system shell - ie. from the commandline.
$:python prog.py --config afilename
You can also do
args = parser.parse_args(['--config', 'afilename'])
this handy during testing.
It also helps to:
import sys
print(sys.argv)
| https://stackoverflow.com/questions/44815875/ |
how to append a singleton numpy array item into a list? | I am attaching the code first which will give you a better idea
` prediction = prediction.data.max(1)[1] #gives a tensor value
prediction = (prediction.cpu().numpy().item()) #converts that tensor into a numpy array
result.append(int_to_word[prediction])`
I am using pytorch for word generation. The line prediction = prediction.data.max(1)[1] gives us the class label with the maximum probability, which turns out to be a tensor value.
prediction = (prediction.cpu().numpy().item()) this statement converts the tensor into a numpy array and the funtion item() extracts the value from the array.
Now when I try to append this value using pattern.append(prediction)
I am getting the following error
'numpy.ndarray' object has no attribute 'append'
I am not able to understand as to why I am getting this error. I have already converted the numpy array into a scalr value, didn't I?? Still, why am I getting that error?
Can anybody please explain and provide a solution to this. I would be most grateful.
| Numpy arrays are immutable with respect to their dimensions. They do not support the append operation. You'll have to declare results as a list, then append your values to your list, and then convert it to a numpy array:
result = []
...
result.append(prediction) # inside some loop
...
result = np.array(result)
| https://stackoverflow.com/questions/44819563/ |
No N-dimensional tranpose in PyTorch | PyTorch's torch.transpose function only transposes 2D inputs. Documentation is here.
On the other hand, Tensorflow's tf.transpose function allows you to transpose a tensor of N arbitrary dimensions.
Can someone please explain why PyTorch does not/cannot have N-dimension transpose functionality? Is this due to the dynamic nature of the computation graph construction in PyTorch versus Tensorflow's Define-then-Run paradigm?
| It's simply called differently in pytorch. torch.Tensor.permute will allow you to swap dimensions in pytorch like tf.transpose does in TensorFlow.
As an example of how you'd convert a 4D image tensor from NHWC to NCHW (not tested, so might contain bugs):
>>> img_nhwc = torch.randn(10, 480, 640, 3)
>>> img_nhwc.size()
torch.Size([10, 480, 640, 3])
>>> img_nchw = img_nhwc.permute(0, 3, 1, 2)
>>> img_nchw.size()
torch.Size([10, 3, 480, 640])
| https://stackoverflow.com/questions/44841654/ |
How to debug(monitoring value of object in other class' function) in Pycharm | I'm running seq2seq code in pycharm in order to study pytorch.
The code has many classes and these classes have many function.
I'd like to monitor value of objects in other function, so I'm running code in console one by one.
Is there any good way to this using debug?
I haven't done debug before.
Please help me..
| I'm not familiar with these tools specifically, but here is how I would approach it. It's also kinda hard to express how to properly use a gui interactively through text, so if you are new to a debugger in general it might be good to start with some tutorials. Jetbrains has some PyCharm debugger tutorials online.
PyCharm debugger tutorial 1
PyCharm debugger tutorial 2
When you are running the debugger, set breakpoints and you can see all of the local variables in the scope to your current object. If you are wanting to monitor 2 places, you could set 2 breakpoints. Or you could stop at one and move forward (look at Step Over, F8 and Step Into, F7 until the second object is available. I think specifically for you I would look at the Debugger, Frames. Essentially you can jump backwards in time from your current breakpoint to where your current function was called, and so on and so forth for ~10 calls. This might get you what you are looking for, but it is somewhat project dependent unfortunately.
| https://stackoverflow.com/questions/44843773/ |
How do you alter the size of a Pytorch Dataset? | Say I am loading MNIST from torchvision.datasets.MNIST, but I only want to load in 10000 images total, how would I slice the data to limit it to only some number of data points? I understand that the DataLoader is a generator yielding data in the size of the specified batch size, but how do you slice datasets?
tr = datasets.MNIST('../data', train=True, download=True, transform=transform)
te = datasets.MNIST('../data', train=False, transform=transform)
train_loader = DataLoader(tr, batch_size=args.batch_size, shuffle=True, num_workers=4, **kwargs)
test_loader = DataLoader(te, batch_size=args.batch_size, shuffle=True, num_workers=4, **kwargs)
| It is important to note that when you create the DataLoader object, it doesnt immediately load all of your data (its impractical for large datasets). It provides you an iterator that you can use to access each sample.
Unfortunately, DataLoader doesnt provide you with any way to control the number of samples you wish to extract. You will have to use the typical ways of slicing iterators.
Simplest thing to do (without any libraries) would be to stop after the required number of samples is reached.
nsamples = 10000
for i, image, label in enumerate(train_loader):
if i > nsamples:
break
# Your training code here.
Or, you could use itertools.islice to get the first 10k samples. Like so.
for image, label in itertools.islice(train_loader, stop=10000):
# your training code here.
| https://stackoverflow.com/questions/44856691/ |
Find number of non-zero elements in a tensor along an aixs | I want to find the number of non-zero elements in a tensor along a particular axis. Is there any PyTorch function which can do this?
I tried to use the nonzero() method in PyTorch.
torch.nonzero(losses).size(0)
Here, lossess is a tensor of shape 64 x 1. When I run the above statement, it gives me the following error.
TypeError: Type Variable doesn't implement stateless method nonzero
But if I run, torch.nonzero(losses.data).size(0), then it works fine. Any clue, why this is happening or what the error means?
| Meaning of the error message - TypeError: Type Variable doesn't implement stateless method nonzero is, we cannot use torch.nonzero() on autograd.Variable but only on simple tensors. Also it should be noted that, tensors are stateless while the Variables are stateful.
| https://stackoverflow.com/questions/44857373/ |
PyTorch - Element-wise multiplication between a variable and a tensor? | As of PyTorch 0.4 this question is no longer valid. In 0.4 Tensors and Variables were merged.
How can I perform element-wise multiplication with a variable and a tensor in PyTorch? With two tensors works fine. With a variable and a scalar works fine. But when attempting to perform element-wise multiplication with a variable and tensor I get:
XXXXXXXXXXX in mul
assert not torch.is_tensor(other)
AssertionError
For example, when running the following:
import torch
x_tensor = torch.Tensor([[1, 2], [3, 4]])
y_tensor = torch.Tensor([[5, 6], [7, 8]])
x_variable = torch.autograd.Variable(x_tensor)
print(x_tensor * y_tensor)
print(x_variable * 2)
print(x_variable * y_tensor)
I would expect the first and last print statements to show similar results. The first two multiplications work as expected, with the error coming up in the third. I have attempted the aliases of * in PyTorch (i.e. x_variable.mul(y_tensor), torch.mul(y_tensor, x_variable), etc.).
It seems that element-wise multiplication between a tensor and a variable is not supported given the error and the code which produces it. Is this correct? Or is there something I'm missing? Thank you!
| Yes, you are correct. Elementwise multiplication (like most other operations) is only supported for Tensor * Tensor or Variable * Variable, but not for Tensor * Variable.
To perform your multiplication above, wrap your Tensor as a Variable which doesn't require gradients. The additional overhead is insignificant.
y_variable = torch.autograd.Variable(y_tensor, requires_grad=False)
x_variable * y_variable # returns Variable
But obviously, only use Variables though, if you actually require automatic differentiation through a graph. Else you can just perform the operation on the Tensors directly as you did in your question.
| https://stackoverflow.com/questions/44875430/ |
Pytorch nn.functional.batch_norm for 2D input | I am currently implementing a model on which I need to change the running mean and standard deviation during test time. As such, I assume the nn.functional.batch_norm would be a better choice than the nn.BatchNorm2d
However, I have batches of images as input, and am currently not sure how to take in the images. How would I apply nn.functional.batch_norm on batches of 2D images?
The current code I have is this, I post this even though this is not correct:
mu = torch.mean(inp[0])
stddev = torch.std(inp[0])
x = nn.functional.batch_norm(inp[0], mu, stddev, training=True, momentum=0.9)
| The key is that 2D batchnorm performs the same normalization for each channel. i.e. if you have a batch of data with shape (N, C, H, W) then your mu and stddev should be shape (C,). If your images do not have a channel dimension, then add one using view.
Warning: if you set training=True then batch_norm computes and uses the appropriate normalization statistics for the argued batch (this means we don't need to calculate the mean and std ourselves). Your argued mu and stddev are supposed to be the running mean and running std for all training batches. These tensors are updated with the new batch statistics in the batch_norm function.
# inp is shape (N, C, H, W)
n_chans = inp.shape[1]
running_mu = torch.zeros(n_chans) # zeros are fine for first training iter
running_std = torch.ones(n_chans) # ones are fine for first training iter
x = nn.functional.batch_norm(inp, running_mu, running_std, training=True, momentum=0.9)
# running_mu and running_std now have new values
If you want to just use your own batch statistics, try this:
# inp is shape (N, C, H, W)
n_chans = inp.shape[1]
reshaped_inp = inp.permute(1,0,2,3).contiguous().view(n_chans, -1) # shape (C, N*W*H)
mu = reshaped_inp.mean(-1)
stddev = reshaped_inp.std(-1)
x = nn.functional.batch_norm(inp, mu, stddev, training=False)
| https://stackoverflow.com/questions/44887446/ |
requires_grad relation to leaf nodes | From the docs:
requires_grad – Boolean indicating whether the Variable has been
created by a subgraph containing any Variable, that requires it. Can
be changed only on leaf Variables
What does it mean by leaf nodes here? Are leaf nodes only the input nodes?
If it can be only changed at the leaf nodes, how can I freeze layers then?
|
Leaf nodes of a graph are those nodes (i.e. Variables) that were not computed directly from other nodes in the graph. For example:
import torch
from torch.autograd import Variable
A = Variable(torch.randn(10,10)) # this is a leaf node
B = 2 * A # this is not a leaf node
w = Variable(torch.randn(10,10)) # this is a leaf node
C = A.mm(w) # this is not a leaf node
If a leaf node requires_grad, all subsequent nodes computed from it will automatically also require_grad. Else, you could not apply the chain rule to calculate the gradient of the leaf node which requires_grad. This is the reason why requires_grad can only be set for leaf nodes: For all others, it can be smartly inferred and is in fact determined by the settings of the leaf nodes used for computing these other variables.
Note that in a typical neural network, all parameters are leaf nodes. They are not computed from any other Variables in the network. Hence, freezing layers using requires_gradis simple. Here, is an example taken from the PyTorch docs:
model = torchvision.models.resnet18(pretrained=True)
for param in model.parameters():
param.requires_grad = False
# Replace the last fully-connected layer
# Parameters of newly constructed modules have requires_grad=True by default
model.fc = nn.Linear(512, 100)
# Optimize only the classifier
optimizer = optim.SGD(model.fc.parameters(), lr=1e-2, momentum=0.9)
Even though, what you really do is freezing the entire gradient computation (which is what you should be doing as it avoids unnecessary computation). Technically, you could leave the requires_grad flag on, and only define your optimizer for a subset of the parameters that you would like to learn.
| https://stackoverflow.com/questions/44913720/ |
How two rows can be swapped in a torch tensor? | var = [[0, 1, -4, 8],
[2, -3, 2, 1],
[5, -8, 7, 1]]
var = torch.Tensor(var)
Here, var is a 3 x 4 (2d) tensor. How the first and second row can be swapped to get the following 2d tensor?
2, -3, 2, 1
0, 1, -4, 8
5, -8, 7, 1
| The other answer does not work, as some dimensions get overwritten before they are copied:
>>> var = [[0, 1, -4, 8],
[2, -3, 2, 1],
[5, -8, 7, 1]]
>>> x = torch.tensor(var)
>>> index = torch.LongTensor([1, 0, 2])
>>> x[index] = x
>>> x
tensor([[ 0, 1, -4, 8],
[ 0, 1, -4, 8],
[ 5, -8, 7, 1]])
For me, it suffices to create a new tensor (with separate underlying storage) to hold the result:
>>> x = torch.tensor(var)
>>> index = torch.LongTensor([1, 0, 2])
>>> y = torch.zeros_like(x)
>>> y[index] = x
Alternatively, you can use index_copy_ (following this explanation in discuss.pytorch.org), although I don't see an advantage for either way at the moment.
| https://stackoverflow.com/questions/44935176/ |
How padded sequences given as packed sequences are dealt by RNN in pytorch? | In pytorch, we can give a packed sequence as an input to the RNN. From official doc, input of an RNN can be as follows.
input (seq_len, batch, input_size): tensor containing the features of the input sequence. The input can also be a packed variable length sequence.
Example
packed = torch.nn.utils.rnn.pack_padded_sequence(embedded, input_lengths)
outputs, hidden = self.rnn(packed, hidden)
outputs, output_lengths = torch.nn.utils.rnn.pad_packed_sequence(outputs)
Here, embedded is the embedded representation of a batch input.
My question is, how the computation is carried out for packed sequences in RNN? How the hidden states are computed for padded sequences in a batch through packed representation?
| For the second question: hidden states at padded sequences will not be computed.
To answer how is that happening, let's first see what pack_padded_sequence does for us:
from torch.nn.utils.rnn import pad_sequence, pad_packed_sequence, pack_padded_sequence
raw = [ torch.ones(25, 300) / 2,
torch.ones(22, 300) / 2.3,
torch.ones(15, 300) / 3.2 ]
padded = pad_sequence(raw) # size: [25, 3, 300]
lengths = torch.as_tensor([25, 22, 15], dtype=torch.int64)
packed = pack_padded_sequence(padded, lengths)
so far we randomly created a three tensor with different length (timestep in the context of RNN) , and we first pad them to the same length, then packed it. Now if we run
print(padded.size())
print(packed.data.size()) # packed.data refers to the "packed" tensor
we will see:
torch.Size([25, 3, 300])
torch.Size([62, 300])
Obviously 62 does not come from 25 * 3. So what pack_padded_sequence does is only keep the meaningful timestep of each batch entry according to the lengths tensor we passed to pack_padded_sequence (i.e. if we passed [25, 25, 25] to it, the size of packed.data would still be [75, 300] even though the raw tensor does not change). In short, rnn would no even see the pad timestep with pack_padded_sequence
And now let's see what's the difference after we pass padded and packed to rnn
rnn = torch.nn.RNN(input_size=300, hidden_size=2)
padded_outp, padded_hn = rnn(padded) # size: [25, 3, 2] / [1, 3, 2]
packed_outp, packed_hn = rnn(packed) # 'PackedSequence' Obj / [1, 3, 2]
undo_packed_outp, _ = pad_packed_sequence(packed_outp)
# return "h_n"
print(padded_hn) # tensor([[[-0.2329, -0.6179], [-0.1158, -0.5430],[ 0.0998, -0.3768]]])
print(packed_hn) # tensor([[[-0.2329, -0.6179], [ 0.5622, 0.1288], [ 0.5683, 0.1327]]]
# the output of last timestep (the 25-th timestep)
print(padded_outp[-1]) # tensor([[[-0.2329, -0.6179], [-0.1158, -0.5430],[ 0.0998, -0.3768]]])
print(undo_packed_outp.data[-1]) # tensor([[-0.2329, -0.6179], [ 0.0000, 0.0000], [ 0.0000, 0.0000]]
The values of padded_hn and packed_hn are different since rnn DOES compute the pad for padded yet not for the 'packed' (PackedSequence object), which also can be observed from the last hidden state: all three batch entry in padded got non-zero last hidden state even if its length is less than 25. But for packed, the last hidden state for shorter data is not computed (i.e. 0)
p.s. another observation:
print([(undo_packed_outp[:, i, :].sum(-1) != 0).sum() for i in range(3)])
would give us [tensor(25), tensor(22), tensor(15)], which align to the actual length of our input.
| https://stackoverflow.com/questions/44942931/ |
Backpropagation without updating weights (DCGAN) | The basic idea is, that I load a trained model (DCGAN) and make a anomaly detection with it on images. For the anomaly detection I have to do some iterations on the test phase to evaluate it, if it is a anomaly or not.
For that I have two Loss-Functions in the test setup, which should be calculating a backpropagation to the generator input and update the latent vector. But it only should update the latent vector, not the weights on the graph.
Is this possible?
Probably, if I use only a pytorch-variable of my latent vector and set the variable output of the generator to
"requires_grad=False"
Like in the docs --> Pytorch
| Yes, you're on the right track there.
You can individually set the requires_grad attribute of your model parameters (more precisely of all leaf nodes in your computational graph). I am not familiar with DCGAN, but I assume the latent vector is a trainable parameter, too (else a back-propagation update makes little sense to me).
The following code snippet adapted from the PyTorch docs might be useful for your purposes:
# Load your model
model = torchvision.models.resnet18(pretrained=True)
# Disable gradient computation for all parameters
for param in model.parameters():
param.requires_grad = False
# Enable gradient computation for a specific layer (here the last fc layer)
for param in model.fc.parameters():
param.requires_grad = True
# Optimize only the the desired parameters (for you latent vectors)
optimizer = optim.SGD(model.fc.parameters(), lr=1e-2, momentum=0.9)
| https://stackoverflow.com/questions/44946190/ |
Why does an attribute has a method? | I use pytorch define a Variable x, compute and get its gradient x.grad, meaning grad is an attribute of instance x. But, I can use x.grad.data.zero_() to set x.grad to zero meaning that data.zero_() is method of x.grad. Why does an attribute has a method? Thanks a lot.
| It will be just an object attribute, for example we can have an attribute of type String. And we all now String have his own methods including the method that prints out the value of itself
| https://stackoverflow.com/questions/44951963/ |
Why RNN need two bias vectors? | In pytorch RNN implementation, there are two biases, b_ih and b_hh.
Why is this? Is it different from using one bias? If yes, how? Will it affect performance or efficiency?
| The formular in Pytorch Document in RNN is self-explained. That is b_ih and b_hh in the equation.
You may think that b_ih is bias for input (which pair with w_ih, weight for input) and b_hh is bias for hidden (pair with w_hh, weight for hidden)
| https://stackoverflow.com/questions/45005163/ |
Include .whl installation in requirements.txt | How do I include this in the requirements.txt file?
For Linux:
pip install http://download.pytorch.org/whl/cu75/torch-0.1.12.post2-cp27-none-linux_x86_64.whl
pip install torchvision
FOR MacOS:
pip install http://download.pytorch.org/whl/torch-0.1.12.post2-cp27-none-macosx_10_7_x86_64.whl
pip install torchvision
| You can use environment markers:
http://download.pytorch.org/whl/cu75/torch-0.1.12.post2-cp27-none-linux_x86_64.whl ; sys_platform == "linux"
http://download.pytorch.org/whl/cu75/torch-0.1.12.post2-cp27-none-linux_x86_64.whl ; sys_platform == "linux2"
http://download.pytorch.org/whl/torch-0.1.12.post2-cp27-none-macosx_10_7_x86_64.whl ; sys_platform == "darwin"
torchvision
(Double Linux entries: linux2 for Python 2, linux for Python 3.)
| https://stackoverflow.com/questions/45018492/ |
Understanding a simple LSTM pytorch | import torch,ipdb
import torch.autograd as autograd
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
rnn = nn.LSTM(input_size=10, hidden_size=20, num_layers=2)
input = Variable(torch.randn(5, 3, 10))
h0 = Variable(torch.randn(2, 3, 20))
c0 = Variable(torch.randn(2, 3, 20))
output, hn = rnn(input, (h0, c0))
This is the LSTM example from the docs. I don't know understand the following things:
What is output-size and why is it not specified anywhere?
Why does the input have 3 dimensions. What does 5 and 3 represent?
What are 2 and 3 in h0 and c0, what do those represent?
Edit:
import torch,ipdb
import torch.autograd as autograd
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
import torch.nn.functional as F
num_layers=3
num_hyperparams=4
batch = 1
hidden_size = 20
rnn = nn.LSTM(input_size=num_hyperparams, hidden_size=hidden_size, num_layers=num_layers)
input = Variable(torch.randn(1, batch, num_hyperparams)) # (seq_len, batch, input_size)
h0 = Variable(torch.randn(num_layers, batch, hidden_size)) # (num_layers, batch, hidden_size)
c0 = Variable(torch.randn(num_layers, batch, hidden_size))
output, hn = rnn(input, (h0, c0))
affine1 = nn.Linear(hidden_size, num_hyperparams)
ipdb.set_trace()
print output.size()
print h0.size()
*** RuntimeError: matrices expected, got 3D, 2D tensors at
| The output for the LSTM is the output for all the hidden nodes on the final layer.
hidden_size - the number of LSTM blocks per layer.
input_size - the number of input features per time-step.
num_layers - the number of hidden layers.
In total there are hidden_size * num_layers LSTM blocks.
The input dimensions are (seq_len, batch, input_size).
seq_len - the number of time steps in each input stream.
batch - the size of each batch of input sequences.
The hidden and cell dimensions are: (num_layers, batch, hidden_size)
output (seq_len, batch, hidden_size * num_directions): tensor containing the output features (h_t) from the last layer of the RNN, for each t.
So there will be hidden_size * num_directions outputs. You didn't initialise the RNN to be bidirectional so num_directions is 1. So output_size = hidden_size.
Edit: You can change the number of outputs by using a linear layer:
out_rnn, hn = rnn(input, (h0, c0))
lin = nn.Linear(hidden_size, output_size)
v1 = nn.View(seq_len*batch, hidden_size)
v2 = nn.View(seq_len, batch, output_size)
output = v2(lin(v1(out_rnn)))
Note: for this answer I assumed that we're only talking about non-bidirectional LSTMs.
Source: PyTorch docs.
| https://stackoverflow.com/questions/45022734/ |
How to simplify DataLoader for Autoencoder in Pytorch | Is there any easier way to set up the dataloader, because input and target data is the same in case of an autoencoder and to load the data during training? The DataLoader always requires two inputs.
Currently I define my dataloader like this:
X_train = rnd.random((300,100))
X_val = rnd.random((75,100))
train = data_utils.TensorDataset(torch.from_numpy(X_train).float(), torch.from_numpy(X_train).float())
val = data_utils.TensorDataset(torch.from_numpy(X_val).float(), torch.from_numpy(X_val).float())
train_loader= data_utils.DataLoader(train, batch_size=1)
val_loader = data_utils.DataLoader(val, batch_size=1)
and train like this:
for epoch in range(50):
for batch_idx, (data, target) in enumerate(train_loader):
data, target = Variable(data), Variable(target).detach()
optimizer.zero_grad()
output = model(data, x)
loss = criterion(output, target)
| Why not subclassing TensorDataset to make it compatible with unlabeled data ?
class UnlabeledTensorDataset(TensorDataset):
"""Dataset wrapping unlabeled data tensors.
Each sample will be retrieved by indexing tensors along the first
dimension.
Arguments:
data_tensor (Tensor): contains sample data.
"""
def __init__(self, data_tensor):
self.data_tensor = data_tensor
def __getitem__(self, index):
return self.data_tensor[index]
And something along these lines for training your autoencoder
X_train = rnd.random((300,100))
train = UnlabeledTensorDataset(torch.from_numpy(X_train).float())
train_loader= data_utils.DataLoader(train, batch_size=1)
for epoch in range(50):
for batch in train_loader:
data = Variable(batch)
optimizer.zero_grad()
output = model(data)
loss = criterion(output, data)
| https://stackoverflow.com/questions/45099554/ |
Indexing second dimension of Tensor using indices | I selected element in my tensor using a tensor of indices. Here the code below I use list of indices 0, 3, 2, 1 to select 11, 15, 2, 5
>>> import torch
>>> a = torch.Tensor([5,2,11, 15])
>>> torch.randperm(4)
0
3
2
1
[torch.LongTensor of size 4]
>>> i = torch.randperm(4)
>>> a[i]
11
15
2
5
[torch.FloatTensor of size 4]
Now, I have
>>> b = torch.Tensor([[5, 2, 11, 15],[5, 2, 11, 15], [5, 2, 11, 15]])
>>> b
5 2 11 15
5 2 11 15
5 2 11 15
[torch.FloatTensor of size 3x4]
Now, I want to use indices to select column 0, 3, 2, 1. In others word, I want a tensor like this
>>> b
11 15 2 5
11 15 2 5
11 15 2 5
[torch.FloatTensor of size 3x4]
| If using pytorch version v0.1.12
For this version there isnt an easy way to do this. Even though pytorch promises tensor manipulation to be exactly like numpy's, there are some capabilities that are still lacking. This is one of them.
Typically you would be able to do this relatively easily if you were working with numpy arrays. Like so.
>>> i = [2, 1, 0, 3]
>>> a = np.array([[5, 2, 11, 15],[5, 2, 11, 15], [5, 2, 11, 15]])
>>> a[:, i]
array([[11, 2, 5, 15],
[11, 2, 5, 15],
[11, 2, 5, 15]])
But the same thing with Tensors will give you an error:
>>> i = torch.LongTensor([2, 1, 0, 3])
>>> a = torch.Tensor([[5, 2, 11, 15],[5, 2, 11, 15], [5, 2, 11, 15]])
>>> a[:,i]
The error:
TypeError: indexing a tensor with an object of type torch.LongTensor. The only supported types are integers, slices, numpy scalars and torch.LongTensor or torch.ByteTensor as the only argument.
What that TypeError is telling you is, if you plan to use a LongTensor or a ByteTensor for indexing, then the only valid syntax is a[<LongTensor>] or a[<ByteTensor>]. Anything other than that will not work.
Because of this limitation, you have two options:
Option 1: Convert to numpy, permute, then back to Tensor
>>> i = [2, 1, 0, 3]
>>> a = torch.Tensor([[5, 2, 11, 15],[5, 2, 11, 15], [5, 2, 11, 15]])
>>> np_a = a.numpy()
>>> np_a = np_a[:,i]
>>> a = torch.from_numpy(np_a)
>>> a
11 2 5 15
11 2 5 15
11 2 5 15
[torch.FloatTensor of size 3x4]
Option 2: Move the dim you want to permute to 0 and then do it
you will move the dim that you are looking to permute, (in your case dim=1) to 0, perform the permutation, and move it back. Its a bit hacky, but it gets the job done.
def hacky_permute(a, i, dim):
a = torch.transpose(a, 0, dim)
a = a[i]
a = torch.transpose(a, 0, dim)
return a
And use it like so:
>>> i = torch.LongTensor([2, 1, 0, 3])
>>> a = torch.Tensor([[5, 2, 11, 15],[5, 2, 11, 15], [5, 2, 11, 15]])
>>> a = hacky_permute(a, i, dim=1)
>>> a
11 2 5 15
11 2 5 15
11 2 5 15
[torch.FloatTensor of size 3x4]
If using pytorch version v0.2.0
Direct indexing using a tensor now works in this version. ie.
>>> i = torch.LongTensor([2, 1, 0, 3])
>>> a = torch.Tensor([[5, 2, 11, 15],[5, 2, 11, 15], [5, 2, 11, 15]])
>>> a[:,i]
11 2 5 15
11 2 5 15
11 2 5 15
[torch.FloatTensor of size 3x4]
| https://stackoverflow.com/questions/45121182/ |
PyTorch RuntimeError : Gradients are not CUDA tensors |
I am getting the following error while doing seq to seq on characters and feeding to LSTM, and decoding to words using attention. The forward propagation is fine but while computing loss.backward() I am getting the following error.
RuntimeError: Gradients aren't CUDA tensors
My train() function is as followed.
def train(input_batch, input_batch_length, target_batch, target_batch_length, batch_size):
# Zero gradients of both optimizers
encoderchar_optimizer.zero_grad()
encoder_optimizer.zero_grad()
decoder_optimizer.zero_grad()
encoder_input = Variable(torch.FloatTensor(len(input_batch), batch_size, 500))
for ix , w in enumerate(input_batch):
w = w.contiguous().view(15, batch_size)
reshaped_input_length = [x[ix] for x in input_batch_length] # [15 ,.. 30 times] * 128
if USE_CUDA:
w = w.cuda()
#reshaped_input_length = Variable(torch.LongTensor(reshaped_input_length)).cuda()
hidden_all , output = encoderchar(w, reshaped_input_length)
encoder_input[ix] = output.transpose(0,1).contiguous().view(batch_size, -1)
if USE_CUDA:
encoder_input = encoder_input.cuda()
temporary_target_batch_length = [15] * batch_size
encoder_hidden_all, encoder_output = encoder(encoder_input, target_batch_length)
decoder_input = Variable(torch.LongTensor([SOS_token] * batch_size))
decoder_hidden = encoder_output
max_target_length = max(temporary_target_batch_length)
all_decoder_outputs = Variable(torch.zeros(max_target_length, batch_size, decoder.output_size))
# Move new Variables to CUDA
if USE_CUDA:
decoder_input = decoder_input.cuda()
all_decoder_outputs = all_decoder_outputs.cuda()
target_batch = target_batch.cuda()
# Run through decoder one time step at a time
for t in range(max_target_length):
decoder_output, decoder_hidden, decoder_attn = decoder(
decoder_input, decoder_hidden, encoder_hidden_all
)
all_decoder_outputs[t] = decoder_output
decoder_input = target_batch[t] # Next input is current target
if USE_CUDA:
decoder_input = decoder_input.cuda()
# Loss calculation and backpropagation
loss = masked_cross_entropy(
all_decoder_outputs.transpose(0, 1).contiguous(), # -> batch x seq
target_batch.transpose(0, 1).contiguous(), # -> batch x seq
target_batch_length
)
loss.backward()
# Clip gradient norms
ecc = torch.nn.utils.clip_grad_norm(encoderchar.parameters(), clip)
ec = torch.nn.utils.clip_grad_norm(encoder.parameters(), clip)
dc = torch.nn.utils.clip_grad_norm(decoder.parameters(), clip)
# Update parameters with optimizers
encoderchar_optimizer.step()
encoder_optimizer.step()
decoder_optimizer.step()
return loss.data[0], ec, dc
Full Stack Trace is here.
RuntimeError Traceback (most recent call last)
<ipython-input-10-9778e12ded02> in <module>()
11 data_target_batch_index= Variable(torch.LongTensor(data_target_batch_index)).transpose(0,1)
12 # Send the data for training
---> 13 loss, ar1, ar2 = train(data_input_batch_index, data_input_batch_length, data_target_batch_index, data_target_batch_length, batch_size)
14
15 # Keep track of loss
<ipython-input-8-9c71c385f8cd> in train(input_batch, input_batch_length, target_batch, target_batch_length, batch_size)
54 target_batch_length
55 )
---> 56 loss.backward()
57
58 # Clip gradient norms
/home/ubuntu/anaconda3/envs/tensorflow/lib/python3.6/site-packages/torch/autograd/variable.py in backward(self, gradient, retain_variables)
144 'or with gradient w.r.t. the variable')
145 gradient = self.data.new().resize_as_(self.data).fill_(1)
--> 146 self._execution_engine.run_backward((self,), (gradient,), retain_variables)
147
148 def register_hook(self, hook):
/home/ubuntu/anaconda3/envs/tensorflow/lib/python3.6/site-packages/torch/autograd/function.py in _do_backward(self, gradients, retain_variables)
207 def _do_backward(self, gradients, retain_variables):
208 self.retain_variables = retain_variables
--> 209 result = super(NestedIOFunction, self)._do_backward(gradients, retain_variables)
210 if not retain_variables:
211 del self._nested_output
/home/ubuntu/anaconda3/envs/tensorflow/lib/python3.6/site-packages/torch/autograd/function.py in backward(self, *gradients)
215 def backward(self, *gradients):
216 nested_gradients = _unflatten(gradients, self._nested_output)
--> 217 result = self.backward_extended(*nested_gradients)
218 return tuple(_iter_None_tensors(result))
219
/home/ubuntu/anaconda3/envs/tensorflow/lib/python3.6/site-packages/torch/nn/_functions/rnn.py in backward_extended(self, grad_output, grad_hy)
314 grad_hy,
315 grad_input,
--> 316 grad_hx)
317
318 if any(self.needs_input_grad[1:]):
/home/ubuntu/anaconda3/envs/tensorflow/lib/python3.6/site-packages/torch/backends/cudnn/rnn.py in backward_grad(fn, input, hx, weight, output, grad_output, grad_hy, grad_input, grad_hx)
371 hidden_size, dcy.size()))
372 if not dhy.is_cuda or not dy.is_cuda or (dcy is not None and not dcy.is_cuda):
--> 373 raise RuntimeError('Gradients aren\'t CUDA tensors')
374
375 check_error(cudnn.lib.cudnnRNNBackwardData(
RuntimeError: Gradients aren't CUDA tensors
any suggestions about why I am doing wrong?
| Make sure that all the objects that inherit nn.Module also call their .cuda(). Make sure to call before you pass any tensor to them. (essentially before training)
For example, (and I am guessing your encoder and decoder are such objects), do this right before you call train().
encoder = encoder.cuda()
decoder = decoder.cuda()
This ensures that all of the model's parameters are initialized in cuda memory.
Edit
In general, whenever you have this kind of error,
RuntimeError: Gradients aren't CUDA tensors
somewhere, (from your model creation, to defining inputs, to finally supplying the outputs to the loss function) you missed specifying a Variable object to be in GPU memory. You will have go through every step in your model, verifying all Variable objects to be in GPU memory.
Additionally, you dont have to call .cuda() on the outputs. Given that the inputs are in gpu's memory, all operations also takes place in gpu's memory, and so are your outputs.
| https://stackoverflow.com/questions/45206561/ |
Is torchvision.datasets.cifar.CIFAR10 a list or not? | When I try to figure out what is inside torchvision.datasets.cifar.CIFAR10, I did some simple code
trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
download=True, transform=transform)
print(trainset[1])
print(trainset[:10])
print(type(trainset))
However, I got some error when I try
print(trainset[:10])
The error info is
TypeError: Cannot handle this data type
I am wondering why I can use trainset[1], but not trainset[:10]?
|
Slicing isnt supported by CIFAR10, which is why you are getting that error. If you want the first 10 you will have to do this instead:
print([trainset[i] for i in range(10)])
More Info
The main reason why you can index an instance of CIFAR10 class is because the class implements __getitem__() function.
So, when you call trainset[i] you are essentially calling trainset.__getitem__(i)
Now, in python3, slicing expressions is also handled through __getitem__() where the slicing expression is passed to __getitem__() as a slice object.
So, trainset[2:10] is equivalent to trainset.__getitem__(slice(2, 10))
And since the two different types of objects being passed to __getitem__
are expected to do entirely different things, you have to handle them explicitly.
Unfortunately it isnt, as you can see from the __getitem__ method implementation of CIFAR10 class:
def __getitem__(self, index):
if self.train:
img, target = self.train_data[index], self.train_labels[index]
else:
img, target = self.test_data[index], self.test_labels[index]
# doing this so that it is consistent with all other datasets
# to return a PIL Image
img = Image.fromarray(img)
if self.transform is not None:
img = self.transform(img)
if self.target_transform is not None:
target = self.target_transform(target)
return img, target
| https://stackoverflow.com/questions/45225694/ |
Is there a way to use an external loss function in pytorch? | A typical skeleton of pytorch neural network has a forward() method, then we compute loss based on outputs of forward pass, and call backward() on that loss to update the gradients. What if my loss is determined externally (e.g. by running simulation in some RL environment). Can I still leverage this typical structure this way?
This might be a bit dumb as we no longer know exactly how much each element of output influences the loss, but perhaps there is some trickery I'm not aware of. Otherwise I'm not sure how neural nets can be used in combination with other RL algorithms.
Thank you!
| In this case it appears easiest to me abstract the forward pass (your policy?) from the loss computation. This is because (as you note) in most scenarios, you will need to obtain a state (from your environment), then compute an action (essentially the forward pass), then feed that action back to the environment to obtain a reward/ loss from your environment.
Of course, you could probably call your environment within the forward pass once you computed an action to then calculate the resultant loss. But why bother? It will get even more complicated (though possible) once you are are taking several steps in your environment until you obtain a reward/ loss.
I would suggest you take a look at the following RL example for an application of policy gradients within openAI gym: https://github.com/pytorch/examples/blob/master/reinforcement_learning/reinforce.py#L43
The essential ideas are:
Create a policy (as an nn.module) that takes in a state and returns
a stochastic policy
Wrap the computation of a policy and the sampling of an action from the policy into one function.
Call this function repeatedly to take steps through your environment, record actions and rewards.
Once an episode is finished, register rewards and perform only now the back-propagation and gradient updates.
While this example is specific to REINFORCE, the general idea of structuring your code is applicable to other RL algorithms. Besides, you'll find two other examples in the same repo.
Hope this helps.
| https://stackoverflow.com/questions/45258655/ |
Numpy/PyTorch method for partial tiling | I seem to recall encountering a Numpy or PyTorch method similar to numpy.tile, except that it allowed partial tiling to reach specified dimensions. So if I had
a = np.array([[5, 6, 7, 8],
[1, 2, 3, 4]])
(or, correspondingly, t = torch.Tensor(a) if this is a PyTorch method),
then
a.mystery_method((3,4))
would produce output
array([[5, 6, 7, 8],
[1, 2, 3, 4],
[5, 6, 7, 8]])
Unfortunately, I now can't find this method in the Numpy or Pytorch docs. Does it actually exist, and what is it if so?
| You can use np.resize -
M = 3 # number of rows for output
np.resize(a,(M,a.shape[1]))
Another way with np.take or simply indexing along the first axis for performance -
np.take(a,np.arange(M)%a.shape[0],axis=0) # with np.take
a[np.arange(M)%a.shape[0]] # with indexing
Runtime test -
In [91]: a = np.random.randint(0,9,(2000,4000))
In [92]: M = 3000
In [93]: %timeit np.resize(a,(M,a.shape[1]))
10 loops, best of 3: 24.1 ms per loop
In [94]: %timeit np.take(a,np.arange(M)%a.shape[0],axis=0)
100 loops, best of 3: 16.2 ms per loop
In [95]: %timeit a[np.arange(M)%a.shape[0]]
100 loops, best of 3: 15.2 ms per loop
| https://stackoverflow.com/questions/45298960/ |
pytorch: how to directly find gradient w.r.t. loss | In theano, it was very easy to get the gradient of some variable w.r.t. a given loss:
loss = f(x, w)
dl_dw = tt.grad(loss, wrt=w)
I get that pytorch goes by a different paradigm, where you'd do something like:
loss = f(x, w)
loss.backwards()
dl_dw = w.grad
The thing is I might not want to do a full backwards propagation through the graph - just along the path needed to get to w.
I know you can define Variables with requires_grad=False if you don't want to backpropagate through them. But then you have to decide that at the time of variable-creation (and the requires_grad=False property is attached to the variable, rather than the call which gets the gradient, which seems odd).
My Question is is there some way to backpropagate on demand (i.e. only backpropagate along the path needed to compute dl_dw, as you would in theano)?
| It turns out that this is reallyy easy. Just use torch.autograd.grad
Example:
import torch
import numpy as np
from torch.autograd import grad
x = torch.autograd.Variable(torch.from_numpy(np.random.randn(5, 4)))
w = torch.autograd.Variable(torch.from_numpy(np.random.randn(4, 3)), requires_grad=True)
y = torch.autograd.Variable(torch.from_numpy(np.random.randn(5, 3)))
loss = ((x.mm(w) - y)**2).sum()
(d_loss_d_w, ) = grad(loss, w)
assert np.allclose(d_loss_d_w.data.numpy(), (x.transpose(0, 1).mm(x.mm(w)-y)*2).data.numpy())
Thanks to JerryLin for answering the question here.
| https://stackoverflow.com/questions/45323240/ |
Running conv2d on tensor [batch, channel, sequence, H,W] in Pytorch | I am working on a video frame data where I am getting input data as tensor of the form [batch,channel,frame_sequence,height, weight] (let denote it by [B,C,S,H,W] for clarity. So each batch basically consists of a consecutive sequence of frame. What I basically want to do is run an encoder (consisting of several conv2d) on each frame ie each [C,H,W] and get it back as [B,C_output,S,H_output,W_output]. Now conv2d expects input as (N,C_in,H_in,W_in) form. I am wondering what's the best way to do this without messing up the order within the 5D tensor.
So far I am considering following way of thoughts:
>>> # B,C,seq,h,w
# 4,2, 5, 3,3
>>> x = Variable(torch.rand(4,2,5,3,3))
>>> x.size()
#torch.Size([4, 2, 5, 3, 3])
>>> x = x.permute(0,2,1,3,4)
>>> x.size() #expected = 4,5,2,3,3 B,seq,C,h,w
#torch.Size([4, 5, 2, 3, 3])
>>> x = x.contiguous().view(-1,2,3,3)
>>> x.size()
#torch.Size([20, 2, 3, 3])
And then run conv2d (encoder) on the updated x and reshape it. But I think It wouldn't preserve the original order of tensor. So, how can I achieve the goal?
|
What you are doing is completely fine. It will preserve the order. You can verify this by visualizing them.
I quickly built this for displaying the images stored in a 4d tensor (where dim=0 is batch) or a 5d tensor (where dim=0 is batch and dim=1 is sequence):
def custom_imshow(tensor):
if tensor.dim() == 4:
count = 1
for i in range(tensor.size(0)):
img = tensor[i].numpy()
plt.subplot(1, tensor.size(0), count)
img = img / 2 + 0.5 # unnormalize
img = np.transpose(img, (1, 2, 0))
count += 1
plt.imshow(img)
plt.axis('off')
if tensor.dim() == 5:
count = 1
for i in range(tensor.size(0)):
for j in range(tensor.size(1)):
img = tensor[i][j].numpy()
plt.subplot(tensor.size(0), tensor.size(1), count)
img = img / 2 + 0.5 # unnormalize
img = np.transpose(img, (1, 2, 0))
plt.imshow(img)
plt.axis('off')
count +=1
Lets say we use the CIFAR-10 dataset (consisting of 32x32x3 size images).
For a tensor x:
>>> x.size()
torch.Size([4, 5, 3, 32, 32])
>>> custom_imshow(x)
After doing x.view(-1, 3, 32, 32):
# x.size() -> torch.Size([4, 5, 3, 32, 32])
>>> x = x.view(-1, 3, 32, 32)
>>> x.size()
torch.Size([20, 3, 32, 32])
>>> custom_imshow(x)
And if you go back to a 5d tensor view:
# x.size() -> torch.Size([20, 3, 32, 32])
>>> x.view(4, 5, 3, 32, 32)
>>> x.size()
torch.Size([4, 5, 3, 32, 32])
>>> custom_imshow(x)
| https://stackoverflow.com/questions/45353497/ |
Replace all nonzero values by zero and all zero values by a specific value | I have a 3d tensor which contains some zero and nonzero values. I want to replace all nonzero values by zero and zero values by a specific value. How can I do that?
|
Pretty much exactly how you would do it using numpy, like so:
tensor[tensor!=0] = 0
In order to replace zeros and non-zeros, you can just chain them together. Just be sure to use a copy of the tensor, since they get modified:
def custom_replace(tensor, on_zero, on_non_zero):
# we create a copy of the original tensor,
# because of the way we are replacing them.
res = tensor.clone()
res[tensor==0] = on_zero
res[tensor!=0] = on_non_zero
return res
And use it like so:
>>>z
(0 ,.,.) =
0 1
1 3
(1 ,.,.) =
0 1
1 0
[torch.LongTensor of size 2x2x2]
>>>out = custom_replace(z, on_zero=5, on_non_zero=0)
>>>out
(0 ,.,.) =
5 0
0 0
(1 ,.,.) =
5 0
0 5
[torch.LongTensor of size 2x2x2]
| https://stackoverflow.com/questions/45384684/ |
Tensor type mismatch when moving to GPU | I'm getting the following error when trying to move my network and tensors to GPU. I've checked that the network parameters are moved to the GPU and check each batch's tensor and move them if they're not already on the GPU. But I'm still getting this issue say that there's a mismatch in the tensor types - one is a torch.cuda.FloatTensor and the other is a torch.FloatTensor? Could someone tell me what I'm doing wrong? Thanks.
My code:
class Train():
def __init__(self, network, training, address):
self.network = network
self.address = address
self.batch_size = training['batch_size']
self.iterations = training['iterations']
self.samples = training['samples']
self.data = training['data']
self.lr = training['lr']
self.noisy_lr = training['nlr']
self.cuda = training['cuda']
self.save = training['save']
self.scale = training['scale']
self.limit = training['limit']
self.replace = training['strategy']
self.optimizer = torch.optim.Adam(self.network.parameters(), lr=self.lr)
def tensor_to_Variable(self, t):
if next(self.network.parameters()).is_cuda and not t.is_cuda:
t = t.cuda()
return Variable(t)
def train(self):
if self.cuda:
self.network.cuda()
dh = DataHandler(self.data)
loss_fn = torch.nn.MSELoss()
losses = []
validate = []
val_size = 100
val_diff = 1
total_val = float(val_size * self.batch_size)
hypos = []
labels = []
# training loop
for i in range(self.iterations):
x, y = dh.get_batch(self.batch_size)
x = self.tensor_to_Variable(x)
y = self.tensor_to_Variable(y)
self.optimizer.zero_grad()
hypo = self.network(x)
loss = loss_fn(hypo, y)
loss.backward()
self.optimizer.step()
class Feedforward(nn.Module):
def __init__(self, topology):
super(Feedforward, self).__init__()
self.input_dim = topology['features']
self.num_hidden = topology['hidden_layers']
self.hidden_dim = topology['hidden_dim']
self.output_dim = topology['output_dim']
self.input_layer = nn.Linear(self.input_dim, self.hidden_dim)
self.hidden_layer = nn.Linear(self.hidden_dim, self.hidden_dim)
self.output_layer = nn.Linear(self.hidden_dim, self.output_dim)
self.dropout_layer = nn.Dropout(p=0.2)
def forward(self, x):
batch_size = x.size()[0]
feat_size = x.size()[1]
input_size = batch_size * feat_size
self.input_layer = nn.Linear(input_size, self.hidden_dim)
hidden = self.input_layer(x.view(1, input_size)).clamp(min=0)
for _ in range(self.num_hidden):
hidden = self.dropout_layer(F.relu(self.hidden_layer(hidden)))
output_size = batch_size * self.output_dim
self.output_layer = nn.Linear(self.hidden_dim, output_size)
return self.output_layer(hidden).view(output_size)
The error:
Traceback (most recent call last):
File "/media/project/train.py", line 78, in train
hypo = self.network(x)
* (torch.cuda.FloatTensor mat1, torch.cuda.FloatTensor mat2)
* (torch.cuda.sparse.FloatTensor mat1, torch.cuda.FloatTensor mat2)
* (float beta, torch.cuda.FloatTensor mat1, torch.cuda.FloatTensor mat2)
* (float alpha, torch.cuda.FloatTensor mat1, torch.cuda.FloatTensor mat2)
* (float beta, torch.cuda.sparse.FloatTensor mat1, torch.cuda.FloatTensor mat2)
* (float alpha, torch.cuda.sparse.FloatTensor mat1, torch.cuda.FloatTensor mat2)
* (float beta, float alpha, torch.cuda.FloatTensor mat1, torch.cuda.FloatTensor mat2)
didn't match because some of the arguments have invalid types: (int, int, torch.cuda.FloatTensor, torch.FloatTensor)
* (float beta, float alpha, torch.cuda.sparse.FloatTensor mat1, torch.cuda.FloatTensor mat2)
didn't match because some of the arguments have invalid types: (int, int, torch.cuda.FloatTensor, torch.FloatTensor)
Stacktrace:
Traceback (most recent call last):
File "smpl.py", line 90, in <module>
main()
File "smpl.py", line 80, in main
trainer.train()
File "/media/mpl/temp/train.py", line 82, in train
hypo = self.network(x)
File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/module.py", line 206, in call
result = self.forward(input, **kwargs)
File "model/network.py", line 35, in forward
hidden = self.input_layer(x.view(1, input_size)).clamp(min=0)
File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/module.py", line 206, in call
result = self.forward(input, *kwargs)
File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/linear.py", line 54, in forward
return self.backend.Linear()(input, self.weight, self.bias)
File "/usr/local/lib/python2.7/dist-packages/torch/nn/_functions/linear.py", line 10, in forward
output.addmm(0, 1, input, weight.t())
TypeError: addmm_ received an invalid combination of arguments - got (int, int, torch.cuda.FloatTensor, torch.FloatTensor), but expected one of: (torch.cuda.FloatTensor mat1, torch.cuda.FloatTensor mat2)
(torch.cuda.sparse.FloatTensor mat1, torch.cuda.FloatTensor mat2) (float beta, torch.cuda.FloatTensor mat1, torch.cuda.FloatTensor mat2)
(float alpha, torch.cuda.FloatTensor mat1, torch.cuda.FloatTensor mat2) (float beta, torch.cuda.sparse.FloatTensor mat1, torch.cuda.FloatTensor mat2)
(float alpha, torch.cuda.sparse.FloatTensor mat1, torch.cuda.FloatTensor mat2) (float beta, float alpha, torch.cuda.FloatTensor mat1, torch.cuda.FloatTensor mat2)
didn't match because some of the arguments have invalid types: (int, int, torch.cuda.FloatTensor, torch.FloatTensor)
* (float beta, float alpha, torch.cuda.sparse.FloatTensor mat1, torch.cuda.FloatTensor mat2)
didn't match because some of the arguments have invalid types: (int, int, torch.cuda.FloatTensor, torch.FloatTensor
| This is happening because you are re-initializing self.input_layer in your forward() function.
The call self.network.cuda() moves all of the model parameters into cuda. Which means any and all the layers you initialize at the creation of your FeedForward object will be moved to cuda memory. But when you reinitialize self.input_layer in your forward() function, you initialize that layer's parameters in cpu and not gpu. Same goes for self.output_layer.
| https://stackoverflow.com/questions/45446983/ |
TypeError: a float is required | l have a pytorch variable :
preds[4,4]
Out[305]:
Variable containing:
-96.7809
[torch.cuda.FloatTensor of size 1 (GPU 0)]
l want to do the following :
import math
x=preds[4,4]
y=maths.exp(x)
z= y / (y+1)
However when l do :
y=maths.exp(x)
l get the following error :
math.exp(preds[4,4])
TypeError: a float is required
How can l transform a torch variable to a float in order to be able to do these operations ?
Thanks
| Indexing a Variable object doesnt convert it into a scalar. Its still a Variable object. However indexing a numpy array does. So converting the Variable object into a numpy and then indexing the way you want it should do the trick.
But there are some small pitfalls when converting a Variable to numpy.
If preds is a Variable stored in cpu memory, you can simply do this.
nparr = preds.data.numpy()
x = nparr[4, 4]
However, if preds is in gpu memory, you will have to first transfer the Variable into cpu memory before you convert it into a numpy object, like so:
preds = preds.cpu()
and then do the same as above.
nparr = preds.data.numpy()
x = nparr[4, 4]
In both cases x is a scalar (a float in your case) and you can use it in any math operations you choose.
Edit:
yes, @mexmex is right, you could also directly index the tensor that is wrapped in the Variable to extract the scalar value at any given index.
Like so:
x = preds.data[4, 4]
| https://stackoverflow.com/questions/45490215/ |
Regression loss functions incorrect | I'm trying a basic averaging example, but the validation and loss don't match and the network fails to converge if I increase the training time. I'm training a network with 2 hidden layers, each 500 units wide on three integers from the range [0,9] with a learning rate of 1e-1, Adam, batch size of 1, and dropout for 3000 iterations and validate every 100 iterations. If the absolute difference between the label and the hypothesis is less than a threshold, here I set the threshold to 1, I consider that correct. Could someone let me know if this is an issue with the choice of loss function, something wrong with Pytorch, or something I'm doing. Below are some plots:
val_diff = 1
acc_diff = torch.FloatTensor([val_diff]).expand(self.batch_size)
Loop 100 times to during validation:
num_correct += torch.sum(torch.abs(val_h - val_y) < acc_diff)
Append after each validation phase:
validate.append(num_correct / total_val)
Here are some examples of the (hypothesis, and labels):
[...(-0.7043088674545288, 6.0), (-0.15691305696964264, 2.6666667461395264),
(0.2827358841896057, 3.3333332538604736)]
I tried six of the loss functions in the API that are typically used for regression:
torch.nn.L1Loss(size_average=False)
torch.nn.L1Loss()
torch.nn.MSELoss(size_average=False)
torch.nn.MSELoss()
torch.nn.SmoothL1Loss(size_average=False)
torch.nn.SmoothL1Loss()
Thanks.
Network code:
class Feedforward(nn.Module):
def __init__(self, topology):
super(Feedforward, self).__init__()
self.input_dim = topology['features']
self.num_hidden = topology['hidden_layers']
self.hidden_dim = topology['hidden_dim']
self.output_dim = topology['output_dim']
self.input_layer = nn.Linear(self.input_dim, self.hidden_dim)
self.hidden_layer = nn.Linear(self.hidden_dim, self.hidden_dim)
self.output_layer = nn.Linear(self.hidden_dim, self.output_dim)
self.dropout_layer = nn.Dropout(p=0.2)
def forward(self, x):
batch_size = x.size()[0]
feat_size = x.size()[1]
input_size = batch_size * feat_size
self.input_layer = nn.Linear(input_size, self.hidden_dim).cuda()
hidden = self.input_layer(x.view(1, input_size)).clamp(min=0)
for _ in range(self.num_hidden):
hidden = self.dropout_layer(F.relu(self.hidden_layer(hidden)))
output_size = batch_size * self.output_dim
self.output_layer = nn.Linear(self.hidden_dim, output_size).cuda()
return self.output_layer(hidden).view(output_size)
Training code:
def train(self):
if self.cuda:
self.network.cuda()
dh = DataHandler(self.data)
# loss_fn = nn.L1Loss(size_average=False)
# loss_fn = nn.L1Loss()
# loss_fn = nn.SmoothL1Loss(size_average=False)
# loss_fn = nn.SmoothL1Loss()
# loss_fn = nn.MSELoss(size_average=False)
loss_fn = torch.nn.MSELoss()
losses = []
validate = []
hypos = []
labels = []
val_size = 100
val_diff = 1
total_val = float(val_size * self.batch_size)
for i in range(self.iterations):
x, y = dh.get_batch(self.batch_size)
x = self.tensor_to_Variable(x)
y = self.tensor_to_Variable(y)
self.optimizer.zero_grad()
loss = loss_fn(self.network(x), y)
loss.backward()
self.optimizer.step()
| It looks like you've misunderstood how layers in pytorch works, here are a few tips:
In your forward when you do nn.Linear(...) you are definining new layers instead of using those you pre-defined in your network __init__. Therefore, it cannot learn anything as weights are constantly reinitalized.
You shouldn't need to call .cuda() inside net.forward(...) since you've already copied the network on gpu in your train by calling self.network.cuda()
Ideally the net.forward(...) input should directly have the shape of the first layer so you won't have to modify it. Here you should have x.size() <=> Linear -- > (Batch_size, Features).
Your forward should look close to this:
def forward(self, x):
x = F.relu(self.input_layer(x))
x = F.dropout(F.relu(self.hidden_layer(x)),training=self.training)
x = self.output_layer(x)
return x
| https://stackoverflow.com/questions/45490265/ |
Any pytorch tools to monitor neural network's training? | Are there any tools to monitor network's training in PyTorch? Like tensorboard in tensorflow.
| I am using tensorboardX. It supports most (if not all) of the features of TensorBoard. I am using the Scalar, Images, Distributions, Histograms and Text. I haven't tried the rest, like audio and graph, but the repo also contains examples for those use cases. The installation can be done easily with pip. It's all explained in the README file of the repo.
There are also other github repos which implement a wrapper for PyTorch (and other languages/frameworks) to tensorboard. As far as I know they support fewer functionalities. But have a look at:
Crayon
Tensorboard-Logger
| https://stackoverflow.com/questions/45519368/ |
Flatten layer of PyTorch build by sequential container | I am trying to build a cnn by sequential container of PyTorch, my problem is I cannot figure out how to flatten the layer.
main = nn.Sequential()
self._conv_block(main, 'conv_0', 3, 6, 5)
main.add_module('max_pool_0_2_2', nn.MaxPool2d(2,2))
self._conv_block(main, 'conv_1', 6, 16, 3)
main.add_module('max_pool_1_2_2', nn.MaxPool2d(2,2))
main.add_module('flatten', make_it_flatten)
What should I put in the "make_it_flatten"?
I tried to flatten the main but it do not work, main do not exist something call view
main = main.view(-1, 16*3*3)
| This might not be exactly what you are looking for, but you can simply create your own nn.Module that flattens any input, which you can then add to the nn.Sequential() object:
class Flatten(nn.Module):
def forward(self, x):
return x.view(x.size()[0], -1)
The x.size()[0] will select the batch dim, and -1 will compute all remaining dims to fit the number of elements, thereby flattening any tensor/Variable.
And using it in nn.Sequential:
main = nn.Sequential()
self._conv_block(main, 'conv_0', 3, 6, 5)
main.add_module('max_pool_0_2_2', nn.MaxPool2d(2,2))
self._conv_block(main, 'conv_1', 6, 16, 3)
main.add_module('max_pool_1_2_2', nn.MaxPool2d(2,2))
main.add_module('flatten', Flatten())
| https://stackoverflow.com/questions/45584907/ |
Which deep learning library support the compression of the deep learning models to be used on the phones? | I want to build an advanced deep learning model (for example: a model that uses attention) and use it on android phones (without training of course) i will only use it for inference.
and i want a library that can do that and can compress the model size to be used on the phone or android.
and do you know any projects or apps similar to my goal?
| There is a Caffe fork called Ristretto. It allows compressing neural nets for lower numerical precision (less than 32 bits per parameter), while keeping high accuracy. MXNet and Tensorflow also have this feature now. Pytorch doesn't have it yet.
These tools allow to reduce the memory required for storing the neural net parameters, but they are not specific to Android.
| https://stackoverflow.com/questions/45605087/ |
End of preview. Expand
in Dataset Viewer.
README.md exists but content is empty.
Use the Edit dataset card button to edit it.
- Downloads last month
- 43