Upload 2 files

linearizer.py

@@ -0,0 +1,221 @@
+import torch
+import numpy as np
+from torch import nn
+from torch.nn import functional as F
+from einops.layers.torch import Rearrange
+import math
+
+
+
+# helper functions
+
+def default(val, default_val):
+    return val if val is not None else default_val
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+# helper classes
+
+class Residual(nn.Module):
+    def __init__(self, fn):
+        super().__init__()
+        self.fn = fn
+    def forward(self, x):
+        return x + self.fn(x)
+
+class PreNorm(nn.Module):
+    def __init__(self, dim, fn):
+        super().__init__()
+        self.fn = fn
+        self.norm = nn.LayerNorm(dim)
+    def forward(self, x):
+        x = self.norm(x)
+        return self.fn(x)
+
+class GELU_(nn.Module):
+    def forward(self, x):
+        return 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
+
+GELU = nn.GELU if hasattr(nn, 'GELU') else GELU_
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout = 0., activation = None, glu = False):
+        super().__init__()
+        activation = default(activation, GELU)
+
+        self.glu = glu
+        self.w1 = nn.Linear(dim, hidden_dim * (2 if glu else 1))
+        self.act = activation()
+        self.dropout = nn.Dropout(dropout)
+        self.w2 = nn.Linear(hidden_dim, dim)
+
+    def forward(self, x, **kwargs):
+        if not self.glu:
+            x = self.w1(x)
+            x = self.act(x)
+        else:
+            x, v = self.w1(x).chunk(2, dim=-1)
+            x = self.act(x) * v
+
+        x = self.dropout(x)
+        x = self.w2(x)
+        return x
+
+class LinformerSelfAttention(nn.Module):
+    def __init__(self, dim, seq_len, k = 16, heads = 4, dim_head = None, one_kv_head = False, share_kv = False, dropout = 0.):
+        super().__init__()
+        
+        assert (dim % heads) == 0, 'dimension must be divisible by the number of heads'
+
+        self.seq_len = seq_len
+        self.k = k
+
+        self.heads = heads
+
+        dim_head = default(dim_head, dim // heads)
+        self.dim_head = dim_head
+
+        self.to_q = nn.Linear(dim, dim_head * heads, bias = False)
+
+        kv_dim = dim_head if one_kv_head else (dim_head * heads)
+        self.to_k = nn.Linear(dim, kv_dim, bias = False)
+        self.proj_k = nn.Parameter(init_(torch.zeros(seq_len, k)))
+
+        self.share_kv = share_kv
+        if not share_kv:
+            self.to_v = nn.Linear(dim, kv_dim, bias = False)
+            self.proj_v = nn.Parameter(init_(torch.zeros(seq_len, k)))
+
+        self.dropout = nn.Dropout(dropout)
+        self.to_out = nn.Linear(dim_head * heads, dim)
+
+    def forward(self, x, context = None, **kwargs):
+        b, n, d, d_h, h, k = *x.shape, self.dim_head, self.heads, self.k
+
+        kv_len = n if context is None else context.shape[1]
+        assert kv_len == self.seq_len, f'the sequence length of the key / values must be {self.seq_len} - {kv_len} given'
+
+        queries = self.to_q(x)
+
+        proj_seq_len = lambda args: torch.einsum('bnd,nk->bkd', *args)
+
+        kv_input = x if context is None else context
+
+        keys = self.to_k(kv_input)
+        values = self.to_v(kv_input) if not self.share_kv else keys
+
+        kv_projs = (self.proj_k, self.proj_v if not self.share_kv else self.proj_k)
+
+        # project keys and values along the sequence length dimension to k
+
+        keys, values = map(proj_seq_len, zip((keys, values), kv_projs))
+
+        # merge head into batch for queries and key / values
+
+        queries = queries.reshape(b, n, h, -1).transpose(1, 2)
+
+        merge_key_values = lambda t: t.reshape(b, k, -1, d_h).transpose(1, 2).expand(-1, h, -1, -1)
+        keys, values = map(merge_key_values, (keys, values))
+
+        # attention
+
+        dots = torch.einsum('bhnd,bhkd->bhnk', queries, keys) * (d_h ** -0.5)
+        attn = dots.softmax(dim=-1)
+        attn = self.dropout(attn)
+        out = torch.einsum('bhnk,bhkd->bhnd', attn, values)
+
+        # split heads
+        out = out.transpose(1, 2).reshape(b, n, -1)
+        return self.to_out(out)
+
+
+
+
+
+
+
+class LinformerBlock(nn.Module):
+    def __init__(self, d_model, d_ffn, seq_len,dropout):
+        super().__init__()
+       
+        self.norm = nn.LayerNorm(d_model) 
+        self.Linformer_unit = LinformerSelfAttention(d_model, seq_len, k = 256, heads = 8, dim_head = None, one_kv_head = False, share_kv = False, dropout=dropout)            
+        self.ffn = FeedForward(d_model,d_ffn,dropout)
+    def forward(self, x):
+        residual = x
+        x = self.norm(x)
+        x = self.Linformer_unit(x)   
+        x = x + residual      
+        residual = x
+        x = self.norm(x)
+        x = self.ffn(x)
+        out = x + residual
+        return out
+
+
+
+
+
+
+class LinearizerGatingUnit(nn.Module):
+    def __init__(self,d_model,d_ffn,seq_len,dropout):
+        super().__init__()
+        self.proj = nn.Linear(d_model,d_model)     
+        self.Linz = LinformerBlock(
+			   d_model, d_ffn, seq_len,dropout
+			)
+
+	
+       
+
+    def forward(self, x):
+        u, v = x, x 
+        u = self.proj(u)  
+        v = self.Linz(v)
+        out = u * v
+        return out
+
+
+class LinearizerBlock(nn.Module):
+    def __init__(self, d_model,d_ffn,seq_len,dropout):
+        super().__init__()
+       
+        self.norm = nn.LayerNorm(d_model)       
+        self.lgu = LinearizerGatingUnit(d_model,d_ffn,seq_len,dropout)
+        self.ffn = FeedForward(d_model,d_ffn,dropout)
+    def forward(self, x):
+        residual = x
+        x = self.norm(x)
+        x = self.lgu(x)   
+        x = x + residual      
+        residual = x
+        x = self.norm(x)
+        x = self.ffn(x)
+        out = x + residual
+        return out
+
+
+
+class Linearizer(nn.Module):
+    def __init__(self, d_model, d_ffn,seq_len, num_layers,dropout):
+        super().__init__()
+        
+        self.model = nn.Sequential(
+            *[LinearizerBlock(d_model,d_ffn,seq_len,dropout) for _ in range(num_layers)]
+        )
+
+    def forward(self, x):
+       
+        return self.model(x)
+
+
+
+
+
+
+
+

train.py

@@ -0,0 +1,197 @@
+#imports
+
+import os
+import csv
+import torch
+from torch import nn 
+from torch.utils.data import DataLoader 
+from torchvision import datasets 
+from torchvision.transforms import ToTensor, Normalize, RandomCrop, RandomHorizontalFlip, Compose 
+from linearizer import Linearizer
+
+# data transforms
+
+transform = Compose([
+RandomCrop(32, padding=4),
+RandomHorizontalFlip(), 
+ToTensor(),
+Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
+
+])
+
+training_data = datasets.CIFAR10(
+                                       root='data',
+                                       train=True,
+                                       download=True,
+                                       transform=transform 
+                                       )
+
+test_data = datasets.CIFAR10(
+                                       root='data',
+                                       train=False,
+                                       download=True,
+                                       transform=transform 
+                                       )                                       
+# create dataloaders  
+                                     
+batch_size = 128
+
+train_dataloader = DataLoader(training_data, batch_size=batch_size,shuffle=True)
+test_dataloader = DataLoader(test_data, batch_size=batch_size)
+
+
+for X, y in test_dataloader:
+    print(f"Shape of X [N,C,H,W]:{X.shape}")
+    print(f"Shape of y:{y.shape}{y.dtype}")
+    break
+
+
+
+# size checking for loading images
+def check_sizes(image_size, patch_size):
+    sqrt_num_patches, remainder = divmod(image_size, patch_size)
+    assert remainder == 0, "`image_size` must be divisibe by `patch_size`"
+    num_patches = sqrt_num_patches ** 2
+    return num_patches
+
+
+
+# create model
+# Get cpu or gpu device for training.
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+print(f"using {device} device") 
+
+# model definition
+
+# model definition
+
+class Linearizer_ImageClassification(Linearizer):
+    def __init__(
+        self,
+        image_size=32,
+        patch_size=4,
+        in_channels=3,
+        num_classes=10,
+        d_model=256,
+        d_ffn=512,
+        seq_len=64,
+        num_layers=4,
+        dropout=0.5
+    ):
+        num_patches = check_sizes(image_size, patch_size)
+        super().__init__(d_model, d_ffn, seq_len, num_layers,dropout)
+        self.patcher = nn.Conv2d(
+            in_channels, d_model, kernel_size=patch_size, stride=patch_size
+        )
+        self.classifier = nn.Linear(d_model, num_classes)
+
+    def forward(self, x):
+        patches = self.patcher(x)
+        batch_size, num_channels, _, _ = patches.shape
+        patches = patches.permute(0, 2, 3, 1)
+        patches = patches.view(batch_size, -1, num_channels)
+        embedding = self.model(patches)
+        embedding = embedding.mean(dim=1) # global average pooling
+        out = self.classifier(embedding)
+        return out
+
+model = Linearizer_ImageClassification().to(device)
+print(model)
+
+# Optimizer
+
+loss_fn = nn.CrossEntropyLoss()
+optimizer = torch.optim.Adam(model.parameters(),lr=1e-3)
+
+
+# Training Loop
+
+def train(dataloader, model, loss_fn, optimizer):
+    size = len(dataloader.dataset)
+    num_batches = len(dataloader)
+    model.train()
+    train_loss = 0
+    correct = 0
+    for batch, (X,y) in enumerate(dataloader):
+        X, y = X.to(device), y.to(device)
+       
+        #compute prediction error
+        pred = model(X)
+        loss = loss_fn(pred,y)
+        
+        # backpropagation
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+        train_loss += loss.item()
+        _, labels = torch.max(pred.data, 1)
+        correct += labels.eq(y.data).type(torch.float).sum()
+
+        
+
+
+        if batch % 100 == 0:
+            loss, current = loss.item(), batch * len(X)
+            print(f"loss: {loss:>7f}   [{current:>5d}/{size:>5d}]")
+
+    train_loss /= num_batches
+    train_accuracy = 100. * correct.item() / size
+    print(train_accuracy)
+    return train_loss,train_accuracy 
+
+
+
+# Test loop
+
+def test(dataloader, model, loss_fn):
+    size = len(dataloader.dataset)            
+    num_batches = len(dataloader)
+    model.eval()
+    test_loss = 0
+    correct = 0
+    with torch.no_grad():
+        for X,y in dataloader:
+            X,y = X.to(device), y.to(device)
+            pred = model(X)
+            test_loss += loss_fn(pred, y).item()
+            correct += (pred.argmax(1) == y).type(torch.float).sum().item()
+    test_loss /= num_batches
+    correct /= size
+    print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")  
+    test_accuracy = 100*correct      
+    return test_loss, test_accuracy
+
+
+
+# apply train and test
+
+logname = "/home/abdullah/Desktop/Proposal_experiments/Linearizer/Experiments_cifar10/logs_linearizer/logs_cifar10.csv"
+if not os.path.exists(logname):
+  with open(logname, 'w') as logfile:
+    logwriter = csv.writer(logfile, delimiter=',')
+    logwriter.writerow(['epoch', 'train loss', 'train acc',
+                        'test loss', 'test acc'])
+
+
+epochs = 100
+for epoch in range(epochs):
+    print(f"Epoch {epoch+1}\n-----------------------------------")
+    train_loss, train_acc = train(train_dataloader, model, loss_fn, optimizer)
+    # learning rate scheduler
+    #if scheduler is not None:
+    #    scheduler.step()
+    test_loss, test_acc = test(test_dataloader, model, loss_fn)
+    with open(logname, 'a') as logfile:
+        logwriter = csv.writer(logfile, delimiter=',')
+        logwriter.writerow([epoch+1, train_loss, train_acc,
+                            test_loss, test_acc])
+print("Done!")
+
+# saving trained model
+
+path = "/home/abdullah/Desktop/Proposal_experiments/Linearizer/Experiments_cifar10/weights_linearizer"
+model_name = "linearizerImageClassification_cifar10"
+torch.save(model.state_dict(), f"{path}/{model_name}.pth")
+print(f"Saved Model State to {path}/{model_name}.pth ")
+