File size: 5,934 Bytes
0bbec58
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8e35bc7
0bbec58
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5993d2f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8e35bc7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0bbec58
 
8e35bc7
 
 
 
 
 
0bbec58
 
 
 
8e35bc7
 
 
 
 
 
 
 
 
 
 
 
 
5993d2f
 
056ab4f
0bbec58
 
5993d2f
 
0bbec58
5993d2f
0bbec58
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
#!/usr/bin/env python
# coding: utf-8

import torch
from torch import nn
import torch.nn.functional as F
from datasets import load_dataset
import fastcore.all as fc
import matplotlib.pyplot as plt
import matplotlib as mpl
import torchvision.transforms.functional as TF
from torch.utils.data import default_collate, DataLoader
import torch.optim as optim
import pickle
get_ipython().run_line_magic('matplotlib', 'inline')
plt.rcParams['figure.figsize'] = [2, 2]


dataset_nm = 'mnist'
x,y = 'image', 'label'
ds = load_dataset(dataset_nm)


def transform_ds(b):
    b[x] = [TF.to_tensor(ele) for ele in b[x]]
    return b

dst = ds.with_transform(transform_ds)
plt.imshow(dst['train'][0]['image'].permute(1,2,0));


bs = 1024
class DataLoaders:
    def __init__(self, train_ds, valid_ds, bs, collate_fn, **kwargs):
        self.train = DataLoader(train_ds, batch_size=bs, shuffle=True, collate_fn=collate_fn, **kwargs)
        self.valid = DataLoader(valid_ds, batch_size=bs*2, shuffle=False, collate_fn=collate_fn, **kwargs)

def collate_fn(b):
    collate = default_collate(b)
    return (collate[x], collate[y])

dls = DataLoaders(dst['train'], dst['test'], bs=bs, collate_fn=collate_fn)
xb,yb = next(iter(dls.train))
xb.shape, yb.shape


class Reshape(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.dim = dim
    
    def forward(self, x):
        return x.reshape(self.dim)


# model definition
def linear_classifier():
    return nn.Sequential(
        Reshape((-1, 784)),
        nn.Linear(784, 50),
        nn.ReLU(),
        nn.Linear(50, 50),
        nn.ReLU(),
        nn.Linear(50, 10)
    )


model = linear_classifier()
lr = 0.1
max_lr = 0.1
epochs = 5
opt = optim.AdamW(model.parameters(), lr=lr)
sched = optim.lr_scheduler.OneCycleLR(opt, max_lr, total_steps=len(dls.train), epochs=epochs)

for epoch in range(epochs):
    for train in (True, False):
        accuracy = 0
        dl = dls.train if train else dls.valid
        for xb,yb in dl:
            preds = model(xb)
            loss = F.cross_entropy(preds, yb)
            if train:
                loss.backward()
                opt.step()
                opt.zero_grad()
            with torch.no_grad():
                accuracy += (preds.argmax(1).detach().cpu() == yb).float().mean()
        if train:
            sched.step()
        accuracy /= len(dl)
        print(f"{'train' if train else 'eval'}, epoch:{epoch+1}, loss: {loss.item():.4f}, accuracy: {accuracy:.4f}")


# def _conv_block(ni, nf, stride, act=act_gr, norm=None, ks=3):
#     return nn.Sequential(conv(ni, nf, stride=1, act=act, norm=norm, ks=ks),
#                          conv(nf, nf, stride=stride, act=None, norm=norm, ks=ks))

# class ResBlock(nn.Module):
#     def __init__(self, ni, nf, stride=1, ks=3, act=act_gr, norm=None):
#         super().__init__()
#         self.convs = _conv_block(ni, nf, stride, act=act, ks=ks, norm=norm)
#         self.idconv = fc.noop if ni==nf else conv(ni, nf, ks=1, stride=1, act=None)
#         self.pool = fc.noop if stride==1 else nn.AvgPool2d(2, ceil_mode=True)
#         self.act = act()

#     def forward(self, x): return self.act(self.convs(x) + self.idconv(self.pool(x)))


def conv(ni, nf, ks=3, s=2, act=nn.ReLU, norm=None):
    layers = [nn.Conv2d(ni, nf, kernel_size=ks, stride=s, padding=ks//2)]
    if norm:
        layers.append(norm)
    if act:
        layers.append(act())
    return nn.Sequential(*layers)

def _conv_block(ni, nf, ks=3, s=2, act=nn.ReLU, norm=None):
    return nn.Sequential(
        conv(ni, nf, ks=ks, s=1, norm=norm, act=act),
        conv(nf, nf, ks=ks, s=s, norm=norm, act=act),
    )

class ResBlock(nn.Module):
    def __init__(self, ni, nf, s=2, ks=3, act=nn.ReLU, norm=None):
        super().__init__()
        self.convs = _conv_block(ni, nf, s=s, ks=ks, act=act, norm=norm)
        self.idconv = fc.noop if ni==nf else conv(ni, nf, ks=1, s=1, act=None)
        self.pool = fc.noop if s==1 else nn.AvgPool2d(2, ceil_mode=True)
        self.act = act()
    
    def forward(self, x):
        return self.act(self.convs(x) + self.idconv(self.pool(x)))


def cnn_classifier():
    return nn.Sequential(
        ResBlock(1, 8, norm=nn.BatchNorm2d(8)),
        ResBlock(8, 16, norm=nn.BatchNorm2d(16)),
        ResBlock(16, 32, norm=nn.BatchNorm2d(32)),
        ResBlock(32, 64, norm=nn.BatchNorm2d(64)),
        ResBlock(64, 64, norm=nn.BatchNorm2d(64)),
        conv(64, 10, act=False),
        nn.Flatten(),
    )


# def cnn_classifier():
#     return nn.Sequential(
#         ResBlock(1, 16, norm=nn.BatchNorm2d(16)),
#         ResBlock(16, 32, norm=nn.BatchNorm2d(32)),
#         ResBlock(32, 64, norm=nn.BatchNorm2d(64)),
#         ResBlock(64, 128, norm=nn.BatchNorm2d(128)),
#         ResBlock(128, 256, norm=nn.BatchNorm2d(256)),
#         ResBlock(256, 256, norm=nn.BatchNorm2d(256)),
#         conv(256, 10, act=False),
#         nn.Flatten(),
#     )


def kaiming_init(m):
    if isinstance(m, (nn.Conv1d, nn.Conv2d, nn.Conv3d)):
        nn.init.kaiming_normal_(m.weight)        


model = cnn_classifier()
model.apply(kaiming_init)
lr = 0.1
max_lr = 0.3
epochs = 5
opt = optim.AdamW(model.parameters(), lr=lr)
sched = optim.lr_scheduler.OneCycleLR(opt, max_lr, total_steps=len(dls.train), epochs=epochs)
for epoch in range(epochs):
    for train in (True, False):
        accuracy = 0
        dl = dls.train if train else dls.valid
        for xb,yb in dl:
            preds = model(xb)
            loss = F.cross_entropy(preds, yb)
            if train:
                loss.backward()
                opt.step()
                opt.zero_grad()
            with torch.no_grad():
                accuracy += (preds.argmax(1).detach().cpu() == yb).float().mean()
        if train:
            sched.step()
        accuracy /= len(dl)
        print(f"{'train' if train else 'eval'}, epoch:{epoch+1}, loss: {loss.item():.4f}, accuracy: {accuracy:.4f}")