Pytorch redundant softmax Medium

1def pytorch_redundant_softmax_noncompliant():
2    import torch
3    from torch import nn
4    from torch.utils.data import DataLoader
5    from torchvision import datasets
6    from torchvision.transforms import ToTensor
7
8    training_data = datasets.FashionMNIST(
9        root="data",
10        train=True,
11        download=True,
12        transform=ToTensor()
13    )
14
15    test_data = datasets.FashionMNIST(
16        root="data",
17        train=False,
18        download=True,
19        transform=ToTensor()
20    )
21
22    train_dataloader = DataLoader(training_data, batch_size=64)
23    test_dataloader = DataLoader(test_data, batch_size=64)
24
25    class NeuralNetwork(nn.Module):
26        def __init__(self):
27            super().__init__()
28            self.flatten = nn.Flatten()
29            self.linear_relu_stack = nn.Sequential(
30                nn.Linear(28 * 28, 512),
31                nn.ReLU(),
32                nn.Linear(512, 512),
33                nn.ReLU(),
34                nn.Linear(512, 10)
35            )
36
37        def forward(self, x):
38            x = self.flatten(x)
39            logits = self.linear_relu_stack(x)
40            # Noncompliant: Softmax used with CrossEntropyLoss.
41            logits = nn.functional.softmax(logits)
42            return logits
43
44    model = NeuralNetwork()
45
46    def train_loop(dataloader, model, loss_fn, optimizer):
47        size = len(dataloader.dataset)
48        for batch, (x, y) in enumerate(dataloader):
49            # Compute prediction and loss
50            pred = model(x)
51            loss = loss_fn(pred, y)
52
53            # Backpropagation
54            optimizer.zero_grad()
55            loss.backward()
56            optimizer.step()
57
58            if batch % 100 == 0:
59                loss, current = loss.item(), (batch + 1) * len(x)
60                print(f"loss: {loss:>7f}  [{current:>5d}/{size:>5d}]")
61
62    def test_loop(dataloader, model, loss_fn):
63        size = len(dataloader.dataset)
64        num_batches = len(dataloader)
65        test_loss, correct = 0, 0
66
67        with torch.no_grad():
68            for x, y in dataloader:
69                pred = model(x)
70                test_loss += loss_fn(pred, y).item()
71                correct += (pred.argmax(1) == y).type(torch.float).sum().item()
72
73        test_loss /= num_batches
74        correct /= size
75        print(f"Test Error: \n Accuracy: {(100 * correct):>0.1f}%, "
76              f"Avg loss: {test_loss:>8f} \n")
77
78    loss_fn = nn.CrossEntropyLoss()
79    learning_rate = 0.05
80    optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
81
82    epochs = 10
83    for t in range(epochs):
84        print(f"Model - Epoch {t + 1}\n-------------------------------")
85        train_loop(train_dataloader, model, loss_fn, optimizer)
86        test_loop(test_dataloader, model, loss_fn)
87
88    print("Done!")

1def pytorch_redundant_softmax_compliant():
2    import torch
3    import torch.nn as nn
4    from transformers import BertModel, BertForSequenceClassification
5    import default_constants
6
7    class BERT(nn.Module):
8        def __init__(self, tokenizer, bert_variant=default_constants.BERT):
9            super(BERT, self).__init__()
10
11            self.num_labels = default_constants.num_labels
12            self.hidden_dim = default_constants.hidden_dim
13            self.dropout_prob = default_constants.dropout_prob
14
15            self.bert = BertModel.from_pretrained(bert_variant)
16            self.bert.resize_token_embeddings(len(tokenizer))
17
18            self.dropout = nn.Dropout(self.dropout_prob).cuda()
19            self.classifier = nn.Linear(
20                self.hidden_dim, self.num_labels).cuda()
21            torch.nn.init.xavier_uniform(self.classifier.weight)
22
23        def forward(
24                self,
25                text,
26                labels,
27                attention_mask=None,
28                token_type_ids=None
29        ):
30            outputs = self.bert(
31                text,
32                attention_mask=attention_mask,
33                token_type_ids=token_type_ids
34            )
35
36            pooled_output = outputs[1]
37
38            pooled_output = self.dropout(pooled_output)
39            logits = self.classifier(pooled_output)
40
41            # Compliant: Softmax is not used with CrossEntropyLoss.
42            loss_fct = nn.CrossEntropyLoss(weight=torch.Tensor(
43                default_constants.class_weight)).cuda(default_constants.DEVICE)
44            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
45
46            return (
47                loss,
48                logits,
49            )