Evaluation of multi-output models and loss weighting

Intermediate Deep Learning with PyTorch

Michal Oleszak

Machine Learning Engineer

Model evaluation

acc_alpha = Accuracy(
    task="multiclass", num_classes=30
)
acc_char = Accuracy(
    task="multiclass", num_classes=964
)


net.eval()
with torch.no_grad():
    for images, labels_alpha, labels_char \
    in dataloader_test:
        out_alpha, out_char = net(images)

        _, pred_alpha = torch.max(out_alpha, 1)
        _, pred_char = torch.max(out_char, 1)

        acc_alpha(pred_alpha, labels_alpha)
        acc_char(pred_char, labels_char)

Set up metric for each output
Iterate over test loader and get outputs
Calculate prediction for each output
Update accuracy metrics
Calculate final accuracy scores

print(f"Alphabet: {acc_alpha.compute()}")
print(f"Character: {acc_char.compute()}")

Alphabet: 0.3166305720806122
Character: 0.24064336717128754

Multi-output training loop revisited

for epoch in range(10):
    for images, labels_alpha, labels_char \
    in dataloader_train:
        optimizer.zero_grad()
        outputs_alpha, outputs_char = net(images)
        loss_alpha = criterion(
          outputs_alpha, labels_alpha
        )
        loss_char = criterion(
          outputs_char, labels_char
        )
        loss = loss_alpha + loss_char
        loss.backward()
        optimizer.step()

Two losses: for alphabets and characters
Final loss defined as sum of alphabet and character losses: loss = loss_alpha + loss_char
Both classification tasks deemed equally important

Varying task importance

Character classification 2 times more important than alphabet classification

Approach 1: Scale more important loss by a factor of 2
```
loss = loss_alpha + loss_char * 2
```

Approach 2: Assign weights that sum to 1

loss = 0.33 * loss_alpha + 0.67 * loss_char

Warning: losses on different scales

Losses must be on the same scale before they are weighted and added
Example tasks:
- Predict house price -> MSE loss
- Predict quality: low, medium, high -> CrossEntropy loss
CrossEntropy is typically in the single-digits
MSE loss can reach tens of thousands
Model would ignore quality assessment task

Solution: Normalize both losses before weighting and adding

loss_price = loss_price / torch.max(loss_price)
loss_quality = loss_quality / torch.max(loss_quality)
loss = 0.7 * loss_price + 0.3 * loss_quality

Let's practice!

Intermediate Deep Learning with PyTorch