Replacing classifier head in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - Replacing classifier head

Which metric matters for Replacing classifier head and WHY

When you replace a classifier head in a model, you want to check how well the new head predicts the correct classes. The key metrics are accuracy to see overall correctness, and precision and recall to understand how well it finds true positives without too many mistakes. This helps you know if the new head is learning properly and making good predictions.

Confusion matrix example

    Confusion Matrix (for 100 samples):

          Predicted
          Pos   Neg
    Actual
    Pos   40    10
    Neg   5     45

    TP = 40, FP = 5, TN = 45, FN = 10

    Precision = TP / (TP + FP) = 40 / (40 + 5) = 0.89
    Recall = TP / (TP + FN) = 40 / (40 + 10) = 0.80
    Accuracy = (TP + TN) / Total = (40 + 45) / 100 = 0.85
    F1 Score = 2 * (Precision * Recall) / (Precision + Recall) ≈ 0.84
    

Precision vs Recall tradeoff with examples

If your new classifier head is used for spam detection, high precision is important. You want to avoid marking good emails as spam (false positives). So, fewer false alarms matter more.

If it is used for medical diagnosis, like cancer detection, high recall is critical. You want to catch as many true cases as possible, even if some false alarms happen.

Replacing the classifier head can change this balance. You must check which metric fits your goal and tune the head accordingly.

What good vs bad metric values look like

Good: Accuracy above 80%, precision and recall both above 75%, and balanced F1 score. This means the new head predicts well and finds most true cases without many mistakes.

Bad: Accuracy below 60%, or very low precision (e.g., 30%) or recall (e.g., 20%). This shows the new head is not learning well or is biased, missing many true cases or making many wrong predictions.

Common pitfalls when evaluating replaced classifier head

• Accuracy paradox: High accuracy can be misleading if classes are imbalanced. For example, if 90% of data is one class, predicting that class always gives 90% accuracy but poor real performance.
• Data leakage: If test data leaks into training, metrics look too good but model fails in real use.
• Overfitting: New head may memorize training data but perform poorly on new data. Watch for big gaps between training and validation metrics.
• Ignoring class balance: Metrics like precision and recall per class matter more than overall accuracy when classes differ in size.

Self-check question

Your model with the replaced classifier head has 98% accuracy but only 12% recall on the positive class (e.g., fraud). Is it good for production? Why or why not?

Answer: No, it is not good. The low recall means the model misses most positive cases (fraud). Even with high accuracy, it fails to find the important cases. For fraud detection, high recall is critical to catch frauds.

Key Result

Replacing the classifier head requires checking accuracy, precision, and recall to ensure balanced and meaningful performance.

Practice

1. What is the main reason to replace the classifier head in a pretrained PyTorch model?

easy

A. To adapt the model to a new task with different output classes

B. To speed up the training by removing layers

C. To reduce the model size by deleting layers

D. To change the input image size the model accepts

Solution

1. Step 1: Understand the classifier head role

   The classifier head is the last layer that decides the output classes based on learned features.

2. Step 2: Reason about adapting to new tasks

   Replacing the classifier head allows the model to output predictions for new classes different from the original training.

3. Final Answer:

   To adapt the model to a new task with different output classes -> Option A

4. Quick Check:

   Classifier head replacement = new task adaptation [OK]

Hint: Classifier head controls output classes, replace for new tasks [OK]

Common Mistakes:

• Thinking replacing head changes input size
• Assuming it reduces model size significantly
• Believing it speeds up training by removing layers

2. Which of the following is the correct way to replace the classifier head of a pretrained ResNet model in PyTorch for 10 output classes?

easy

A. model.fc = nn.Linear(2048, 10)

B. model.classifier = nn.Linear(2048, 10)

C. model.fc = nn.Linear(512, 10)

D. model.head = nn.Linear(512, 10)

Solution

1. Step 1: Identify ResNet classifier attribute

   ResNet models use model.fc as the classifier head.

2. Step 2: Check input feature size for ResNet

   ResNet50 and similar have 2048 features before the classifier, so input size is 2048.

3. Final Answer:

   model.fc = nn.Linear(2048, 10) -> Option A

4. Quick Check:

   ResNet classifier = model.fc with 2048 input features [OK]

Hint: ResNet classifier is model.fc with 2048 input features [OK]

Common Mistakes:

• Using wrong attribute like model.classifier or model.head
• Using wrong input size like 512 instead of 2048
• Confusing ResNet with other models like VGG

3. Given the code below, what will be the output shape of the model's final layer after replacement?

import torch
import torch.nn as nn
from torchvision import models

model = models.resnet18(pretrained=True)
model.fc = nn.Linear(512, 5)

input_tensor = torch.randn(1, 3, 224, 224)
output = model(input_tensor)
print(output.shape)

medium

A. torch.Size([1, 1000])

B. torch.Size([1, 512])

C. torch.Size([1, 5])

D. torch.Size([3, 224, 224])

Solution

1. Step 1: Understand the replaced classifier output size

   The new classifier layer outputs 5 values per input (5 classes).

2. Step 2: Check input batch size and output shape

   Input batch size is 1, so output shape is (1, 5).

3. Final Answer:

   torch.Size([1, 5]) -> Option C

4. Quick Check:

   Output shape = (batch_size, output_classes) = (1, 5) [OK]

Hint: Output shape matches batch size and new class count [OK]

Common Mistakes:

• Expecting original 1000 classes output
• Confusing feature size with output size
• Misreading input tensor shape as output

4. You tried replacing the classifier head of a pretrained model with model.fc = nn.Linear(1024, 10) but got a runtime error during training. What is the most likely cause?

medium

A. The model.fc attribute does not exist in pretrained models

B. The output size 10 is too large for the model

C. You forgot to call model.eval() before training

D. The input feature size 1024 does not match the model's actual output features

Solution

1. Step 1: Check input feature size for classifier

   The input size to the new Linear layer must match the output features of the previous layer.

2. Step 2: Identify mismatch causing runtime error

   If 1024 is incorrect, the model will raise size mismatch errors during forward pass.

3. Final Answer:

   The input feature size 1024 does not match the model's actual output features -> Option D

4. Quick Check:

   Input size mismatch causes runtime error [OK]

Hint: Match Linear input size to previous layer output features [OK]

Common Mistakes:

• Assuming output size causes error
• Confusing eval mode with training errors
• Thinking model.fc is missing in pretrained models

5. You want to fine-tune a pretrained ResNet50 on a dataset with 15 classes. Which code snippet correctly replaces the classifier head and freezes all layers except the new head?

hard

A. model = models.resnet50(pretrained=True)
model.fc = nn.Linear(2048, 15)
for param in model.parameters():
param.requires_grad = False

B. model = models.resnet50(pretrained=True)
for param in model.parameters():
param.requires_grad = False
model.fc = nn.Linear(2048, 15)

C. model = models.resnet50(pretrained=True)
for param in model.fc.parameters():
param.requires_grad = False
model.fc = nn.Linear(2048, 15)

D. model = models.resnet50(pretrained=True)
model.fc = nn.Linear(512, 15)
for param in model.parameters():
param.requires_grad = True

Solution

1. Step 1: Freeze all existing model parameters

   Set param.requires_grad = False for all parameters to prevent updates during training.

2. Step 2: Replace classifier head with correct input/output sizes

   ResNet50's classifier input size is 2048; output size is 15 for new classes.

3. Step 3: Ensure new head parameters are trainable

   By replacing model.fc after freezing, new layer parameters default to requires_grad=True.

4. Final Answer:

   Freeze all params, then replace head with nn.Linear(2048, 15) -> Option B

5. Quick Check:

   Freeze old layers, replace head with correct sizes [OK]

Hint: Freeze before replacing head to keep new layer trainable [OK]

Common Mistakes:

• Freezing after replacing head disables new layer training
• Using wrong input size 512 instead of 2048
• Not freezing any layers when fine-tuning