Forward pass computation in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - Forward pass computation

Which metric matters for Forward Pass Computation and WHY

The forward pass is the step where the model makes predictions by passing input data through its layers. The key metric here is loss, which measures how far the model's predictions are from the true answers. A lower loss means the model is doing better at predicting. Accuracy can also be used to see how many predictions are correct, but loss gives a more detailed view of performance during the forward pass.

Confusion Matrix Example

For classification tasks, the confusion matrix shows how well the forward pass predictions match the true labels. Here is an example with 100 samples:

      | Predicted Positive | Predicted Negative |
      |--------------------|--------------------|
      | True Positive (TP): 40 | False Positive (FP): 5 |
      | False Negative (FN): 10 | True Negative (TN): 45 |
    

This matrix helps calculate precision, recall, and accuracy, which reflect the quality of the forward pass predictions.

Precision vs Recall Tradeoff in Forward Pass

During the forward pass, the model outputs predictions that can be tuned to favor precision or recall depending on the task:

• Precision is important when false alarms are costly. For example, in spam detection, you want to avoid marking good emails as spam.
• Recall is important when missing positive cases is costly. For example, in disease detection, you want to catch as many sick patients as possible.

The forward pass outputs probabilities that can be thresholded to balance precision and recall.

Good vs Bad Metric Values for Forward Pass

Good forward pass results have:

• Low loss values (close to 0)
• High accuracy (close to 1 or 100%)
• Balanced precision and recall depending on the task

Bad forward pass results show:

• High loss values (far from 0)
• Low accuracy (close to random guessing)
• Very low precision or recall, indicating poor prediction quality

Common Pitfalls in Forward Pass Metrics

• Accuracy paradox: High accuracy can be misleading if classes are imbalanced.
• Data leakage: If test data leaks into training, forward pass metrics look unrealistically good.
• Overfitting indicators: Very low training loss but high validation loss means the model memorizes training data but fails on new data.

Self Check

Your model has 98% accuracy but only 12% recall on fraud cases. Is it good for production?

Answer: No. Even though accuracy is high, the model misses most fraud cases (low recall). This means many frauds go undetected, which is risky. For fraud detection, recall is critical to catch as many frauds as possible.

Key Result

Loss is the key metric during forward pass to measure prediction quality; accuracy, precision, and recall help interpret results depending on the task.

Practice

1. What does the forward pass in a PyTorch model do?

easy

A. It moves input data through the model to get predictions.

B. It updates the model weights during training.

C. It saves the model to disk.

D. It initializes the model parameters.

Solution

1. Step 1: Understand the role of forward pass

   The forward pass takes input data and processes it through the model layers to produce output predictions.

2. Step 2: Differentiate from other processes

   Updating weights happens in backward pass, saving is separate, and initialization is before training.

3. Final Answer:

   It moves input data through the model to get predictions. -> Option A

4. Quick Check:

   Forward pass = data to predictions [OK]

Hint: Forward pass = input to output predictions [OK]

Common Mistakes:

• Confusing forward pass with weight update
• Thinking forward pass saves the model
• Mixing initialization with forward pass

2. Which of the following is the correct way to run a forward pass on a PyTorch model named model with input tensor x?

easy

A. output = model.forward(x)

B. output = model.predict(x)

C. output = model.run(x)

D. output = model(x)

Solution

1. Step 1: Recall PyTorch model call syntax

   In PyTorch, calling model(x) runs the forward pass internally.

2. Step 2: Identify incorrect options

   Directly calling model.forward(x) is discouraged; model.run and model.predict are not standard methods.

3. Final Answer:

   output = model(x) -> Option D

4. Quick Check:

   Use model(x) to run forward pass [OK]

Hint: Use model(input) to run forward pass [OK]

Common Mistakes:

• Calling model.forward(x) directly
• Using non-existent methods like model.run or model.predict
• Confusing syntax with other frameworks

3. Given the following PyTorch code, what is the output tensor shape?

import torch
import torch.nn as nn

class SimpleNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 2)
    def forward(self, x):
        return self.linear(x)

model = SimpleNet()
x = torch.randn(3, 4)
output = model(x)
print(output.shape)

medium

A. torch.Size([3, 2])

B. torch.Size([2, 3])

C. torch.Size([4, 2])

D. torch.Size([3, 4])

Solution

1. Step 1: Analyze input and layer dimensions

   The input tensor has shape (3, 4), meaning batch size 3 and feature size 4. The linear layer maps 4 features to 2 outputs.

2. Step 2: Compute output shape

   Applying linear layer to each batch item changes feature size from 4 to 2, so output shape is (3, 2).

3. Final Answer:

   torch.Size([3, 2]) -> Option A

4. Quick Check:

   Input batch 3, features 4 -> output features 2 [OK]

Hint: Output shape = (batch, output features) [OK]

Common Mistakes:

• Mixing batch and feature dimensions
• Assuming output shape equals input shape
• Confusing input and output feature sizes

4. Identify the error in this PyTorch forward pass code snippet:

import torch
import torch.nn as nn

class Net(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc = nn.Linear(5, 3)
    def forward(self, x):
        return self.fc(x)

model = Net()
x = torch.randn(4, 4)
output = model(x)
print(output)

medium

A. Missing activation function after Linear layer.

B. Model is not moved to GPU before forward pass.

C. Input tensor shape does not match Linear layer input size.

D. Forward method is missing return statement.

Solution

1. Step 1: Check input tensor shape vs Linear layer input

   The Linear layer expects input features of size 5, but input tensor has shape (4, 4), so feature size is 4, which mismatches.

2. Step 2: Confirm other options

   GPU usage is optional, activation functions are not mandatory, and forward method correctly returns output.

3. Final Answer:

   Input tensor shape does not match Linear layer input size. -> Option C

4. Quick Check:

   Input features must match Linear input size [OK]

Hint: Check input feature size matches Linear layer input [OK]

Common Mistakes:

• Ignoring input size mismatch errors
• Assuming GPU is required for forward pass
• Thinking activation is always needed

5. You have a PyTorch model with two linear layers chained in the forward pass:

class ChainNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(3, 4)
        self.fc2 = nn.Linear(4, 2)
    def forward(self, x):
        x = self.fc1(x)
        x = torch.relu(x)
        x = self.fc2(x)
        return x

model = ChainNet()
x = torch.tensor([[1.0, 2.0, 3.0]])
output = model(x)
print(output)

What is the shape of output and why?

hard

A. torch.Size([1, 4]) because the first layer outputs 4 features.

B. torch.Size([1, 2]) because input batch size is 1 and final layer outputs 2 features.

C. torch.Size([3, 2]) because input has 3 features and output has 2 features.

D. torch.Size([4, 2]) because the two layers output sizes multiply.

Solution

1. Step 1: Analyze input and layer sizes

   Input tensor shape is (1, 3) -- batch size 1, feature size 3. First layer maps 3 to 4 features, second layer maps 4 to 2 features.

2. Step 2: Determine output shape after forward pass

   After passing through both layers, output shape is (1, 2) because batch size stays the same and final layer outputs 2 features.

3. Final Answer:

   torch.Size([1, 2]) because input batch size is 1 and final layer outputs 2 features. -> Option B

4. Quick Check:

   Batch size unchanged, output features = final layer output [OK]

Hint: Output shape = (batch size, last layer output features) [OK]

Common Mistakes:

• Multiplying output sizes instead of chaining
• Confusing batch size with feature size
• Ignoring activation does not change shape