Feature extraction strategy in PyTorch - Model Pipeline Trace

Model Pipeline - Feature extraction strategy

This pipeline shows how a model uses feature extraction to learn from images. It takes raw images, extracts important features using a pre-trained network, then trains a simple classifier on these features to recognize new images.

Data Flow - 5 Stages

1Raw Image Input

1000 images x 3 channels x 64 height x 64 width→Collect raw color images→1000 images x 3 channels x 64 height x 64 width

An image of a cat represented as 3 color channels with 64x64 pixels

↓

2Preprocessing

1000 images x 3 channels x 64 height x 64 width→Normalize pixel values to range 0-1→1000 images x 3 channels x 64 height x 64 width

Pixel values scaled from 0-255 to 0.0-1.0

↓

3Feature Extraction

1000 images x 3 channels x 64 height x 64 width→Pass images through pre-trained CNN (e.g., ResNet18) without final layer→1000 images x 512 features

Each image converted to a 512-length feature vector summarizing important patterns

↓

4Classifier Training

800 images x 512 features→Train a simple linear layer on extracted features→800 images x 10 classes

Model learns to map features to one of 10 categories

↓

5Validation

200 images x 512 features→Evaluate classifier on unseen features→200 images x 10 classes

Model predicts class probabilities for new images

Training Trace - Epoch by Epoch

Loss
2.0 |****
1.5 |*** 
1.0 |**  
0.5 |*   
0.0 +----
      1 2 3 4 5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.8	0.40	Model starts learning, accuracy is low
2	1.2	0.60	Loss decreases, accuracy improves
3	0.9	0.72	Model learns better features for classification
4	0.7	0.80	Training converges, accuracy stabilizes
5	0.6	0.83	Final epoch with good accuracy

Prediction Trace - 4 Layers

Layer 1: Input Image

Layer 2: Feature Extractor (Pre-trained CNN)

Layer 3: Classifier Linear Layer

Layer 4: Softmax Activation

Model Quiz - 3 Questions

What is the main purpose of the feature extraction step?

A To randomly shuffle image pixels

B To increase the image size for better training

C To convert images into a smaller set of meaningful features

D To directly predict the class labels

Key Insight

Feature extraction uses a pre-trained network to turn complex images into simpler feature vectors. Training a small classifier on these features is faster and effective, showing how reusing learned knowledge helps new tasks.

Practice

1. What is the main purpose of using a pre-trained model for feature extraction in PyTorch?

easy

A. To replace the optimizer with a new one

B. To use learned features from a large dataset and avoid training from scratch

C. To train all layers from random weights

D. To increase the size of the dataset automatically

Solution

1. Step 1: Understand feature extraction concept

   Feature extraction uses a model already trained on a large dataset to get useful features without training all layers again.

2. Step 2: Identify the main benefit

   This saves time and resources by reusing learned knowledge instead of starting from scratch.

3. Final Answer:

   To use learned features from a large dataset and avoid training from scratch -> Option B

4. Quick Check:

   Feature extraction = reuse learned features [OK]

Hint: Pre-trained means reuse, not retrain all layers [OK]

Common Mistakes:

• Thinking feature extraction means training all layers
• Confusing feature extraction with data augmentation
• Believing optimizer changes are part of feature extraction

2. Which PyTorch code snippet correctly freezes all layers of a pre-trained model except the final layer?

easy

A. for param in model.parameters(): param.requires_grad = True
model.fc = nn.Linear(512, 10)

B. model.fc.requires_grad = False
for param in model.parameters(): param.requires_grad = True

C. for param in model.parameters(): param.requires_grad = False
model.fc = nn.Linear(512, 10)

D. model.fc = nn.Linear(512, 10)
for param in model.parameters(): param.requires_grad = False

Solution

1. Step 1: Freeze all layers by setting requires_grad to false

   The loop disables gradient updates for all parameters to keep pre-trained weights fixed.

2. Step 2: Replace the final layer with a new one to train

   Assigning a new linear layer to model.fc allows training only this layer for the new task.

3. Final Answer:

   for param in model.parameters(): param.requires_grad = False model.fc = nn.Linear(512, 10) -> Option C

4. Quick Check:

   Freeze all except final layer = for param in model.parameters(): param.requires_grad = False model.fc = nn.Linear(512, 10) [OK]

Hint: Freeze first, then replace final layer [OK]

Common Mistakes:

• Not freezing layers before replacing final layer
• Freezing final layer instead of others
• Setting requires_grad true for all parameters

3. Given this PyTorch code for feature extraction, what will be the output shape of features?

import torch
import torchvision.models as models
model = models.resnet18(pretrained=True)
model.fc = torch.nn.Identity()
input_tensor = torch.randn(4, 3, 224, 224)
features = model(input_tensor)
print(features.shape)

medium

A. torch.Size([4, 512])

B. torch.Size([4, 1000])

C. torch.Size([4, 3, 224, 224])

D. torch.Size([4, 2048])

Solution

1. Step 1: Understand model modification

   Replacing model.fc with Identity removes the final classification layer, so output is the feature vector before classification.

2. Step 2: Know ResNet18 feature size

   ResNet18 outputs a 512-dimensional vector before the final fc layer for each input image.

3. Final Answer:

   torch.Size([4, 512]) -> Option A

4. Quick Check:

   ResNet18 features = 512 dims [OK]

Hint: Identity layer outputs feature vector size [OK]

Common Mistakes:

• Assuming output is 1000 classes without removing fc
• Confusing batch size with feature dimension
• Expecting 2048 features from ResNet18 (it's 512)

4. Identify the error in this feature extraction code snippet and select the fix:

model = models.resnet50(pretrained=True)
for param in model.parameters():
    param.requires_grad = False
model.fc = nn.Linear(2048, 5)
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)

# Training loop here

medium

A. No error; code is correct

B. Set requires_grad=True for model.fc parameters after replacement

C. Use Adam optimizer instead of SGD

D. Remove freezing of parameters to train all layers

Solution

1. Step 1: Check freezing timing

   The loop freezes existing parameters before replacing model.fc, so the new fc layer's parameters are created with requires_grad=True by default.

2. Step 2: Verify optimizer behavior

   Optimizer only updates parameters where requires_grad=True, which are the new fc parameters; backbone remains frozen.

3. Final Answer:

   No error; code is correct -> Option A

4. Quick Check:

   New layer params unfrozen by default [OK]

Hint: New layers have requires_grad=True by default [OK]

Common Mistakes:

• Assuming freezing all parameters includes new layers
• Changing optimizer without fixing requires_grad
• Removing freezing unnecessarily

5. You want to use a pre-trained ResNet34 to classify 3 classes in your dataset. You freeze all layers except the last one. However, your training accuracy stays very low. What is the best next step to improve feature extraction performance?

hard

A. Reduce batch size to 1 to improve gradient estimates

B. Increase learning rate to 1.0 for faster training

C. Replace the optimizer with SGD without momentum

D. Unfreeze some deeper layers to fine-tune features for your task

Solution

1. Step 1: Understand freezing impact

   Freezing all but last layer may limit model's ability to adapt features to new classes, causing low accuracy.

2. Step 2: Fine-tune some deeper layers

   Unfreezing some layers closer to output allows the model to adjust features better for your specific dataset.

3. Final Answer:

   Unfreeze some deeper layers to fine-tune features for your task -> Option D

4. Quick Check:

   Fine-tune layers = better adaptation [OK]

Hint: Fine-tune layers if frozen model underperforms [OK]

Common Mistakes:

• Increasing learning rate too much causes instability
• Changing optimizer without addressing feature adaptation
• Reducing batch size unnecessarily