Why pre-trained models save time in Computer Vision - Why Metrics Matter

Metrics & Evaluation - Why pre-trained models save time

Which metric matters for this concept and WHY

When using pre-trained models, the key metric to watch is training time and accuracy. Pre-trained models save time because they start with knowledge from previous training. This means they need fewer steps to learn your task well. Watching accuracy ensures the model still performs well after fine-tuning.

Confusion matrix or equivalent visualization (ASCII)

    Example confusion matrix after fine-tuning a pre-trained model:

          Predicted
          +-----+-----+
          | Pos | Neg |
    +-----+-----+-----+
    | Pos |  85 |  10 |
    | Neg |  15 |  90 |
    +-----+-----+-----+

    Total samples = 200
    TP = 85, FP = 15, FN = 10, TN = 90

    This shows good accuracy and balanced errors after less training time.
    

Precision vs Recall tradeoff with concrete examples

Pre-trained models help balance precision and recall faster. For example:

• High precision: The model correctly identifies objects without many false alarms. Useful in quality control where mistakes are costly.
• High recall: The model finds most objects, even if some are wrong. Useful in safety checks where missing an object is bad.

Pre-trained models start with good features, so you can quickly adjust this balance with less data and time.

What "good" vs "bad" metric values look like for this use case

Good: Accuracy above 85% after a few training epochs, with balanced precision and recall around 80% or higher. Training time is short because the model already knows useful features.

Bad: Accuracy below 70%, or very low recall or precision, meaning the model is not learning well. Training takes a long time because the model starts from scratch.

Metrics pitfalls (accuracy paradox, data leakage, overfitting indicators)

• Accuracy paradox: High accuracy can be misleading if data is unbalanced. For example, if most images are background, the model might guess background and get high accuracy but fail to detect objects.
• Data leakage: Using test images in training can make metrics look better than reality.
• Overfitting: Very high training accuracy but low test accuracy means the model memorized training data and won't generalize well.

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

No, it is not good for fraud detection. The high accuracy likely comes from many normal cases. The very low recall means the model misses most fraud cases, which is dangerous. For fraud, recall is critical because missing fraud is costly.

Key Result

Pre-trained models reduce training time while maintaining good accuracy and balanced precision-recall.

Practice

1. Why do pre-trained models save time in computer vision projects?

easy

A. They require more data to train from scratch

B. They eliminate the need for any data preprocessing

C. They always produce perfect results without any training

D. They reuse features learned from large datasets, reducing training time

Solution

1. Step 1: Understand what pre-trained models do

   Pre-trained models have already learned useful features from large datasets, so you don't start from zero.

2. Step 2: Connect this to time saved

   Since the model already knows many features, you spend less time training it on your own data.

3. Final Answer:

   They reuse features learned from large datasets, reducing training time -> Option D

4. Quick Check:

   Pre-trained models reuse features = B [OK]

Hint: Pre-trained means already learned features reused [OK]

Common Mistakes:

• Thinking pre-trained models need more data
• Believing they need no training at all
• Assuming they remove all preprocessing

2. Which of the following is the correct way to load a pre-trained model in Python using PyTorch?

easy

A. model = torchvision.models.resnet50(pretrained=True)

B. model = torchvision.models.resnet50(pretrained=False)

C. model = torchvision.load_model('resnet50')

D. model = torch.load('resnet50_pretrained')

Solution

1. Step 1: Recall PyTorch syntax for loading pre-trained models

   In PyTorch, you use torchvision.models with pretrained=True to load a pre-trained model.

2. Step 2: Check options for correctness

   model = torchvision.models.resnet50(pretrained=True) uses the correct function and argument. model = torchvision.models.resnet50(pretrained=False) loads without pre-training. Options C and D are incorrect function calls.

3. Final Answer:

   model = torchvision.models.resnet50(pretrained=True) -> Option A

4. Quick Check:

   PyTorch pre-trained load = A [OK]

Hint: Use pretrained=True to load pre-trained models in PyTorch [OK]

Common Mistakes:

• Using pretrained=False by mistake
• Calling non-existent functions like torchvision.load_model
• Trying to load model weights incorrectly

3. Consider this Python code using TensorFlow to load a pre-trained MobileNetV2 model and predict on an input image:

import tensorflow as tf
model = tf.keras.applications.MobileNetV2(weights='imagenet')
import numpy as np
input_data = np.random.rand(1, 224, 224, 3).astype('float32')
predictions = model.predict(input_data)
print(predictions.shape)

What will be the printed output shape?

medium

A. (224, 224, 3)

B. (1, 1000)

C. (1, 224, 224, 3)

D. (1000,)

Solution

1. Step 1: Understand MobileNetV2 output shape

   MobileNetV2 pre-trained on ImageNet outputs predictions for 1000 classes, so output shape is (batch_size, 1000).

2. Step 2: Check input batch size and output

   Input batch size is 1, so output shape is (1, 1000).

3. Final Answer:

   (1, 1000) -> Option B

4. Quick Check:

   Output shape = (batch, 1000 classes) = A [OK]

Hint: Output shape matches batch size and number of classes [OK]

Common Mistakes:

• Confusing input shape with output shape
• Ignoring batch dimension
• Expecting output shape to match input image size

4. You tried to fine-tune a pre-trained model but got an error: AttributeError: 'Sequential' object has no attribute 'fc'. What is the likely cause?

medium

A. You used a model architecture without an 'fc' layer and tried to access it

B. You forgot to load pre-trained weights

C. You passed wrong input shape to the model

D. You used the wrong optimizer

Solution

1. Step 1: Understand the error message

   The error says the model has no attribute 'fc', which usually means the model architecture does not have a fully connected layer named 'fc'.

2. Step 2: Connect error to cause

   Trying to access or modify 'fc' layer on a Sequential model that doesn't have it causes this error.

3. Final Answer:

   You used a model architecture without an 'fc' layer and tried to access it -> Option A

4. Quick Check:

   Missing 'fc' layer attribute = D [OK]

Hint: Check if model has the layer before accessing it [OK]

Common Mistakes:

• Assuming all models have 'fc' layer
• Ignoring error details
• Blaming optimizer or input shape wrongly

5. You want to use a pre-trained model to classify images of cats and dogs but your dataset has only 500 images. Which approach saves the most time while achieving good accuracy?

hard

A. Use a pre-trained model without any fine-tuning and directly predict

B. Train a new model from scratch with random weights on your 500 images

C. Use a pre-trained model and fine-tune only the last layer on your dataset

D. Manually label more images before training any model

Solution

1. Step 1: Consider dataset size and training time

   With only 500 images, training from scratch is slow and likely inaccurate.

2. Step 2: Use pre-trained model fine-tuning

   Fine-tuning only the last layer uses learned features and adapts to your task quickly and efficiently.

3. Final Answer:

   Use a pre-trained model and fine-tune only the last layer on your dataset -> Option C

4. Quick Check:

   Fine-tune last layer for small data = C [OK]

Hint: Fine-tune last layer for small datasets [OK]

Common Mistakes:

• Training from scratch with little data
• Skipping fine-tuning and expecting perfect results
• Spending time labeling more data unnecessarily