Compose transforms in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - Compose transforms

Which metric matters for Compose transforms and WHY

Compose transforms are used to prepare data before training a model. The key metric to check here is data consistency and correctness. This means the transformed data should still represent the original information well, without errors or distortions.

For example, if you normalize images, the pixel values should be scaled correctly. If you flip or crop images, the labels should still match the content. So, metrics like data integrity checks or visual inspection are important.

In practice, after applying Compose transforms, you want to see if your model's training loss decreases and accuracy improves. This shows the transforms help the model learn better.

Confusion matrix or equivalent visualization

Compose transforms do not directly produce predictions, so no confusion matrix applies here.

Instead, you can visualize sample images before and after transforms to check correctness.

Original Image:  [Image of a cat]
After Compose:   [Flipped, normalized image of the same cat]
Label: Cat
    

Precision vs Recall tradeoff with concrete examples

Compose transforms affect data quality, which indirectly impacts precision and recall of the model.

If transforms are too aggressive (e.g., too much cropping), important features may be lost, causing the model to miss true positives (lower recall).

If transforms are too weak or inconsistent, the model may learn noise, causing false positives (lower precision).

Example: For a face detector, if Compose transforms randomly flip images but labels are not adjusted, the model may confuse left and right faces, hurting precision.

What "good" vs "bad" metric values look like for Compose transforms

Good Compose transforms lead to:

• Stable or improved training loss over epochs
• Improved validation accuracy
• Consistent data samples that match labels

Bad Compose transforms cause:

• Training loss that does not decrease or fluctuates wildly
• Validation accuracy that drops or is unstable
• Visual mismatch between transformed data and labels

Metrics pitfalls

• Data leakage: Applying transforms that use test data statistics can leak information and inflate metrics.
• Overfitting: Overly complex transforms may cause the model to memorize transformed data, hurting generalization.
• Incorrect label alignment: Some transforms (like flipping) require label changes; forgetting this causes wrong labels and poor metrics.
• Ignoring normalization: Not normalizing data properly can slow training and reduce accuracy.

Self-check question

Your model trained with Compose transforms has 98% accuracy but only 12% recall on fraud cases. Is it good for production? Why not?

Answer: No, it is not good. High accuracy can be misleading if the data is imbalanced (few fraud cases). The very low recall means the model misses most fraud cases, which is dangerous. You need to improve recall to catch more fraud, even if accuracy drops a bit.

Key Result

Compose transforms impact data quality, which affects model training loss and accuracy; correct transforms improve these metrics and maintain label consistency.

Practice

1. What is the main purpose of Compose in PyTorch transforms?

easy

A. To chain multiple image transforms into one step

B. To train a neural network model

C. To load datasets from disk

D. To visualize images in a grid

Solution

1. Step 1: Understand Compose functionality

   Compose is used to combine several image transformations so they run one after another.

2. Step 2: Compare with other options

   Training models, loading datasets, and visualization are different tasks not handled by Compose.

3. Final Answer:

   To chain multiple image transforms into one step -> Option A

4. Quick Check:

   Compose chains transforms = C [OK]

Hint: Compose chains transforms in order for easy preprocessing [OK]

Common Mistakes:

• Thinking Compose trains models
• Confusing Compose with dataset loading
• Assuming Compose visualizes images

2. Which of the following is the correct way to create a Compose transform that converts images to tensor and normalizes them?

easy

A. transforms.Compose(transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,)))

B. transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])

C. transforms.Compose{transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))}

D. transforms.Compose([transforms.Normalize((0.5,), (0.5,)), transforms.ToTensor()])

Solution

1. Step 1: Check Compose syntax

   Compose takes a list of transforms inside square brackets [].

2. Step 2: Verify order and syntax

   transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))]) uses correct list syntax and order: ToTensor first, then Normalize. transforms.Compose(transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))) misses list brackets, C uses curly braces, D reverses order which is incorrect.

3. Final Answer:

   transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))]) -> Option B

4. Quick Check:

   Compose needs list and correct order = A [OK]

Hint: Use square brackets [] inside Compose and correct transform order [OK]

Common Mistakes:

• Omitting list brackets in Compose
• Using curly braces instead of brackets
• Reversing transform order

3. What will be the output shape of the tensor after applying this Compose transform on a 28x28 grayscale image?

transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Resize((14, 14))
])
img_tensor = transform(image)

medium

A. [3, 14, 14]

B. [1, 28, 28]

C. [1, 14, 14]

D. [3, 28, 28]

Solution

1. Step 1: Understand ToTensor output

   For a grayscale image, ToTensor() converts it to a tensor with shape [channels, height, width] where channels=1.

2. Step 2: Apply Resize transform

   Resize((14,14)) changes height and width to 14 each, channels remain 1.

3. Final Answer:

   [1, 14, 14] -> Option C

4. Quick Check:

   Grayscale tensor shape after resize = [1,14,14] [OK]

Hint: Grayscale images have 1 channel; Resize changes height and width [OK]

Common Mistakes:

• Assuming 3 channels for grayscale
• Ignoring Resize effect on shape
• Confusing height and width order

4. Identify the error in this Compose transform code:

transform = transforms.Compose([
    transforms.Normalize((0.5,), (0.5,)),
    transforms.ToTensor()
])

medium

A. ToTensor should be replaced with ToPILImage

B. Normalize parameters are incorrect

C. Compose requires a tuple, not a list

D. Normalize should come after ToTensor

Solution

1. Step 1: Check transform order

   Normalize expects a tensor input, so ToTensor() must come first to convert image to tensor.

2. Step 2: Validate other options

   Parameters for Normalize are correct, Compose accepts list, and ToPILImage is not needed here.

3. Final Answer:

   Normalize should come after ToTensor -> Option D

4. Quick Check:

   Normalize after ToTensor = A [OK]

Hint: Always put ToTensor before Normalize in Compose [OK]

Common Mistakes:

• Placing Normalize before ToTensor
• Confusing list and tuple for Compose
• Replacing ToTensor incorrectly

5. You want to apply a Compose transform that first randomly flips an image horizontally, then converts it to a tensor, and finally normalizes it. Which code snippet correctly implements this?

hard

A. transforms.Compose([transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])

B. transforms.Compose([transforms.ToTensor(), transforms.RandomHorizontalFlip(), transforms.Normalize((0.5,), (0.5,))])

C. transforms.Compose([transforms.Normalize((0.5,), (0.5,)), transforms.RandomHorizontalFlip(), transforms.ToTensor()])

D. transforms.Compose([transforms.RandomHorizontalFlip(), transforms.Normalize((0.5,), (0.5,)), transforms.ToTensor()])

Solution

1. Step 1: Order of transforms

   RandomHorizontalFlip works on PIL images, so it must come before ToTensor which converts to tensor.

2. Step 2: Normalize after ToTensor

   Normalize requires tensor input, so it must come after ToTensor.

3. Step 3: Verify options

   transforms.Compose([transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))]) has correct order: flip, then tensor, then normalize. Others have wrong order causing errors or wrong behavior.

4. Final Answer:

   transforms.Compose([transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))]) -> Option A

5. Quick Check:

   Flip before ToTensor, Normalize last = D [OK]

Hint: Flip before ToTensor; Normalize last in Compose chain [OK]

Common Mistakes:

• Placing Normalize before ToTensor
• Applying flip after ToTensor (wrong input type)
• Mixing order causing runtime errors