PTCA sample questions

Question 1

A researcher is implementing a custom regularization penalty. They need to extract an intermediate feature map from the forward pass, perform some non-differentiable numpy operations on it to calculate a logging metric, and then continue the standard backward pass using the original feature map.

Which PyTorch mechanism should be used to extract the feature map for the logging metric without breaking or modifying the computation graph?

Accepted Answer

Use feature_map.detach() to create a new tensor that shares storage but is completely disconnected from the computation graph.

Answer

Call feature_map.clone() and pass it to numpy.

Answer

Wrap the logging operation in a with torch.no_grad(): block.

Answer

Set feature_map.requires_grad = False temporarily during the logging step.

Question 2

True or False: Calling `model.eval()` automatically disables gradient computation during the forward pass, effectively acting as a substitute for `torch.no_grad()` during inference.

Accepted Answer

False

Answer

True

Question 3

During the setup phase of a deep learning pipeline, an architect configures the model and optimizer. Which of the following sequences represents the correct best practice for initializing an optimizer in relation to moving the model to a target hardware device (e.g., GPU)?

Accepted Answer

Call model.to(device), then initialize the optimizer with model.parameters().

Answer

Initialize the optimizer with model.parameters(), then call model.to(device).

Answer

Call optimizer.to(device) immediately after initializing it with model parameters.

Answer

The order does not matter as PyTorch optimizers automatically track parameter device migrations.

Question 4

A computer vision model requires blending a feature map tensor `A` with a per-channel bias tensor `B`. Tensor `A` has the shape `(16, 3, 256, 256)` representing (batch, channels, height, width). Tensor `B` has the shape `(3, 1, 1)`. When the operation `C = A + B` is executed, what will be the resulting shape of tensor `C` based on PyTorch broadcasting semantics?

Accepted Answer

(16, 3, 256, 256)

Answer

(16, 6, 256, 256)

Answer

A RuntimeError will be thrown due to shape mismatch.

Answer

(3, 16, 256, 256)

Question 5

To compute the gradients of the loss with respect to all tensors in the computation graph that have requires_grad=True, you must call the ________ method on the final loss tensor.

Accepted Answer

backward()

Answer

step()

Answer

compute_grad()

Answer

autograd()

Question 6

A data scientist observes that their model's training loss is behaving erratically, oscillating wildly and occasionally exploding to infinity, despite a very small learning rate. They review their core training loop:

```python
for data, target in dataloader:
output = model(data)
loss = criterion(output, target)
loss.backward()
optimizer.step()
```

What critical omission in this training loop is causing the erratic loss behavior?

Accepted Answer

The gradients are not being explicitly cleared, causing them to accumulate indefinitely across batches.

Answer

The model was not set to train mode using model.train() inside the loop.

Answer

The data and target tensors were not moved to the GPU.

Answer

The loss tensor must be detached from the graph before calling backward().

Free PTCA sample questions

True or False: Calling `model.eval()` automatically disables gradient computation during the forward pass, effectively acting as a substitute for `torch.no_grad()` during inference.

During the setup phase of a deep learning pipeline, an architect configures the model and optimizer. Which of the following sequences represents the correct best practice for initializing an optimizer in relation to moving the model to a target hardware device (e.g., GPU)?

To compute the gradients of the loss with respect to all tensors in the computation graph that have requires_grad=True, you must call the ________ method on the final loss tensor.

6 more free samples are waiting