Question 1 of 20

A 3×3 convolution with stride 1 and padding 1 on a H×W feature map outputs size:

Question 2 of 20

Receptive field of a deep CNN layer refers to:

Question 3 of 20

Batch normalization during training primarily helps by:

Question 4 of 20

ResNet skip connections help very deep networks by:

Question 5 of 20

Random horizontal flip during training is safe for natural pet photos because:

Question 6 of 20

ColorJitter during training helps generalization by:

Question 7 of 20

Freezing early CNN layers during fine-tuning is common because:

Question 8 of 20

When your dataset is small and similar to ImageNet, a good strategy is:

Question 9 of 20

RandomResizedCrop during training but CenterCrop at validation ensures:

Question 10 of 20

Global average pooling before a classifier head:

Question 11 of 20

Domain shift between pretraining and deployment data means:

Question 12 of 20

Applying heavy augmentation at inference time would:

Question 13 of 20

Discriminative learning rates in fine-tuning mean:

Question 14 of 20

1×1 convolutions in networks like Inception/ResNet are used to:

Question 15 of 20

Weight decay (L2 regularization) during CNN training:

Question 16 of 20

Early stopping during fine-tuning uses validation loss to:

Question 17 of 20

CrossEntropyLoss with ResNet classifier outputs expects:

Question 18 of 20

If train accuracy is 99% but val accuracy is 62%, the first suspects are:

Question 19 of 20

Replacing ResNet's final `fc` layer when fine-tuning is necessary because:

Question 20 of 20

Exporting a PyTorch classifier to ONNX enables: