PyTorch Training Loop with AMP and Clipping | Dataford Interview Questions

Implement a custom training loop in PyTorch that incorporates gradient clipping and mixed-precision training. For dependency-free grading, implement the numerical core with JSON-compatible linear-model weights and batches, preserving the required order: scale gradients, unscale them, clip their norm, then update parameters. Return the final weights and average loss for each epoch using custom_training_loop(weights, batches, learning_rate, max_grad_norm, epochs, loss_scale).

Constraints

1 <= len(weights) <= 100
1 <= len(batches) <= 100
Each batch contains at least one feature row
Every feature row has len(weights) values
Each batch has the same feature width as weights
learning_rate > 0
max_grad_norm > 0
loss_scale > 0
1 <= epochs <= 100

Function Signature

def custom_training_loop(weights, batches, learning_rate, max_grad_norm, epochs, loss_scale):

Your question is PyTorch Training Loop with AMP and Clipping. Start with the requirements in the Question tab.

Run and submit as often as you like. When you're ready, talk me through your approach or go straight to the code.

Problem

Constraints

1 <= len(weights) <= 100
1 <= len(batches) <= 100
Each batch contains at least one feature row
Every feature row has len(weights) values
Each batch has the same feature width as weights
learning_rate > 0
max_grad_norm > 0
loss_scale > 0
1 <= epochs <= 100

Function Signature

def custom_training_loop(weights, batches, learning_rate, max_grad_norm, epochs, loss_scale):

Your question is PyTorch Training Loop with AMP and Clipping. Start with the requirements in the Question tab.

Run and submit as often as you like. When you're ready, talk me through your approach or go straight to the code.