protomotions.utils.torch_utils module

PyTorch utility functions.

Includes helper functions for gradient computation, tensor conversion, and seeding.

protomotions.utils.torch_utils.grad_norm(params)[source]

Compute L2 norm of gradients across all parameters.

Parameters:

params – List of parameters with gradients.

Returns:

Scalar tensor with gradient norm.

protomotions.utils.torch_utils.to_torch(x, dtype=<Mock object>, device='cuda:0', requires_grad=False)[source]

Convert data to PyTorch tensor with specified dtype and device.

Parameters:

  • x – Data to convert.

  • dtype – PyTorch data type.

  • device – Target device.

  • requires_grad – Whether tensor requires gradients.

Returns:

PyTorch tensor.

Return type:

MockTensor

protomotions.utils.torch_utils.seeding(seed=0, torch_deterministic=False)[source]

Set random seeds for reproducibility.

Parameters:

  • seed – Integer seed value.

  • torch_deterministic – If True, configure PyTorch for deterministic execution.

Returns:

The seed used.