utils

Clip the global gradient norm of parameters in place.

Clip the global gradient norm of parameters in place.

Clamp every gradient element to $[-\text{clip\_value}, \text{clip\_value}]$ in place.

Compute the $\ell_p$ norm of tensors, taken jointly.

Flatten an iterable of parameters into a single 1-D Tensor.

Copy a flat 1-D tensor back into the supplied parameters in place.

Reverse weight_norm and restore a plain leaf parameter.

Reparametrise a weight as direction times magnitude (Salimans & Kingma 2016).

Reverse spectral_norm and restore a plain leaf parameter.

Constrain a weight's spectral norm via power iteration (Miyato et al. 2018).

Pack a padded $(T, B, *)$ batch into a PackedSequence.

One-shot pack-from-list — equivalent to pad_sequence then pack_padded_sequence.

Inverse of pack_padded_sequence — produce a padded tensor.

Stack a list of variable-length tensors into a padded batch.

Copy all parameter / buffer values from source into dest.

Fold a BatchNorm layer into the preceding Conv weights (inference-only).

Low-level form of fuse_conv_bn_eval operating on raw tensors.

Fold a BatchNorm1d into the preceding Linear weights (inference-only).

Low-level form of fuse_linear_bn_eval operating on raw tensors.

Construct a module while skipping its parameter initialisation.