|
|
import torch |
|
|
import numpy as np |
|
|
from jaxtyping import jaxtyped |
|
|
import typeguard |
|
|
import typing |
|
|
from functorch.dim import tree_map |
|
|
import torch |
|
|
|
|
|
|
|
|
@jaxtyped(typechecker=typeguard.typechecked) |
|
|
def torch_to_numpy(tensor: typing.Any) -> np.ndarray | float: |
|
|
""" |
|
|
Convert a torch tensor to a numpy array. |
|
|
""" |
|
|
if isinstance(tensor, torch.Tensor): |
|
|
return tensor.detach().cpu().numpy() |
|
|
else: |
|
|
return tensor |
|
|
|
|
|
|
|
|
@jaxtyped(typechecker=typeguard.typechecked) |
|
|
def torch_dict_to_numpy(d: dict) -> dict: |
|
|
return tree_map(torch_to_numpy, d) |
|
|
|
|
|
|
|
|
@jaxtyped(typechecker=typeguard.typechecked) |
|
|
def compute_grad_norm(model: torch.nn.Module, grads: None) -> float: |
|
|
|
|
|
total_norm = 0 |
|
|
if grads is not None: |
|
|
for p in grads: |
|
|
param_norm = p.norm(2) |
|
|
total_norm += param_norm.item() ** 2 |
|
|
total_norm = total_norm ** (1.0 / 2) |
|
|
return total_norm |
|
|
|
|
|
for p in model.parameters(): |
|
|
param_norm = p.grad.data.norm(2) |
|
|
total_norm += param_norm.item() ** 2 |
|
|
total_norm = total_norm ** (1.0 / 2) |
|
|
return total_norm |
|
|
|
|
|
|
|
|
def is_numpy(x: typing.Any) -> bool: |
|
|
return isinstance(x, np.ndarray) |
|
|
|
