DL-Art-School/codes/utils/loss_accumulator.py

24 lines
827 B
Python

import torch
# Utility class that stores detached, named losses in a rotating buffer for smooth metric outputting.
class LossAccumulator:
def __init__(self, buffer_sz=50):
self.buffer_sz = buffer_sz
self.buffers = {}
def add_loss(self, name, tensor):
if name not in self.buffers.keys():
self.buffers[name] = (0, torch.zeros(self.buffer_sz))
i, buf = self.buffers[name]
# Can take tensors or just plain python numbers.
if isinstance(tensor, torch.Tensor):
buf[i] = tensor.detach().cpu()
else:
buf[i] = tensor
self.buffers[name] = ((i+1) % self.buffer_sz, buf)
def as_dict(self):
result = {}
for k, v in self.buffers.items():
result["loss_" + k] = torch.mean(v[1])
return result