PointMLP/part_segmentation/util.py
2021-10-04 12:39:33 -04:00

39 lines
862 B
Python

import numpy as np
import torch
import torch.nn.functional as F
def cal_loss(pred, gold, smoothing=True):
''' Calculate cross entropy loss, apply label smoothing if needed. '''
gold = gold.contiguous().view(-1)
if smoothing:
eps = 0.2
n_class = pred.size(1)
one_hot = torch.zeros_like(pred).scatter(1, gold.view(-1, 1), 1)
one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)
log_prb = F.log_softmax(pred, dim=1)
loss = -(one_hot * log_prb).sum(dim=1).mean()
else:
loss = F.cross_entropy(pred, gold, reduction='mean')
return loss
class IOStream():
def __init__(self, path):
self.f = open(path, 'a')
def cprint(self, text):
print(text)
self.f.write(text+'\n')
self.f.flush()
def close(self):
self.f.close()