util.py

"""
@Origin : util.py by Yue Wang
@Contact: yuewangx@mit.edu
@Time: 4/5/19 3:47 PM

modified by {Sanghyeok Lee, Sihyeon Kim}
@Contact: {cat0626, sh_bs15}@korea.ac.kr
@File: util.py
@Time: 2021.09.29
"""


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)
        
        #Move to model.py
        #log_prb = F.log_softmax(pred, dim=1)

        loss = -(one_hot * pred).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()