import torch
import torch.nn as nn
import torch.nn.functional as F
import itertools

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

PAD_token = 0

def zeroPadding(l, fillvalue=PAD_token):
 # 输入：[[1, 1, 1], [2, 2], [3]]
 # 返回：[(1, 2, 3), (1, 2, 0), (1, 0, 0)] 返回已经是转置后的 [L, B]
 return list(itertools.zip_longest(*l, fillvalue=fillvalue))


def binaryMatrix(l):
 # 将targets里非pad部分标记为1，pad部分标记为0
 m = []
 for i, seq in enumerate(l):
 m.append([])
 for token in seq:
  if token == PAD_token:
  m[i].append(0)
  else:
  m[i].append(1)
 return m

inputs = [[1, 2, 3], [4, 5], [6]] # 输入句，一个batch，需要按照长度从大到小排好序
inputs_lengths = [3, 2, 1]
targets = [[1, 2], [1, 2, 3], [1]] # 目标句，这里的长度是不确定的，mask是针对targets的
inputs_batch = torch.LongTensor(zeroPadding(inputs))
inputs_lengths = torch.LongTensor(inputs_lengths)
targets_batch = torch.LongTensor(zeroPadding(targets))
targets_mask = torch.ByteTensor(binaryMatrix(zeroPadding(targets))) # 注意这里是ByteTensor
print(inputs_batch)
print(targets_batch)
print(targets_mask)

tensor([[ 1, 4, 6],
 [ 2, 5, 0],
 [ 3, 0, 0]])
tensor([[ 1, 1, 1],
 [ 2, 2, 0],
 [ 0, 3, 0]])
tensor([[ 1, 1, 1],
 [ 1, 1, 0],
 [ 0, 1, 0]], dtype=torch.uint8)

outputs1 = torch.FloatTensor([[0.2, 0.1, 0.7], [0.3, 0.6, 0.1], [0.4, 0.5, 0.1]])
print(outputs1)

tensor([[ 0.2000, 0.1000, 0.7000],
 [ 0.3000, 0.6000, 0.1000],
 [ 0.4000, 0.5000, 0.1000]])

out[i][j][k] = input[index[i][j][k]][j][k] # if dim == 0
out[i][j][k] = input[i][index[i][j][k]][k] # if dim == 1
out[i][j][k] = input[i][j][index[i][j][k]] # if dim == 2

# 收集每行的第2个元素
temp = torch.gather(outputs1, 1, torch.LongTensor([[1], [1], [1]]))
print(temp)

tensor([[ 0.1000],
 [ 0.6000],
 [ 0.5000]])

mask = temp.ge(0.5) # 大于等于0.5
print(mask)
print(torch.masked_select(temp, temp.ge(0.5)))

tensor([[ 0],
 [ 1],
 [ 1]], dtype=torch.uint8)
tensor([ 0.6000, 0.5000])

# 计算一个batch内的平均负对数似然损失，即只考虑mask为1的元素
def maskNLLLoss(inp, target, mask):
 nTotal = mask.sum()
 # 收集目标词的概率，并取负对数
 crossEntropy = -torch.log(torch.gather(inp, 1, target.view(-1, 1)))
 # 只保留mask中值为1的部分，并求均值
 loss = crossEntropy.masked_select(mask).mean()
 loss = loss.to(DEVICE)
 return loss, nTotal.item()

# 计算预测的第一行和targets的第一行的loss
maskNLLLoss(outputs1, targets_batch[0], targets_mask[0])

(tensor(1.1689, device='cuda:0'), 3)