KL 散度

算KL散度的時候要注意前后順序以及加log

import torhch.nn as nn
d_loss = nn.KLDivLoss(reduction=reduction_kd)(F.log_softmax(y / T, dim=1),F.softmax(teacher_scores / T, dim=1)) * T * T

蒸餾loss T 為

L2 loss

import torch.nn.functional as F
F.mse_loss(teacher_patience.float(), student_patience.float()).half()

做標準化處理

if normalized_patience:teacher_patience = F.normalize(teacher_patience, p=2, dim=2)student_patience = F.normalize(student_patience, p=2, dim=2)

L2 范數

CEloss

分類問題

nll_loss = F.cross_entropy(y, labels, reduction=reduction_nll)

CTCLoss

計算連續(未分割)時間序列和目標序列之間的損失.
torch.nn.CTCLoss(blank=0, reduction=‘mean’, zero_infinity=False)

>>> # Target are to be padded
>>> T = 50      # Input sequence length
>>> C = 20      # Number of classes (including blank)
>>> N = 16      # Batch size
>>> S = 30      # Target sequence length of longest target in batch (padding length)
>>> S_min = 10  # Minimum target length, for demonstration purposes
>>>
>>> # Initialize random batch of input vectors, for *size = (T,N,C)
>>> input = torch.randn(T, N, C).log_softmax(2).detach().requires_grad_()
>>>
>>> # Initialize random batch of targets (0 = blank, 1:C = classes)
>>> target = torch.randint(low=1, high=C, size=(N, S), dtype=torch.long)
>>>
>>> input_lengths = torch.full(size=(N,), fill_value=T, dtype=torch.long)
>>> target_lengths = torch.randint(low=S_min, high=S, size=(N,), dtype=torch.long)
>>> ctc_loss = nn.CTCLoss()
>>> loss = ctc_loss(input, target, input_lengths, target_lengths)

圖片卷積-》序列做loss
31,1，64,256 -》 31,512,3，65-》31,65，512,3-》polling-》31,65，512》》liner》31,65,103》TNC》65,31,103

AdaptiveAvgPool2d

自適應池化，最后設置要輸出的H和W，B和N（C）不變