21、Transformer Masked loss原理精讲及其PyTorch逐行实现

取个名字真难呐2025-01-13 13:21

1. Transformer结构图

2. python

python 复制代码

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

torch.set_printoptions(precision=3, sci_mode=False)

if __name__ == "__main__":
    run_code = 0
    batch_size = 2
    seq_length = 3
    vocab_size = 4
    logits = torch.randn(batch_size,seq_length,vocab_size)
    print(f"logits=\n{logits}")
    logits_t = logits.transpose(-1,-2)
    print(f"logits_t=\n{logits_t}")

    label = torch.randint(0,vocab_size,(batch_size,seq_length))
    print(f"label=\n{label}")
    result_none = F.cross_entropy(logits_t,label,reduction="none")
    print(f"result_none=\n{result_none}")
    result_none_mean = torch.mean(result_none)
    result_mean = F.cross_entropy(logits_t,label)
    print(f"result_mean=\n{result_mean}")
    print(f"result_none_mean={result_none_mean}")

python 复制代码

logits=
tensor([[[ 0.477,  2.017,  1.016, -0.299],
         [-0.189,  0.321, -0.885,  1.418],
         [ 0.027, -0.606,  0.079, -0.491]],

        [[ 1.911,  1.643, -0.327,  0.185],
         [-0.031, -1.463, -0.073,  1.391],
         [-0.710,  0.811,  1.521,  0.033]]])
logits_t=
tensor([[[ 0.477, -0.189,  0.027],
         [ 2.017,  0.321, -0.606],
         [ 1.016, -0.885,  0.079],
         [-0.299,  1.418, -0.491]],

        [[ 1.911, -0.031, -0.710],
         [ 1.643, -1.463,  0.811],
         [-0.327, -0.073,  1.521],
         [ 0.185,  1.391,  0.033]]])
label=
tensor([[0, 0, 0],
        [3, 0, 0]])
result_none=
tensor([[2.059, 2.098, 1.157],
        [2.444, 1.848, 2.832]])
result_mean=
2.0730881690979004
result_none_mean=2.0730881690979004