ප්රයිමර් EZ විචලනයන්

12import torch
13from torch import nn
14
15from labml_helpers.module import Module
16from labml_nn.transformers import MultiHeadAttention

19class SpatialDepthWiseSharedConvolution(Module):

26    def __init__(self, kernel_size: int = 3):
27        super().__init__()
28        self.kernel_size = kernel_size

33        self.conv = nn.Conv1d(in_channels=1, out_channels=1, kernel_size=(kernel_size,), padding=(kernel_size - 1,))

35    def forward(self, x: torch.Tensor):

41        seq_len, batch_size, heads, d_k = x.shape

43        x = x.permute(1, 2, 3, 0)

45        x = x.view(batch_size * heads * d_k, 1, seq_len)

48        x = self.conv(x)

50        x = x[:, :, :-(self.kernel_size - 1)]

52        x = x.view(batch_size, heads, d_k, seq_len)

54        x = x.permute(3, 0, 1, 2)

57        return x

60class MultiDSharedConvHeadAttention(MultiHeadAttention):

68    def __init__(self, heads: int, d_model: int, dropout_prob: float = 0.1):
69        super().__init__(heads, d_model, dropout_prob)

76        self.query = nn.Sequential(self.query, SpatialDepthWiseSharedConvolution())
77        self.key = nn.Sequential(self.key, SpatialDepthWiseSharedConvolution())
78        self.value = nn.Sequential(self.value, SpatialDepthWiseSharedConvolution())

81class SpatialDepthWisePerHeadConvolution(Module):

86    def __init__(self, heads: int, d_k: int, kernel_size: int = 3):

91        super().__init__()
92        self.kernel_size = kernel_size

98        self.conv = nn.Conv1d(in_channels=d_k * heads, out_channels=d_k * heads,
99                              kernel_size=(kernel_size,), padding=(kernel_size - 1,), groups=d_k * heads)

101    def forward(self, x: torch.Tensor):

107        seq_len, batch_size, heads, d_k = x.shape

109        x = x.permute(1, 2, 3, 0)

111        x = x.view(batch_size, heads * d_k, seq_len)

114        x = self.conv(x)

116        x = x[:, :, :-(self.kernel_size - 1)]

118        x = x.view(batch_size, heads, d_k, seq_len)

120        x = x.permute(3, 0, 1, 2)

123        return x

126class MultiDPHConvHeadAttention(MultiHeadAttention):

134    def __init__(self, heads: int, d_model: int, dropout_prob: float = 0.1):
135        super().__init__(heads, d_model, dropout_prob)

142        self.query = nn.Sequential(self.query, SpatialDepthWisePerHeadConvolution(heads, self.d_k))
143        self.key = nn.Sequential(self.key, SpatialDepthWisePerHeadConvolution(heads, self.d_k))
144        self.value = nn.Sequential(self.value, SpatialDepthWisePerHeadConvolution(heads, self.d_k))