relative transformer fix for encoder decoder

transformers/relative_mha.py

@@ -1,66 +1,58 @@
-import copy
+"""
+Implementation of "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+https://arxiv.org/abs/1901.02860
+"""
 
 import torch
 from torch import nn
 
 from labml.helpers.pytorch.module import Module
+from labml.logger import inspect
 from transformers.mha import MultiHeadAttention
 
 
-class PrepareForMultiHeadAttention(Module):
-    def __init__(self, d_model: int, heads: int, d_k: int):
-        super().__init__()
-        self.linear = nn.Linear(d_model, heads * d_k, bias=False)
-        self.heads = heads
-        self.d_k = d_k
+def relative_shift(x: torch.Tensor):
+    zero_pad = x.new_zeros(x.shape[0], 1, *x.shape[2:])
+    x_padded = torch.cat([x, zero_pad], dim=1)
 
-    def __call__(self, x: torch.Tensor):
-        seq_len, batch_size, _ = x.shape
+    x_padded = x_padded.view(x.shape[1] + 1, x.shape[0], *x.shape[2:])
 
-        x = self.linear(x)
-        x = x.view(seq_len, batch_size, self.heads, self.d_k)
+    x = x_padded[:-1].view_as(x)
 
-        return x
+    return x
 
 
 class RelativeMultiHeadAttention(MultiHeadAttention):
-    @staticmethod
-    def _rel_shift(x: torch.Tensor):
-        zero_pad = torch.zeros((x.shape[0], 1, *x.shape[2:]),
-                               device=x.device, dtype=x.dtype)
-        x_padded = torch.cat([zero_pad, x], dim=1)
-
-        x_padded = x_padded.view(x.shape[1] + 1, x.shape[0], *x.shape[2:])
-
-        x = x_padded[1:].view_as(x)
-
-        ones = torch.ones((x.size(0), x.size(1)), device=x.device)
-        lower_triangle = torch.tril(ones, x.size(1) - x.size(0))
-        x = x * lower_triangle[:, :, None, None]
-
-        return x
-
     def __init__(self, heads: int, d_model: int, dropout_prob: float = 0.1):
-        super().__init__(heads, d_model, dropout_prob)
-        self.max_key_len = 2 ** 12
+        super().__init__(heads, d_model, dropout_prob, False)
+        self.P = 2 ** 12
 
-        self.key_pos_embeddings = nn.Parameter(
-            torch.zeros((self.max_key_len, heads, self.d_k)),
-            requires_grad=True)
-        self.query_pos_bias = nn.Parameter(
-            torch.zeros((heads, self.d_k)),
-            requires_grad=True)
-        self.key_pos_bias = nn.Parameter(
-            torch.zeros((self.max_key_len, heads)),
-            requires_grad=True)
+        self.key_pos_embeddings = nn.Parameter(torch.zeros((self.P * 2, heads, self.d_k)), requires_grad=True)
+        self.query_pos_bias = nn.Parameter(torch.zeros((heads, self.d_k)), requires_grad=True)
+        self.key_pos_bias = nn.Parameter(torch.zeros((self.P * 2, heads)), requires_grad=True)
 
-    def get_scores(self, query: torch.Tensor,
-                   key: torch.Tensor, ):
-        key_len = key.shape[0]
+    def get_scores(self, query: torch.Tensor, key: torch.Tensor):
+        key_pos_emb = self.key_pos_embeddings[self.P - query.shape[0]:self.P + key.shape[0]]
+        key_pos_bias = self.key_pos_bias[self.P - query.shape[0]:self.P + key.shape[0]]
 
         ac = torch.einsum('ibhd,jbhd->ijbh', query + self.query_pos_bias[None, None, :, :], key)
-        b = torch.einsum('ibhd,jhd->ijbh', query, self.key_pos_embeddings[-key_len:])
-        d = self.key_pos_bias[None, -key_len:, None, :]
-        bd = self._rel_shift(b + d)
+        b = torch.einsum('ibhd,jhd->ijbh', query, key_pos_emb)
+        d = key_pos_bias[None, :, None, :]
+        bd = relative_shift(b + d)
+        bd = bd[:, -key.shape[0]:]
 
         return ac + bd
+
+
+def _test_relative_shift():
+    x = torch.arange(1, 6)[None, :, None, None].repeat(5, 1, 1, 1)
+    inspect(x[:, :, 0, 0])
+    inspect(relative_shift(x)[:, :, 0, 0])
+
+    x = torch.arange(1, 6)[None, :, None, None].repeat(3, 1, 1, 1)
+    inspect(x[:, :, 0, 0])
+    inspect(relative_shift(x)[:, :, 0, 0])
+
+
+if __name__ == '__main__':
+    _test_relative_shift()