- add star-transformer

fastNLP/modules/encoder/star_transformer.py

@@ -0,0 +1,146 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+import numpy as NP
+
+
+class StarTransformer(nn.Module):
+    """Star-Transformer Encoder part。
+    paper: https://arxiv.org/abs/1902.09113
+
+    :param hidden_size: int, 输入维度的大小。同时也是输出维度的大小。
+    :param num_layers: int, star-transformer的层数
+    :param num_head: int，head的数量。
+    :param head_dim: int, 每个head的维度大小。
+    :param dropout: float dropout 概率
+    :param max_len: int or None, 如果为int，输入序列的最大长度，
+                    模型会为属于序列加上position embedding。
+                    若为None，忽略加上position embedding的步骤
+    """
+    def __init__(self, hidden_size, num_layers, num_head, head_dim, dropout=0.1, max_len=None):
+        super(StarTransformer, self).__init__()
+        self.iters = num_layers
+
+        self.norm = nn.ModuleList([nn.LayerNorm(hidden_size) for _ in range(self.iters)])
+        self.ring_att = nn.ModuleList(
+            [MSA1(hidden_size, nhead=num_head, head_dim=head_dim, dropout=dropout)
+                for _ in range(self.iters)])
+        self.star_att = nn.ModuleList(
+            [MSA2(hidden_size, nhead=num_head, head_dim=head_dim, dropout=dropout)
+                for _ in range(self.iters)])
+
+        if max_len is not None:
+            self.pos_emb = self.pos_emb = nn.Embedding(max_len, hidden_size)
+        else:
+            self.pos_emb = None
+
+    def forward(self, data, mask):
+        """
+        :param FloatTensor data: [batch, length, hidden] the input sequence
+        :param ByteTensor mask: [batch, length] the padding mask for input, in which padding pos is 0
+        :return: [batch, length, hidden] the output sequence
+                 [batch, hidden] the global relay node
+        """
+        def norm_func(f, x):
+            # B, H, L, 1
+            return f(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
+
+        B, L, H = data.size()
+        smask = torch.cat([torch.zeros(B, 1, ).byte().to(mask), mask], 1)
+
+        embs = data.permute(0, 2, 1)[:,:,:,None] # B H L 1
+        if self.pos_emb:
+            P = self.pos_emb(torch.arange(L, dtype=torch.long, device=embs.device)\
+                    .view(1, L)).permute(0, 2, 1).contiguous()[:, :, :, None]  # 1 H L 1
+            embs = embs + P
+
+        nodes = embs
+        relay = embs.mean(2, keepdim=True)
+        ex_mask = mask[:, None, :, None].expand(B, H, L, 1)
+        r_embs = embs.view(B, H, 1, L)
+        for i in range(self.iters):
+            ax = torch.cat([r_embs, relay.expand(B, H, 1, L)], 2)
+            nodes = nodes + F.leaky_relu(self.ring_att[i](norm_func(self.norm[i], nodes), ax=ax))
+            relay = F.leaky_relu(self.star_att[i](relay, torch.cat([relay, nodes], 2), smask))
+
+            nodes = nodes.masked_fill_(ex_mask, 0)
+
+        nodes = nodes.view(B, H, L).permute(0, 2, 1)
+
+        return nodes, relay.view(B, H)
+
+
+class MSA1(nn.Module):
+    def __init__(self, nhid, nhead=10, head_dim=10, dropout=0.1):
+        super(MSA1, self).__init__()
+        # Multi-head Self Attention Case 1, doing self-attention for small regions
+        # Due to the architecture of GPU, using hadamard production and summation are faster than dot production when unfold_size is very small
+        self.WQ = nn.Conv2d(nhid, nhead * head_dim, 1)
+        self.WK = nn.Conv2d(nhid, nhead * head_dim, 1)
+        self.WV = nn.Conv2d(nhid, nhead * head_dim, 1)
+        self.WO = nn.Conv2d(nhead * head_dim, nhid, 1)
+
+        self.drop = nn.Dropout(dropout)
+
+        # print('NUM_HEAD', nhead, 'DIM_HEAD', head_dim)
+        self.nhid, self.nhead, self.head_dim, self.unfold_size = nhid, nhead, head_dim, 3
+
+    def forward(self, x, ax=None):
+        # x: B, H, L, 1, ax : B, H, X, L append features
+        nhid, nhead, head_dim, unfold_size = self.nhid, self.nhead, self.head_dim, self.unfold_size
+        B, H, L, _ = x.shape
+
+        q, k, v = self.WQ(x), self.WK(x), self.WV(x)  # x: (B,H,L,1)
+
+        if ax is not None:
+            aL = ax.shape[2]
+            ak = self.WK(ax).view(B, nhead, head_dim, aL, L)
+            av = self.WV(ax).view(B, nhead, head_dim, aL, L)
+        q = q.view(B, nhead, head_dim, 1, L)
+        k = F.unfold(k.view(B, nhead * head_dim, L, 1), (unfold_size, 1), padding=(unfold_size // 2, 0))\
+                .view(B, nhead, head_dim, unfold_size, L)
+        v = F.unfold(v.view(B, nhead * head_dim, L, 1), (unfold_size, 1), padding=(unfold_size // 2, 0))\
+                .view(B, nhead, head_dim, unfold_size, L)
+        if ax is not None:
+            k = torch.cat([k, ak], 3)
+            v = torch.cat([v, av], 3)
+
+        alphas = self.drop(F.softmax((q * k).sum(2, keepdim=True) / NP.sqrt(head_dim), 3))  # B N L 1 U
+        att = (alphas * v).sum(3).view(B, nhead * head_dim, L, 1)
+
+        ret = self.WO(att)
+
+        return ret
+
+
+class MSA2(nn.Module):
+    def __init__(self, nhid, nhead=10, head_dim=10, dropout=0.1):
+        # Multi-head Self Attention Case 2, a broadcastable query for a sequence key and value
+        super(MSA2, self).__init__()
+        self.WQ = nn.Conv2d(nhid, nhead * head_dim, 1)
+        self.WK = nn.Conv2d(nhid, nhead * head_dim, 1)
+        self.WV = nn.Conv2d(nhid, nhead * head_dim, 1)
+        self.WO = nn.Conv2d(nhead * head_dim, nhid, 1)
+
+        self.drop = nn.Dropout(dropout)
+
+        # print('NUM_HEAD', nhead, 'DIM_HEAD', head_dim)
+        self.nhid, self.nhead, self.head_dim, self.unfold_size = nhid, nhead, head_dim, 3
+
+    def forward(self, x, y, mask=None):
+        # x: B, H, 1, 1, 1 y: B H L 1
+        nhid, nhead, head_dim, unfold_size = self.nhid, self.nhead, self.head_dim, self.unfold_size
+        B, H, L, _ = y.shape
+
+        q, k, v = self.WQ(x), self.WK(y), self.WV(y)
+
+        q = q.view(B, nhead, 1, head_dim)  # B, H, 1, 1 -> B, N, 1, h
+        k = k.view(B, nhead, head_dim, L)  # B, H, L, 1 -> B, N, h, L
+        v = k.view(B, nhead, head_dim, L).permute(0, 1, 3, 2)  # B, H, L, 1 -> B, N, L, h
+        pre_a = torch.matmul(q, k) / NP.sqrt(head_dim)
+        if mask is not None:
+            pre_a = pre_a.masked_fill(mask[:, None, None, :], -float('inf'))
+        alphas = self.drop(F.softmax(pre_a, 3))  # B, N, 1, L
+        att = torch.matmul(alphas, v).view(B, -1, 1, 1)  # B, N, 1, h -> B, N*h, 1, 1
+        return self.WO(att)
+

test/modules/test_other_modules.py

@@ -3,6 +3,7 @@ import unittest
 import torch
 
 from fastNLP.modules.other_modules import GroupNorm, LayerNormalization, BiLinear, BiAffine
+from fastNLP.modules.encoder.star_transformer import StarTransformer
 
 
 class TestGroupNorm(unittest.TestCase):
@@ -49,3 +50,12 @@ class TestBiAffine(unittest.TestCase):
         encoder_input = torch.randn((batch_size, decoder_length, 10))
         y = layer(decoder_input, encoder_input)
         self.assertEqual(tuple(y.shape), (batch_size, 25, encoder_length, 1))
+
+class TestStarTransformer(unittest.TestCase):
+    def test_1(self):
+        model = StarTransformer(num_layers=6, hidden_size=100, num_head=8, head_dim=20, max_len=100)
+        x = torch.rand(16, 45, 100)
+        mask = torch.ones(16, 45).byte()
+        y, yn = model(x, mask)
+        self.assertEqual(tuple(y.size()), (16, 45, 100))
+        self.assertEqual(tuple(yn.size()), (16, 100))