[FEATURE]basic QANet

Author: CoyoteLeo

config.py

@@ -12,6 +12,7 @@
 SQUAD_VERSION = 'v1.1'
 flags.DEFINE_string('squad_version', SQUAD_VERSION, '')
 flags.DEFINE_string("mode", "debug", "train/debug/test")
+flags.DEFINE_string("run_name", "", "")
 
 # data
 DATA_DIR = os.path.join(BASE_DIR, 'data', 'squad', SQUAD_VERSION)
@@ -47,9 +48,11 @@
 flags.DEFINE_integer('emb_encoder_conv_num', 4, "")
 flags.DEFINE_integer('emb_encoder_conv_kernel_size', 7, "")
 flags.DEFINE_integer('emb_encoder_block_num', 1, "")
+flags.DEFINE_integer('emb_encoder_ff_depth', 3, "")
 flags.DEFINE_integer('output_encoder_conv_num', 2, "")
 flags.DEFINE_integer('output_encoder_conv_kernel_size', 5, "")
 flags.DEFINE_integer('output_encoder_block_num', 7, "")
+flags.DEFINE_integer('output_encoder_ff_depth', 2, "")
 flags.DEFINE_integer('attention_head_num', 8, "")
 
 # train & test config
@@ -59,7 +62,7 @@
 flags.DEFINE_integer('checkpoint', 1400, "")
 flags.DEFINE_float('lr', 0.001, "")
 flags.DEFINE_integer('lr_warm_up_steps', 1000, "")
-flags.DEFINE_float('adam_beta1', 0.9, "")
+flags.DEFINE_float('adam_beta1', 0.8, "")
 flags.DEFINE_float('adam_beta2', 0.999, "")
 flags.DEFINE_float('adam_eps', 1e-7, "")
 flags.DEFINE_float('adam_decay', 5e-8, "")
@@ -81,5 +84,3 @@
     os.makedirs(DATA_DIR)
 if not os.path.exists(RESULT_DIR):
     os.makedirs(RESULT_DIR)
-if not os.path.exists(LOG_DIR):
-    os.makedirs(LOG_DIR)

main.py

@@ -1,4 +1,6 @@
 import math
+import os
+
 import numpy as np
 import torch
 import torch.cuda
@@ -192,7 +194,8 @@ def train_entry(config):
 
 
 def main(*args, **kwarg):
-    # runner = Runner(loss=nn.CrossEntropyLoss())
+    if not os.path.exists(LOG_DIR) and config.mode == 'train':
+        os.makedirs(LOG_DIR)
     if config.mode == "data":
         preproc(config)
     elif config.mode == "train":
@@ -206,6 +209,7 @@ def main(*args, **kwarg):
     else:
         print("Unknown mode")
         exit(0)
+    print(config.run_name)
 
 
 if __name__ == '__main__':

models.py

@@ -70,9 +70,9 @@ def forward(self, cemb: torch.Tensor, wemb: torch.Tensor):
         emb = torch.cat((cemb, wemb), dim=1)
         emb = self.highway(emb)
 
-        emb = F.relu(self.resizer(emb.transpose(1, 2)))
-        emb = F.dropout(emb, p=config.word_emb_dropout, training=self.training)
-        emb = self.norm(emb).transpose(1, 2)
+        emb = F.relu(self.norm(self.resizer(emb.transpose(1, 2))))
+        emb = F.dropout(emb, p=config.layer_dropout, training=self.training)
+        emb = emb.transpose(1, 2)
 
         return emb
 
@@ -141,40 +141,6 @@ def mask_logits(target, mask):
     return target + (-1e30) * (1 - mask)
 
 
-class AttentionBlock(nn.Module):
-    def __init__(self, hidden_size, head_number):
-        super(AttentionBlock, self).__init__()
-        self.self_attention = nn.MultiheadAttention(hidden_size, head_number,
-                                                    dropout=config.layer_dropout)
-        self.attention_layer_norm = nn.LayerNorm(hidden_size)
-        self.feedforward_norm1 = nn.LayerNorm(hidden_size)
-        self.feedforward1 = nn.Linear(hidden_size, hidden_size)
-        self.feedforward_norm2 = nn.LayerNorm(hidden_size)
-        self.feedforward2 = nn.Linear(hidden_size, hidden_size)
-        nn.init.kaiming_normal_(self.feedforward1.weight, nonlinearity='relu')
-        nn.init.kaiming_normal_(self.feedforward2.weight, nonlinearity='relu')
-
-    def forward(self, x, mask):
-        raw = x
-        x = x.permute(2, 0, 1)
-        x, _ = self.self_attention(x, x, x, key_padding_mask=(mask.bool() == False))
-        x = x.permute(1, 2, 0)
-        x = F.dropout(x, config.layer_dropout, training=self.training)
-        x = self.attention_layer_norm(raw.transpose(1, 2) + x.transpose(1, 2)).transpose(1, 2)
-
-        raw = x
-        x = self.feedforward1(x.transpose(1, 2)).transpose(1, 2)
-        x = F.dropout(F.relu(x), config.layer_dropout, training=self.training)
-        x = self.feedforward_norm1(raw.transpose(1, 2) + x.transpose(1, 2)).transpose(1, 2)
-
-        raw = x
-        x = self.feedforward2(x.transpose(1, 2)).transpose(1, 2)
-        x = F.dropout(F.relu(x), config.layer_dropout, training=self.training)
-        x = self.feedforward_norm2(raw.transpose(1, 2) + x.transpose(1, 2)).transpose(1, 2)
-
-        return x
-
-
 class EncoderBlock(nn.Module):
     """
     input:
@@ -184,31 +150,66 @@ class EncoderBlock(nn.Module):
         x: shape [batch_size, hidden_size, max length] => [8, 128, 400]
     """
 
-    def __init__(self, conv_number, hidden_size, kernel_size, head_number):
+    def __init__(self, conv_num, hidden_size, kernel_size, head_number, ff_depth):
         super(EncoderBlock, self).__init__()
-        self.conv_number = conv_number
+        self.conv_num = conv_num
+        self.ff_depth = ff_depth
+        self.total_layer = self.conv_num + self.ff_depth + 1  # one => atten
+
         self.position_encoder = PositionEncoder(hidden_size)
+
+        self.conv_norm_list = nn.ModuleList(
+            [nn.LayerNorm(hidden_size) for _ in range(self.conv_num)]
+        )
         self.conv_list = nn.ModuleList([
             DepthwiseSeparableConv(hidden_size, hidden_size, kernel_size)
-            for _ in range(self.conv_number)
+            for _ in range(self.conv_num)
         ])
 
-        self.conv_norm_list = nn.ModuleList(
-            [nn.LayerNorm(hidden_size) for _ in range(self.conv_number)]
+        self.atten_layer_norm = nn.LayerNorm(hidden_size)
+        self.self_atten = nn.MultiheadAttention(
+            hidden_size,
+            head_number,
+            dropout=(1 - (self.conv_num + 1) / self.total_layer) * config.layer_dropout
         )
-        self.self_attention = AttentionBlock(hidden_size, head_number)
+
+        self.ff_norm_list = nn.ModuleList(
+            [nn.LayerNorm(hidden_size) for _ in range(self.ff_depth)]
+        )
+        self.ff = nn.ModuleList(
+            [nn.Linear(hidden_size, hidden_size) for _ in range(self.ff_depth)]
+        )
+        for layer in self.ff:
+            nn.init.kaiming_normal_(layer.weight, nonlinearity='relu')
 
     def forward(self, x, mask):
         x = self.position_encoder(x)
-        for i in range(self.conv_number):
-            raw = x.transpose(1, 2)
-            x = self.conv_list[i](x)
-            x = F.dropout(x, config.layer_dropout * (i + 1) / self.conv_number,
-                          training=self.training)
-            x = self.conv_norm_list[i](x.transpose(1, 2) + raw).transpose(1, 2)
-
-        x = self.self_attention(x, mask)
+        for i in range(self.conv_num):
+            raw = x
+            x = self.conv_list[i](self.conv_norm_list[i](x.transpose(1, 2)).transpose(1, 2))
+            x = F.dropout(
+                input=x,
+                p=(1 - (i + 1) / self.total_layer) * config.layer_dropout,
+                training=self.training
+            )
+            x = raw + x
 
+        raw = x
+        x = self.atten_layer_norm(x.transpose(1, 2)).transpose(1, 2)
+        x = x.permute(2, 0, 1)
+        x, _ = self.self_atten(x, x, x, key_padding_mask=(mask.bool() == False))
+        x = x.permute(1, 2, 0)
+        x = raw + x
+
+        for i in range(self.ff_depth):
+            raw = x
+            x = F.relu(self.ff[i](self.ff_norm_list[i](x.transpose(1, 2)))).transpose(1, 2)
+            x = F.dropout(
+                input=x,
+                p=(1 - (self.conv_num + 1 + (i + 1)) / self.total_layer) * config.layer_dropout,
+                training=self.training
+            )
+            x = raw + x
         return x
 
 
@@ -254,10 +255,7 @@ def forward(self, C, Q, cmask, qmask):
 
         A = torch.bmm(S_row_sofmax, Q)
         B = torch.bmm(torch.bmm(S_row_sofmax, S_column_softmax.transpose(1, 2)), C)
-        output = torch.cat((C, A, torch.mul(C, A), torch.mul(C, B)), dim=2)
-        output = F.dropout(output, p=config.layer_dropout, training=self.training)
-        output = self.resizer(output)
-        output = F.relu(output)
+        output = F.relu(self.resizer(torch.cat((C, A, torch.mul(C, A), torch.mul(C, B)), dim=2)))
         output = F.dropout(output, p=config.layer_dropout, training=self.training)
         output = output.transpose(1, 2)
         return output
@@ -301,19 +299,21 @@ def __init__(self, word_mat, char_mat):
         self.char_embedding = nn.Embedding.from_pretrained(torch.tensor(char_mat), freeze=False)
         self.embedding = Embedding(word_mat.shape[1], char_mat.shape[1], config.global_hidden_size)
         emb_encoder_block = EncoderBlock(
-            conv_number=config.emb_encoder_conv_num,
+            conv_num=config.emb_encoder_conv_num,
             hidden_size=config.global_hidden_size,
             kernel_size=config.emb_encoder_conv_kernel_size,
-            head_number=config.attention_head_num
+            head_number=config.attention_head_num,
+            ff_depth=config.emb_encoder_ff_depth,
         )
         self.emb_encoder = nn.ModuleList(
             [emb_encoder_block for _ in range(config.emb_encoder_block_num)])
         self.cq_attention = CQAttention(hidden_size=config.global_hidden_size)
         output_encoder_block = EncoderBlock(
-            conv_number=config.output_encoder_conv_num,
+            conv_num=config.output_encoder_conv_num,
             hidden_size=config.global_hidden_size,
             kernel_size=config.output_encoder_conv_kernel_size,
-            head_number=config.attention_head_num
+            head_number=config.attention_head_num,
+            ff_depth=config.output_encoder_ff_depth,
         )
         self.output_encoder = nn.ModuleList(
             [output_encoder_block for _ in range(config.output_encoder_block_num)])