-update DPCNN & train script

-use spacy tokenizer for yelp data
-add set_rng_seed

fastNLP/modules/aggregator/attention.py

@@ -19,7 +19,7 @@ class DotAttention(nn.Module):
         补上文档
     """
     
-    def __init__(self, key_size, value_size, dropout=0):
+    def __init__(self, key_size, value_size, dropout=0.0):
         super(DotAttention, self).__init__()
         self.key_size = key_size
         self.value_size = value_size
@@ -37,7 +37,7 @@ class DotAttention(nn.Module):
         """
         output = torch.matmul(Q, K.transpose(1, 2)) / self.scale
         if mask_out is not None:
-            output.masked_fill_(mask_out, -1e8)
+            output.masked_fill_(mask_out, -1e18)
         output = self.softmax(output)
         output = self.drop(output)
         return torch.matmul(output, V)
@@ -67,9 +67,8 @@ class MultiHeadAttention(nn.Module):
         self.k_in = nn.Linear(input_size, in_size)
         self.v_in = nn.Linear(input_size, in_size)
         # follow the paper, do not apply dropout within dot-product
-        self.attention = DotAttention(key_size=key_size, value_size=value_size, dropout=0)
+        self.attention = DotAttention(key_size=key_size, value_size=value_size, dropout=dropout)
         self.out = nn.Linear(value_size * num_head, input_size)
-        self.drop = TimestepDropout(dropout)
         self.reset_parameters()
     
     def reset_parameters(self):
@@ -105,7 +104,7 @@ class MultiHeadAttention(nn.Module):
         
         # concat all heads, do output linear
         atte = atte.permute(1, 2, 0, 3).contiguous().view(batch, sq, -1)
-        output = self.drop(self.out(atte))
+        output = self.out(atte)
         return output
 
 

reproduction/text_classification/data/yelpLoader.py

@@ -8,11 +8,20 @@ from fastNLP.io.base_loader import DataInfo
 from fastNLP.io.embed_loader import EmbeddingOption
 from fastNLP.io.file_reader import _read_json
 from typing import Union, Dict
-from reproduction.Star_transformer.datasets import EmbedLoader
 from reproduction.utils import check_dataloader_paths
 
 
-def clean_str(sentence, char_lower=False):
+def get_tokenizer():
+    try:
+        import spacy
+        en = spacy.load('en')
+        print('use spacy tokenizer')
+        return lambda x: [w.text for w in en.tokenizer(x)]
+    except Exception as e:
+        print('use raw tokenizer')
+        return lambda x: x.split()
+
+def clean_str(sentence, tokenizer, char_lower=False):
     """
     heavily borrowed from github
     https://github.com/LukeZhuang/Hierarchical-Attention-Network/blob/master/yelp-preprocess.ipynb
@@ -23,7 +32,7 @@ def clean_str(sentence, char_lower=False):
         sentence = sentence.lower()
     import re
     nonalpnum = re.compile('[^0-9a-zA-Z?!\']+')
-    words = sentence.split()
+    words = tokenizer(sentence)
     words_collection = []
     for word in words:
         if word in ['-lrb-', '-rrb-', '<sssss>', '-r', '-l', 'b-']:
@@ -65,6 +74,7 @@ class yelpLoader(JsonLoader):
         self.fine_grained = fine_grained
         self.tag_v = tag_v
         self.lower = lower
+        self.tokenizer = get_tokenizer()
 
     '''
     def _load_json(self, path):
@@ -109,7 +119,7 @@ class yelpLoader(JsonLoader):
             all_count += 1
             if len(row) == 2:
                 target = self.tag_v[row[0] + ".0"]
-                words = clean_str(row[1], self.lower)
+                words = clean_str(row[1], self.tokenizer, self.lower)
                 if len(words) != 0:
                     ds.append(Instance(words=words, target=target))
                     real_count += 1

reproduction/text_classification/model/dpcnn.py

@@ -3,22 +3,27 @@ import torch.nn as nn
 from fastNLP.modules.utils import get_embeddings
 from fastNLP.core import Const as C
 
+
 class DPCNN(nn.Module):
-    def __init__(self, init_embed, num_cls, n_filters=256, kernel_size=3, n_layers=7, embed_dropout=0.1, dropout=0.1):
+    def __init__(self, init_embed, num_cls, n_filters=256,
+                 kernel_size=3, n_layers=7, embed_dropout=0.1, cls_dropout=0.1):
         super().__init__()
-        self.region_embed = RegionEmbedding(init_embed, out_dim=n_filters, kernel_sizes=[3, 5, 9])
+        self.region_embed = RegionEmbedding(
+            init_embed, out_dim=n_filters, kernel_sizes=[1, 3, 5])
         embed_dim = self.region_embed.embedding_dim
         self.conv_list = nn.ModuleList()
         for i in range(n_layers):
             self.conv_list.append(nn.Sequential(
                 nn.ReLU(),
-                nn.Conv1d(n_filters, n_filters, kernel_size, padding=kernel_size//2),
-                nn.Conv1d(n_filters, n_filters, kernel_size, padding=kernel_size//2),
+                nn.Conv1d(n_filters, n_filters, kernel_size,
+                          padding=kernel_size//2),
+                nn.Conv1d(n_filters, n_filters, kernel_size,
+                          padding=kernel_size//2),
             ))
         self.pool = nn.MaxPool1d(kernel_size=3, stride=2, padding=1)
         self.embed_drop = nn.Dropout(embed_dropout)
         self.classfier = nn.Sequential(
-            nn.Dropout(dropout),
+            nn.Dropout(cls_dropout),
             nn.Linear(n_filters, num_cls),
         )
         self.reset_parameters()
@@ -57,7 +62,8 @@ class RegionEmbedding(nn.Module):
         super().__init__()
         if kernel_sizes is None:
             kernel_sizes = [5, 9]
-        assert isinstance(kernel_sizes, list), 'kernel_sizes should be List(int)'
+        assert isinstance(
+            kernel_sizes, list), 'kernel_sizes should be List(int)'
         self.embed = get_embeddings(init_embed)
         try:
             embed_dim = self.embed.embedding_dim
@@ -69,14 +75,14 @@ class RegionEmbedding(nn.Module):
                 nn.Conv1d(embed_dim, embed_dim, ksz, padding=ksz // 2),
             ))
         self.linears = nn.ModuleList([nn.Conv1d(embed_dim, out_dim, 1)
-                                      for _ in range(len(kernel_sizes) + 1)])
+                                      for _ in range(len(kernel_sizes))])
         self.embedding_dim = embed_dim
 
     def forward(self, x):
         x = self.embed(x)
         x = x.transpose(1, 2)
         # B, C, L
-        out = self.linears[0](x)
+        out = 0
         for conv, fc in zip(self.region_embeds, self.linears[1:]):
             conv_i = conv(x)
             out = out + fc(conv_i)

reproduction/text_classification/train_dpcnn.py

@@ -1,40 +1,44 @@
 # 首先需要加入以下的路径到环境变量，因为当前只对内部测试开放，所以需要手动申明一下路径
 
-from torch.optim.lr_scheduler import CosineAnnealingLR
 import torch.cuda
+from fastNLP.core.utils import cache_results
 from torch.optim import SGD
+from torch.optim.lr_scheduler import CosineAnnealingLR, LambdaLR
 from fastNLP.core.trainer import Trainer
 from fastNLP import CrossEntropyLoss, AccuracyMetric
 from fastNLP.modules.encoder.embedding import StaticEmbedding, CNNCharEmbedding, StackEmbedding
 from reproduction.text_classification.model.dpcnn import DPCNN
-from .data.yelpLoader import yelpLoader
-from fastNLP.io.dataset_loader import SSTLoader
+from data.yelpLoader import yelpLoader
 import torch.nn as nn
 from fastNLP.core import LRScheduler
 from fastNLP.core.const import Const as C
-import sys
+from fastNLP.core.vocabulary import VocabularyOption
+from utils.util_init import set_rng_seeds
 import os
 os.environ['FASTNLP_BASE_URL'] = 'http://10.141.222.118:8888/file/download/'
 os.environ['FASTNLP_CACHE_DIR'] = '/remote-home/hyan01/fastnlp_caches'
 os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
 
-sys.path.append('../..')
-
 
 # hyper
 
 class Config():
-    model_dir_or_name = "en-base-uncased"
-    embedding_grad = False,
+    seed = 12345
+    model_dir_or_name = "dpcnn-yelp-p"
+    embedding_grad = True
     train_epoch = 30
     batch_size = 100
     num_classes = 2
     task = "yelp_p"
     #datadir = '/remote-home/yfshao/workdir/datasets/SST'
-    datadir = '/remote-home/ygwang/yelp_polarity'
+    datadir = '/remote-home/yfshao/workdir/datasets/yelp_polarity'
     #datafile = {"train": "train.txt", "dev": "dev.txt", "test": "test.txt"}
     datafile = {"train": "train.csv",  "test": "test.csv"}
     lr = 1e-3
+    src_vocab_op = VocabularyOption()
+    embed_dropout = 0.3
+    cls_dropout = 0.1
+    weight_decay = 1e-4
 
     def __init__(self):
         self.datapath = {k: os.path.join(self.datadir, v)
@@ -43,15 +47,23 @@ class Config():
 
 ops = Config()
 
+set_rng_seeds(ops.seed)
+print('RNG SEED: {}'.format(ops.seed))
 
 # 1.task相关信息：利用dataloader载入dataInfo
 
 #datainfo=SSTLoader(fine_grained=True).process(paths=ops.datapath, train_ds=['train'])
-datainfo = yelpLoader(fine_grained=True, lower=True).process(
-    paths=ops.datapath, train_ds=['train'])
-print(len(datainfo.datasets['train']))
-print(len(datainfo.datasets['test']))
-
+@cache_results(ops.model_dir_or_name+'-data-cache')
+def load_data():
+    datainfo = yelpLoader(fine_grained=True, lower=True).process(
+        paths=ops.datapath, train_ds=['train'], src_vocab_op=ops.src_vocab_op)
+    for ds in datainfo.datasets.values():
+        ds.apply_field(len, C.INPUT, C.INPUT_LEN)
+        ds.set_input(C.INPUT, C.INPUT_LEN)
+        ds.set_target(C.TARGET)
+    return datainfo
+
+datainfo = load_data()
 
 # 2.或直接复用fastNLP的模型
 
@@ -59,43 +71,50 @@ vocab = datainfo.vocabs['words']
 # embedding = StackEmbedding([StaticEmbedding(vocab), CNNCharEmbedding(vocab, 100)])
 #embedding = StaticEmbedding(vocab)
 embedding = StaticEmbedding(
-    vocab, model_dir_or_name='en-word2vec-300', requires_grad=True)
+    vocab, model_dir_or_name='en-word2vec-300', requires_grad=ops.embedding_grad,
+    normalize=False
+)
+
+print(len(datainfo.datasets['train']))
+print(len(datainfo.datasets['test']))
+print(datainfo.datasets['train'][0])
 
 print(len(vocab))
 print(len(datainfo.vocabs['target']))
 
-model = DPCNN(init_embed=embedding, num_cls=ops.num_classes)
-
+model = DPCNN(init_embed=embedding, num_cls=ops.num_classes,
+              embed_dropout=ops.embed_dropout, cls_dropout=ops.cls_dropout)
+print(model)
 
 # 3. 声明loss,metric,optimizer
 loss = CrossEntropyLoss(pred=C.OUTPUT, target=C.TARGET)
 metric = AccuracyMetric(pred=C.OUTPUT, target=C.TARGET)
 optimizer = SGD([param for param in model.parameters() if param.requires_grad == True],
-                lr=ops.lr, momentum=0.9, weight_decay=0)
+                lr=ops.lr, momentum=0.9, weight_decay=ops.weight_decay)
 
 callbacks = []
 callbacks.append(LRScheduler(CosineAnnealingLR(optimizer, 5)))
+# callbacks.append
+#     LRScheduler(LambdaLR(optimizer, lambda epoch: ops.lr if epoch <
+#                          ops.train_epoch * 0.8 else ops.lr * 0.1))
+# )
+
+# callbacks.append(
+#     FitlogCallback(data=datainfo.datasets, verbose=1)
+# )
 
 device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
 
 print(device)
 
-for ds in datainfo.datasets.values():
-    ds.apply_field(len, C.INPUT, C.INPUT_LEN)
-    ds.set_input(C.INPUT, C.INPUT_LEN)
-    ds.set_target(C.TARGET)
-
-
 # 4.定义train方法
-def train(model, datainfo, loss, metrics, optimizer, num_epochs=ops.train_epoch):
-    trainer = Trainer(datainfo.datasets['train'], model, optimizer=optimizer, loss=loss,
-                      metrics=[metrics],
-                      dev_data=datainfo.datasets['test'], device=device,
-                      check_code_level=-1, batch_size=ops.batch_size, callbacks=callbacks,
-                      n_epochs=num_epochs)
+trainer = Trainer(datainfo.datasets['train'], model, optimizer=optimizer, loss=loss,
+                  metrics=[metric],
+                  dev_data=datainfo.datasets['test'], device=device,
+                  check_code_level=-1, batch_size=ops.batch_size, callbacks=callbacks,
+                  n_epochs=ops.train_epoch, num_workers=4)
 
-    print(trainer.train())
 
 
 if __name__ == "__main__":
-    train(model, datainfo, loss, metric, optimizer)
+    print(trainer.train())

reproduction/text_classification/utils/util_init.py

@@ -0,0 +1,11 @@
+import numpy
+import torch
+import random
+
+
+def set_rng_seeds(seed):
+    random.seed(seed)
+    numpy.random.seed(seed)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    # print('RNG_SEED {}'.format(seed))