Replication of STOA for Reuters Dataset (#152)

* Add ReutersTrainer, ReutersEvaluator options in Factory classes

* Add Reuters to Kim-CNN command line arguments

* Fix SST dataset path according to changes in Kim-CNN args

The dataset path in args.py was made to point at the dataset folder rather than dataset/SST folder. Hence SST folder was added to paths in the SST dataset class

* Add Reuters dataset class, and support in __main__

* Add Reuters dataset trainers and evaluators

* Remove debug print statement in reuters_evaluator

* Fix rounding bug in reuters_trainer and reuters_evaluator

* Add LSTM for baseline text classification measurements

* Add eval metrics for lstm_baseline

* Set batch_first param in lstm_baseline

* Remove onnx args from lstm_baseline

* Pack padded sequences in LSTM_baseline

* Add TensorBoardX support for Reuters trainer

* Add Arxiv Academic Paper Dataset (AAPD)

* Add Hidden Bottleneck Layer to BiLSTM

* Fix packing of padded tensors in Reuters

* Add cmdline args for Hidden Bottleneck Layer for BiLSTM

* Include pre-padding lengths in AAPD dataset

* Remove duplication of preprocessing code in AAPD

* Remove batch_size condition in ReutersTrainer

common/dataset.py

@@ -12,6 +12,7 @@ from datasets.snli import SNLI
 from datasets.sts2014 import STS2014
 from datasets.quora import Quora
 from datasets.reuters import Reuters
+from datasets.aapd import AAPD
 
 class UnknownWordVecCache(object):
     """
@@ -92,6 +93,11 @@ class DatasetFactory(object):
             train_loader, dev_loader, test_loader = Reuters.iters(dataset_root, word_vectors_file, word_vectors_dir, batch_size, device=device, unk_init=UnknownWordVecCache.unk)
             embedding = nn.Embedding.from_pretrained(Reuters.TEXT_FIELD.vocab.vectors)
             return Reuters, embedding, train_loader, test_loader, dev_loader
+        elif dataset_name == 'aapd':
+            dataset_root = os.path.join(castor_dir, os.pardir, 'Castor-data', 'datasets', 'AAPD/')
+            train_loader, dev_loader, test_loader = AAPD.iters(dataset_root, word_vectors_file, word_vectors_dir, batch_size, device=device, unk_init=UnknownWordVecCache.unk)
+            embedding = nn.Embedding.from_pretrained(AAPD.TEXT_FIELD.vocab.vectors)
+            return AAPD, embedding, train_loader, test_loader, dev_loader
         else:
             raise ValueError('{} is not a valid dataset.'.format(dataset_name))
 

common/evaluation.py

@@ -26,6 +26,7 @@ class EvaluatorFactory(object):
         'pit2015': PIT2015Evaluator,
         'twitterurl': PIT2015Evaluator,
         'Reuters': ReutersEvaluator,
+        'AAPD': ReutersEvaluator,
         'SNLI': SNLIEvaluator,
         'sts2014': STS2014Evaluator,
         'Quora': QuoraEvaluator

common/evaluators/reuters_evaluator.py

@@ -14,11 +14,14 @@ class ReutersEvaluator(Evaluator):
         total_loss = 0
 
         for batch_idx, batch in enumerate(self.data_loader):
-            scores = self.model(batch.text)
+            scores = self.model(batch.text[0], lengths=batch.text[1])
+            scores_rounded = F.sigmoid(scores).round().long()
+
             # Using binary accuracy
-            for tensor1, tensor2 in zip(F.sigmoid(scores).round().long(), batch.label):
+            for tensor1, tensor2 in zip(scores_rounded, batch.label):
                 if np.array_equal(tensor1, tensor2):
                     n_dev_correct += 1
+
             total_loss += F.binary_cross_entropy_with_logits(scores, batch.label.float(), size_average=False).item()
 
         accuracy = 100. * n_dev_correct / len(self.data_loader.dataset.examples)

common/train.py

@@ -26,6 +26,7 @@ class TrainerFactory(object):
         'pit2015': PIT2015Trainer,
         'twitterurl': PIT2015Trainer,
         'Reuters': ReutersTrainer,
+        'AAPD': ReutersTrainer,
         'snli': SNLITrainer,
         'sts2014': STS2014Trainer,
         'quora': QuoraTrainer

common/trainers/reuters_trainer.py

@@ -1,12 +1,13 @@
 import time
-import os
 
+import datetime
+import numpy as np
+import os
 import torch
 import torch.nn.functional as F
-import numpy as np
+from tensorboardX import SummaryWriter
 
 from .trainer import Trainer
-from utils.serialization import save_checkpoint
 
 
 class ReutersTrainer(Trainer):
@@ -21,6 +22,7 @@ class ReutersTrainer(Trainer):
         self.log_template = ' '.join(
             '{:>6.0f},{:>5.0f},{:>9.0f},{:>5.0f}/{:<5.0f} {:>7.0f}%,{:>8.6f},{},{:12.4f},{}'.split(','))
         self.dev_log_template = ' '.join('{:>6.0f},{:>5.0f},{:>9.0f},{:>5.0f}/{:<5.0f} {:>7.0f}%,{:>8.6f},{:8.6f},{:12.4f},{:12.4f}'.split(','))
+        self.writer = SummaryWriter(log_dir="tensorboard_logs/" + datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S"))
 
     def train_epoch(self, epoch):
         self.train_loader.init_epoch()
@@ -29,7 +31,7 @@ class ReutersTrainer(Trainer):
             self.iterations += 1
             self.model.train()
             self.optimizer.zero_grad()
-            scores = self.model(batch.text)
+            scores = self.model(batch.text[0], lengths=batch.text[1])
             # Using binary accuracy
             for tensor1, tensor2 in zip(F.sigmoid(scores).round().long(), batch.label):
                 if np.array_equal(tensor1, tensor2):
@@ -44,6 +46,11 @@ class ReutersTrainer(Trainer):
             # Evaluate performance on validation set
             if self.iterations % self.dev_log_interval == 1:
                 dev_acc, dev_loss = self.dev_evaluator.get_scores()[0]
+                niter = epoch * len(self.train_loader) + batch_idx
+                self.writer.add_scalar('Train/Loss', loss.data[0], niter)
+                self.writer.add_scalar('Dev/Loss', dev_loss, niter)
+                self.writer.add_scalar('Train/Accuracy', train_acc, niter)
+                self.writer.add_scalar('Dev/Accuracy', dev_acc, niter)
                 print(self.dev_log_template.format(time.time() - self.start,
                       epoch, self.iterations, 1 + batch_idx, len(self.train_loader),
                       100. * (1 + batch_idx) / len(self.train_loader), loss.item(),