Add README (#110)

* Make *QA/MSRVID work with VDPWI

* Add README

common/trainers/msrvid_trainer.py

@@ -56,7 +56,8 @@ class MSRVIDTrainer(Trainer):
         return left_out_val_a, left_out_val_b, left_out_val_ext_feats, left_out_val_labels
 
     def train(self, epochs):
-        scheduler = ReduceLROnPlateau(self.optimizer, mode='max', factor=self.lr_reduce_factor, patience=self.patience)
+        if self.lr_reduce_factor != 1 and self.lr_reduce_factor != None:
+            scheduler = ReduceLROnPlateau(self.optimizer, mode='max', factor=self.lr_reduce_factor, patience=self.patience)
         epoch_times = []
         prev_loss = -1
         best_dev_score = -1
@@ -100,7 +101,8 @@ class MSRVIDTrainer(Trainer):
                     self.writer.add_scalar('msrvid/dev/kl_div_loss', val_kl_div_loss, epoch)
                 break
 
-            scheduler.step(pearson_r)
+            if scheduler is not None:
+                scheduler.step(pearson_r)
 
             end = time.time()
             duration = end - start

common/trainers/qa_trainer.py

@@ -40,7 +40,9 @@ class QATrainer(Trainer):
         return total_loss
 
     def train(self, epochs):
-        scheduler = ReduceLROnPlateau(self.optimizer, mode='max', factor=self.lr_reduce_factor, patience=self.patience)
+        scheduler = None
+        if self.lr_reduce_factor != 1 and self.lr_reduce_factor != None:
+            scheduler = ReduceLROnPlateau(self.optimizer, mode='max', factor=self.lr_reduce_factor, patience=self.patience)
         epoch_times = []
         prev_loss = -1
         best_dev_score = -1
@@ -72,6 +74,7 @@ class QATrainer(Trainer):
                 break
 
             prev_loss = new_loss
-            scheduler.step(mean_average_precision)
+            if scheduler is not None:
+                scheduler.step(mean_average_precision)
 
         self.logger.info('Training took {:.2f} minutes overall...'.format(sum(epoch_times) / 60))

vdpwi/README.md

@@ -0,0 +1,48 @@
+# VDPWI PyTorch Implementation
+
+This is a PyTorch implementation of the following paper
+
+* Hua He and Jimmy Lin. [Pairwise Word Interaction Modeling with Deep Neural Networks for Semantic Similarity Measurement.](http://www.aclweb.org/anthology/N16-1108) *Proceedings of the 15th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL/HLT 2016)*, pages 937-948.
+
+
+Please ensure you have followed instructions in the main [README](../README.md) doc before running any further commands in this doc.
+
+## SICK Dataset
+
+To run VDPWI on the SICK dataset, use the following command. If you have any problems running it check the Troubleshooting section below.
+
+```
+python -m vdpwi vdpwi.sick.model.castor --dataset sick --epochs 19 --epsilon 1e-7
+```
+
+## MSRVID Dataset
+
+To run VDPWI on the MSRVID dataset, use the following command:
+```
+python -m vdpwi vdpwi.msrvid.model.castor --dataset msrvid --batch-size 16 --epochs 32 --regularization 0.0025
+```
+
+## TrecQA Dataset
+
+To run VDPWI on (Raw) TrecQA, you first need to run `./get_trec_eval.sh` in `utils` under the repo root while inside the `utils` directory. This will download and compile the official `trec_eval` tool used for evaluation.
+
+Then, you can run:
+```
+python -m vdpwi vdpwi.trecqa.model --dataset trecqa --epochs 5 --regularization 0.0005 --eps 0.1
+```
+
+The paper results are reported in [Noise-Contrastive Estimation for Answer Selection with Deep Neural Networks](https://dl.acm.org/citation.cfm?id=2983872).
+
+## WikiQA Dataset
+
+You also need `trec_eval` for this dataset, similar to TrecQA.
+
+Then, you can run:
+```
+python -m vdpwi vdpwi.wikiqa.model --epochs 10 --dataset wikiqa --batch-size 64 --lr 0.0004 --regularization 0.02
+```
+
+To see all options available, use
+```
+python -m vdpwi --help
+```

vdpwi/__main__.py

@@ -33,17 +33,17 @@ if __name__ == '__main__':
     parser.add_argument('--skip-training', help='will load pre-trained model', action='store_true')
     parser.add_argument('--device', type=int, default=0, help='GPU device, -1 for CPU (default: 0)')
     parser.add_argument('--sparse-features', action='store_true', default=False, help='use sparse features (default: false)')
-    parser.add_argument('--batch-size', type=int, default=64, help='input batch size for training (default: 64)')
+    parser.add_argument('--batch-size', type=int, default=16, help='input batch size for training (default: 64)')
     parser.add_argument('--epochs', type=int, default=10, help='number of epochs to train (default: 10)')
-    parser.add_argument('--optimizer', type=str, default='adam', help='optimizer to use: adam or sgd (default: adam)')
+    parser.add_argument('--optimizer', type=str, default='rmsprop', help='optimizer to use: adam, sgd, or rmsprop (default: adam)')
     parser.add_argument('--lr', type=float, default=5E-4, help='learning rate (default: 0.001)')
-    parser.add_argument('--lr-reduce-factor', type=float, default=1, help='learning rate reduce factor after plateau (default: 0.3)')
+    parser.add_argument('--lr-reduce-factor', type=float, default=0.3, help='learning rate reduce factor after plateau (default: 0.3)')
     parser.add_argument('--patience', type=float, default=2, help='learning rate patience after seeing plateau (default: 2)')
     parser.add_argument('--momentum', type=float, default=0.1, help='momentum (default: 0.1)')
     parser.add_argument('--epsilon', type=float, default=1e-8, help='Adam epsilon (default: 1e-8)')
     parser.add_argument('--log-interval', type=int, default=10, help='how many batches to wait before logging training status (default: 10)')
     parser.add_argument('--regularization', type=float, default=1E-5, help='Regularization for the optimizer (default: 0.00001)')
-    parser.add_argument('--hidden-units', type=int, default=150, help='number of hidden units in the RNN')
+    parser.add_argument('--hidden-units', type=int, default=250, help='number of hidden units in the RNN')
     parser.add_argument('--seed', type=int, default=1, help='random seed (default: 1)')
     parser.add_argument('--tensorboard', action='store_true', default=False, help='use TensorBoard to visualize training (default: false)')
     parser.add_argument('--run-label', type=str, help='label to describe run')
@@ -98,7 +98,7 @@ if __name__ == '__main__':
     elif args.optimizer == 'sgd':
         optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.regularization)
     elif args.optimizer == "rmsprop":
-        optimizer = optim.RMSprop(model.parameters(), lr=args.lr, momentum=args.momentum, alpha=config.decay, 
+        optimizer = optim.RMSprop(model.parameters(), lr=args.lr, momentum=args.momentum, alpha=args.decay, 
             weight_decay=args.regularization)
     else:
         raise ValueError('optimizer not recognized: it should be one of adam, sgd, or rmsprop')

vdpwi/model.py

@@ -32,7 +32,7 @@ class ResNet(nn.Module):
                 x += old_x
                 old_x = x
         x = torch.mean(x.view(x.size(0), x.size(1), -1), 2)
-        return self.output(x)
+        return F.log_softmax(self.output(x), 1)
 
 class VDPWIConvNet(nn.Module):
     def __init__(self, config):