Training results

.gitignore

@@ -161,4 +161,6 @@ cython_debug/
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 #.idea/
 
-zemo*.py
+nohup.out
+*.pt
+predictions/

cfimdb-classifier.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:e1c66df3c0ce0e4326519041f49707f102df5f680de5ded1b5125ba689a9d141
-size 438045778

classifier.py

@@ -48,6 +48,7 @@ class BertSentimentClassifier(torch.nn.Module):
                 param.requires_grad = True
 
         # Create any instance variables you need to classify the sentiment of BERT embeddings.
+        self.dropout = torch.nn.Dropout(config.hidden_dropout_prob)
         self.classifier = torch.nn.Linear(config.hidden_size, self.num_labels)
 
 
@@ -67,7 +68,7 @@ class BertSentimentClassifier(torch.nn.Module):
         cls_token_output = sequence_output[:, 0, :]  # The first token is [CLS]
 
         # Pass the [CLS] token representation through the classifier.
-        logits = self.classifier(cls_token_output)
+        logits = self.classifier(self.dropout(cls_token_output))
 
         return logits
 
@@ -180,7 +181,7 @@ def model_eval(dataloader, model, device):
     y_pred = []
     sents = []
     sent_ids = []
-    for step, batch in enumerate(tqdm(dataloader, desc=f'eval', disable=TQDM_DISABLE)):
+    for step, batch in enumerate(tqdm(dataloader, desc=f'eval', leave=False, disable=TQDM_DISABLE)):
         b_ids, b_mask, b_labels, b_sents, b_sent_ids = batch['token_ids'],batch['attention_mask'],  \
                                                         batch['labels'], batch['sents'], batch['sent_ids']
 
@@ -209,7 +210,7 @@ def model_test_eval(dataloader, model, device):
     y_pred = []
     sents = []
     sent_ids = []
-    for step, batch in enumerate(tqdm(dataloader, desc=f'eval', disable=TQDM_DISABLE)):
+    for step, batch in enumerate(tqdm(dataloader, desc=f'eval', leave=False, disable=TQDM_DISABLE)):
         b_ids, b_mask, b_sents, b_sent_ids = batch['token_ids'],batch['attention_mask'],  \
                                                          batch['sents'], batch['sent_ids']
 
@@ -277,7 +278,7 @@ def train(args):
         model.train()
         train_loss = 0
         num_batches = 0
-        for batch in tqdm(train_dataloader, desc=f'train-{epoch}', disable=TQDM_DISABLE):
+        for batch in tqdm(train_dataloader, desc=f'train-{epoch}', leave=False, disable=TQDM_DISABLE):
             b_ids, b_mask, b_labels = (batch['token_ids'],
                                        batch['attention_mask'], batch['labels'])
 

predictions/last-linear-layer-cfimdb-dev-out.csv

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c3f994587376345ea6a1e80a7946d5889259f6a427989c71e0b45de28ea4545d
-size 7621

predictions/last-linear-layer-cfimdb-test-out.csv

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:7ebedf210c8973e02648e96152e253daa2385b230a48da151812a58d80178536
-size 15154

predictions/last-linear-layer-sst-dev-out.csv

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:22412dead5299ffb8fae45448f240cb135e3ad5dc04cea96975e893bdd719ba8
-size 34157

predictions/last-linear-layer-sst-test-out.csv

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:3455d6637e5ecd118c31e48534d92298da3c865ed11ad93e2aadc09fcc743666
-size 68536

sst-classifier.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:62f6282ea608a997c1b43071cedcb1c4ba454b420305c7b15138aa9d7f70103d
-size 438072793

trainings/last-layer-no-dropout.txt

@@ -0,0 +1,38 @@
+Training Sentiment Classifier on SST...
+load 8544 data from data/ids-sst-train.csv
+load 1101 data from data/ids-sst-dev.csv
+Epoch 0: train loss :: 1.429, train acc :: 0.454, dev acc :: 0.441
+Epoch 1: train loss :: 1.297, train acc :: 0.467, dev acc :: 0.431
+Epoch 2: train loss :: 1.253, train acc :: 0.489, dev acc :: 0.450
+Epoch 3: train loss :: 1.233, train acc :: 0.491, dev acc :: 0.455
+Epoch 4: train loss :: 1.214, train acc :: 0.501, dev acc :: 0.450
+Epoch 5: train loss :: 1.211, train acc :: 0.511, dev acc :: 0.465
+Epoch 6: train loss :: 1.199, train acc :: 0.515, dev acc :: 0.478
+Epoch 7: train loss :: 1.192, train acc :: 0.518, dev acc :: 0.481
+Epoch 8: train loss :: 1.191, train acc :: 0.513, dev acc :: 0.467
+Epoch 9: train loss :: 1.191, train acc :: 0.505, dev acc :: 0.448
+Evaluating on SST...
+load model from sst-classifier.pt
+load 1101 data from data/ids-sst-dev.csv
+DONE DEV
+DONE Test
+dev acc :: 0.481
+Training Sentiment Classifier on cfimdb...
+load 1707 data from data/ids-cfimdb-train.csv
+load 245 data from data/ids-cfimdb-dev.csv
+Epoch 0: train loss :: 0.574, train acc :: 0.821, dev acc :: 0.829
+Epoch 1: train loss :: 0.466, train acc :: 0.866, dev acc :: 0.857
+Epoch 2: train loss :: 0.419, train acc :: 0.872, dev acc :: 0.873
+Epoch 3: train loss :: 0.386, train acc :: 0.878, dev acc :: 0.833
+Epoch 4: train loss :: 0.373, train acc :: 0.899, dev acc :: 0.849
+Epoch 5: train loss :: 0.357, train acc :: 0.893, dev acc :: 0.865
+Epoch 6: train loss :: 0.342, train acc :: 0.905, dev acc :: 0.873
+Epoch 7: train loss :: 0.334, train acc :: 0.906, dev acc :: 0.873
+Epoch 8: train loss :: 0.345, train acc :: 0.892, dev acc :: 0.824
+Epoch 9: train loss :: 0.321, train acc :: 0.888, dev acc :: 0.820
+Evaluating on cfimdb...
+load model from cfimdb-classifier.pt
+load 245 data from data/ids-cfimdb-dev.csv
+DONE DEV
+DONE Test
+dev acc :: 0.873

trainings/last-layer-w-dropout.txt

@@ -0,0 +1,38 @@
+Training Sentiment Classifier on SST...              
+load 8544 data from data/ids-sst-train.csv
+load 1101 data from data/ids-sst-dev.csv
+Epoch 0: train loss :: 1.458, train acc :: 0.460, dev acc :: 0.442
+Epoch 1: train loss :: 1.331, train acc :: 0.472, dev acc :: 0.440
+Epoch 2: train loss :: 1.288, train acc :: 0.476, dev acc :: 0.447
+Epoch 3: train loss :: 1.269, train acc :: 0.490, dev acc :: 0.457
+Epoch 4: train loss :: 1.252, train acc :: 0.485, dev acc :: 0.446
+Epoch 5: train loss :: 1.242, train acc :: 0.487, dev acc :: 0.447
+Epoch 6: train loss :: 1.235, train acc :: 0.511, dev acc :: 0.472
+Epoch 7: train loss :: 1.235, train acc :: 0.512, dev acc :: 0.465
+Epoch 8: train loss :: 1.235, train acc :: 0.512, dev acc :: 0.472
+Epoch 9: train loss :: 1.227, train acc :: 0.509, dev acc :: 0.475
+Evaluating on SST...
+load model from sst-classifier.pt
+load 1101 data from data/ids-sst-dev.csv
+DONE DEV
+DONE Test
+dev acc :: 0.475
+Training Sentiment Classifier on cfimdb...
+load 1707 data from data/ids-cfimdb-train.csv
+load 245 data from data/ids-cfimdb-dev.csv
+Epoch 0: train loss :: 0.590, train acc :: 0.819, dev acc :: 0.849
+Epoch 1: train loss :: 0.510, train acc :: 0.826, dev acc :: 0.845
+Epoch 2: train loss :: 0.459, train acc :: 0.848, dev acc :: 0.853
+Epoch 3: train loss :: 0.438, train acc :: 0.880, dev acc :: 0.857
+Epoch 4: train loss :: 0.413, train acc :: 0.876, dev acc :: 0.869
+Epoch 5: train loss :: 0.406, train acc :: 0.890, dev acc :: 0.833
+Epoch 6: train loss :: 0.401, train acc :: 0.893, dev acc :: 0.845
+Epoch 7: train loss :: 0.403, train acc :: 0.870, dev acc :: 0.861
+Epoch 8: train loss :: 0.393, train acc :: 0.879, dev acc :: 0.865
+Epoch 9: train loss :: 0.407, train acc :: 0.895, dev acc :: 0.873
+Evaluating on cfimdb...
+load model from cfimdb-classifier.pt
+load 245 data from data/ids-cfimdb-dev.csv
+DONE DEV
+DONE Test
+dev acc :: 0.873