Merge branch 'test'

setup.py

@@ -1,7 +1,8 @@
 import os
 import setuptools
 
-current_dir = os.path.abspath(os.path.dirname(__file__))
+current_dir = os.path.dirname(os.path.abspath("__file__"))
+
 
 ## requirements
 with open(os.path.join(current_dir, "README.md"), "r") as fh:
@@ -31,4 +32,5 @@
     ],
     python_requires=">=3.6",
     install_requires=required,
+    include_package_data=True,
 )

twittersentiment/TwitterSentiment.py

@@ -12,8 +12,7 @@
 from sklearn.model_selection import train_test_split
 
 
-from twittersentiment.utils.model import TwitterModel as Model
-from twittersentiment.utils.model import TwitterModel
+from twittersentiment.utils.model import TweetModel
 from twittersentiment.utils.preprocessing import Preprocess
 from twittersentiment.utils.data import TweetDataset
 
@@ -35,19 +34,21 @@ def __init__(self):
 
         self.models = {
             "twitter-en": model(
-                url="https://github.com/shahules786/ML_EXP/releases/download/sample/best_model.pt",
-                model=Model,
+                url="/home/shahul/Downloads/classifier(1).pt",
+                model=TweetModel,
             )
         }
 
-        self.tokenizer_path = os.path.split(os.__file__)[0]
+        self.tokenizer_path = os.path.join(
+            os.path.dirname(os.path.abspath("__file__")), "twittersentiment", "utils", "tokenizer.pickle"
+        )
 
     def load_pretrained(self, model_name="twitter-en"):
 
-        state_dict = torch.hub.load_state_dict_from_url(
-            self.models[model_name].url, progress=True, map_location=DEVICE
-        )
-        # state_dict = torch.load("/home/shahul/Downloads/classifier.pt", map_location=DEVICE)
+        # state_dict = torch.hub.load_state_dict_from_url(
+        #    self.models[model_name].url, progress=True, map_location=DEVICE
+        # )
+        state_dict = torch.load("/home/shahul/Downloads/classifier (1).pt", map_location=DEVICE)
         self.model = self.models[model_name].model(state_dict["embedding.weight"].numpy())
         self.model.load_state_dict(state_dict)
 
@@ -89,7 +90,7 @@ def train(self, X, y, path, epochs=5, lr=1e-3, batch_size=32, test_size=0.15, na
             "valid": DataLoader(test_data, batch_size=batch_size, shuffle=False),
         }
 
-        model = TwitterModel(embedding_matrix).to(DEVICE)
+        model = TweetModel(embedding_matrix).to(DEVICE)
         loss_fn = torch.nn.BCELoss().cuda()
         optimizer = torch.optim.Adam(model.parameters())
 

twittersentiment/utils/model.py

@@ -3,46 +3,31 @@
 import torch.nn.functional as F
 
 
-class TwitterModel(nn.Module):
-    """
-    Model Architecture
-    """
+class TweetModel(nn.Module):
+    def __init__(self, embedding_matrix, lstm_hidden_size=200, gru_hidden_size=128):
 
-    def __init__(self, embedding_matrix, lstm_hidden_size=128, gru_hidden_size=64):
-
-        super(TwitterModel, self).__init__()
+        super(TweetModel, self).__init__()
         self.embedding = nn.Embedding(*embedding_matrix.shape)
         self.embedding.weight = nn.Parameter(torch.tensor(embedding_matrix, dtype=torch.float32))
-        self.embedding.weight.requires_grad = False
-        self.embedding_dropout = nn.Dropout2d(0.2)
-
-        self.lstm = nn.LSTM(embedding_matrix.shape[1], lstm_hidden_size, bidirectional=True, batch_first=True)
-        self.lstm2 = nn.LSTM(lstm_hidden_size * 2, gru_hidden_size, bidirectional=True, batch_first=True)
-
-        self.Linear1 = nn.Linear(gru_hidden_size * 4, 1)
-
-    def apply_spatial_dropout(self, h_embedding):
+        self.embedding.weight.requires_grad = True
+        self.embedding_dropout = nn.Dropout2d(0.1)
 
-        h_embedding = h_embedding.transpose(1, 2).unsqueeze(2)
-        h_embedding = self.embedding_dropout(h_embedding).squeeze(2).transpose(1, 2)
-        return h_embedding
+        self.gru = nn.GRU(
+            embedding_matrix.shape[1], gru_hidden_size, num_layers=1, bidirectional=True, batch_first=True
+        )
 
-    def flatten_parameters(self):
-
-        self.lstm.flatten_parameters()
-        self.lstm2.flatten_parameters()
+        self.dropout2 = nn.Dropout(0.25)
+        self.Linear1 = nn.Linear(gru_hidden_size * 5, 16)
+        self.Linear2 = nn.Linear(16, 1)
 
     def forward(self, x):
-
         h_embedding = self.embedding(x)
-        h_embedding = self.apply_spatial_dropout(h_embedding)
-
-        h_lstm, _ = self.lstm(h_embedding)
-        h_lstm, _ = self.lstm2(h_lstm)
 
-        avg_pool = torch.mean(h_lstm, 1)
-        max_pool, _ = torch.max(h_lstm, 1)
-        concat = torch.cat((avg_pool, max_pool), 1)
+        x, (x_h, x_c) = self.gru(h_embedding)
 
-        out = torch.sigmoid(self.Linear1(concat))
+        avg_pool = torch.mean(x, 1)
+        max_pool, _ = torch.max(x, 1)
+        concat = torch.cat((avg_pool, x_h, max_pool), 1)
+        concat = self.Linear1(concat)
+        out = torch.sigmoid(self.Linear2(concat))
         return out

twittersentiment/utils/preprocessing.py

@@ -34,7 +34,8 @@ def tokenizer(corpus, path, mode="train"):
 
         """
 
-        path = os.path.join(path, "new_tokenizer.pickle")
+        if not path.endswith(".pickle"):
+            path = os.path.join(path, "tokenizer.pickle")
 
         if mode == "train":
             tokenizer_obj = Tokenizer()