bidirectional_lstm_with_MEAN_STD

config.yml

@@ -11,6 +11,12 @@ Dataset:
   _ENCODING_INDEX: "I"  # h(short) with 2 bytes should be sufficient
   _ENCODING_OFFSET_FORMAT: ""
   _ENCODING_ENDIAN: "<"
+  Transform: dataloader.composite_transform(dataloader.transform_remove_space_time(), dataloader.transform_normalize_with_train_statistics(MEAN, STD))
+  # Available transforms:
+  # dataloader.transform_remove_space_time()
+  # dataloader.transform_normalize_with_train_statistics(MEAN, STD)
+  # dataloader.transform_min_max_scaling(MIN, MAX)
+  # dataloader.composite_transform(dataloader.transform_remove_space_time(), dataloader.transform_min_max_scaling(MIN, MAX))
 
 # Data Transformation
 ApproximativeStats: True
@@ -31,6 +37,7 @@ ApproximativeMini: "torch.tensor([ 4.1479e+01,  6.0000e+00,  1.0182e+00,  1.2623
          7.9218e+00,  1.0000e-11,  3.7171e+01,  2.5584e+00,  1.2075e+01,
         -1.2436e+00, -9.9256e-01, -8.8131e-01])"
 
+
 #Optimizer selection
 Optimizer: Adam # in {Adam}
 
@@ -40,7 +47,7 @@ Training:
   
 #Model selection
 Model:
-  Name: CNN1D
+  Name: BidirectionalLSTM
   #choose in {LinearRegression, BidirectionalLSTM, RNN}
 
 #Model parameters selection
@@ -51,9 +58,10 @@ LinearRegression:
   Initialization: init_he
 
 BidirectionalLSTM:
-  HiddenSize: 16
-  NumLayers: 1
-  Dropout: 0.2
+  HiddenSize: 32
+  NumLayers: 4
+  LSTMDropout: 0
+  FFNDropout: 0.2
   NumFFN: 8
   Initialization: None
 

main.py

@@ -59,6 +59,8 @@ if __name__ == "__main__":
     log_freq        = int(cfg["Wandb"]["log_freq"])
     log_interval    = int(cfg["Wandb"]["log_interval"])
 
+    dataset_transform = cfg["Dataset"]["Transform"]
+
     if not args.no_wandb:
         wandb.init(entity = "wherephytoplankton", project = "Kaggle phytoplancton", config = {"batch_size": batch_size, "epochs": epochs})
 
@@ -93,8 +95,8 @@ if __name__ == "__main__":
         valid_ratio,
         overwrite_index =  True,
         max_num_samples=max_num_samples,
-        train_transform=dataloader.composite_transform(dataloader.transform_remove_space_time(), dataloader.transform_min_max_scaling(MIN, MAX)),
-        valid_transform=dataloader.composite_transform(dataloader.transform_remove_space_time(), dataloader.transform_min_max_scaling(MIN, MAX))
+        train_transform= eval(dataset_transfom),
+        valid_transform=eval(dataset_transfom)
     )
 
     if use_cuda :
@@ -133,4 +135,4 @@ if __name__ == "__main__":
 
     utils.write_summary(logdir, network, optimizer, val_loss)
     
-    create_submission.create_submission(network, dataloader.composite_transform(dataloader.transform_remove_space_time(), dataloader.transform_min_max_scaling(MIN, MAX)), device, rootDir, logdir)
+    create_submission.create_submission(network, eval(dataset_transfom), device, rootDir, logdir)

model.py

@@ -76,10 +76,11 @@ class BidirectionalLSTM(nn.Module):
         super(BidirectionalLSTM, self).__init__()
         self.hidden_size = cfg["BidirectionalLSTM"]["HiddenSize"]
         self.num_layers =  cfg["BidirectionalLSTM"]["NumLayers"]
-        self.dropout    =  cfg["BidirectionalLSTM"]["Dropout"]
+        self.LSTM_dropout    =  cfg["BidirectionalLSTM"]["LSTMDropout"]
+        self.FFN_dropout    = cfg["DidirectionalLSTM"]["FFNDropout"]
         self.num_ffn    = cfg["BidirectionalLSTM"]["NumFFN"]
 
-        self.lstm = nn.LSTM(input_size, self.hidden_size, self.num_layers, batch_first = True, bidirectional =True, dropout = 0.2)
+        self.lstm = nn.LSTM(input_size, self.hidden_size, self.num_layers, batch_first = True, bidirectional =True, dropout = self.LSTM_dropout)
         self.fc = nn.Sequential()
         
         for layer in range(self.num_ffn):
@@ -92,7 +93,7 @@ class BidirectionalLSTM(nn.Module):
             )
             self.fc.add_module(
                 f"dropout_{layer}", 
-                nn.Dropout(p=self.dropout)
+                nn.Dropout(p=self.FFN_dropout)
             )
         self.fc.add_module(
             "last_linear",
@@ -120,16 +121,15 @@ class CNN1D(torch.nn.Module):
     def __init__(self, cfg, num_inputs):
         super(CNN1D, self).__init__()
         self.model = torch.nn.Sequential(
-            *conv_block(num_inputs, 6, 0.01),
-            *conv_block(64, 7, 0.01),
-            *conv_block(128,8,0.01)
+            *conv_block(num_inputs, 32),
+            *conv_block(32, 128)
         )
         self.avg_pool = torch.nn.AdaptiveAvgPool1d(1)
 
-        self.fc = nn.Sequential(
-            nn.Linear(256, 32),
+        self.ffn = nn.Sequential(
+            nn.Linear(128, 128),
             nn.ReLU(),
-            nn.Linear(32,cfg["Dataset"]["num_days"])
+            nn.Linear(128,cfg["Dataset"]["num_days"])
         )
 
     def forward(self, x):
@@ -137,25 +137,21 @@ class CNN1D(torch.nn.Module):
 
         out = self.model(x)
 
+        print(f"This is after CNN : {out}")
         out = self.avg_pool(out)
 
         out = out.view([out.shape[0], -1])
 
-        out = self.fc(out)
+        #print(f"{out.shape} this is out.shape")
+        out = self.ffn(out)
         out = out.view([out.shape[0], out.shape[1], 1])
         return out
 
-def conv_block(in_channels, power, dropout_p):
+def conv_block(in_channels, out_channels):
     return [
-        torch.nn.Conv1d(in_channels, 2**power, 16),
-        #torch.nn.BatchNorm1d(2**power),
-        torch.nn.LeakyReLU(),
-        torch.nn.Dropout(p=dropout_p),
-        torch.nn.Conv1d(2**power, 2**power, 8),
-        torch.nn.BatchNorm1d(2**power),
-        torch.nn.LeakyReLU(),
-        #torch.nn.Dropout(p=dropout_p),
-        torch.nn.MaxPool1d(2, stride = 1)
+        torch.nn.Conv1d(in_channels, out_channels, kernel_size = 3, stride = 1, padding = 1),
+        torch.nn.ReLU(),
+        torch.nn.BatchNorm1d(out_channels),
     ]
 
 # Initialization

wandb/debug-internal.log

-run-20230204_013952-5w9xw0aw/logs/debug-internal.log
\ No newline at end of file
+run-20230204_131220-udsbqzli/logs/debug-internal.log
\ No newline at end of file

wandb/debug.log

-run-20230204_013952-5w9xw0aw/logs/debug.log
\ No newline at end of file
+run-20230204_131220-udsbqzli/logs/debug.log
\ No newline at end of file

wandb/latest-run

-run-20230204_013952-5w9xw0aw
\ No newline at end of file
+run-20230204_131220-udsbqzli
\ No newline at end of file