non-vanishing RNN

config.yml

@@ -56,8 +56,8 @@ BidirectionalLSTM:
 
 RNN:
   HiddenSize: 35
-  NumLayers: 4
-  NumFFN: 15
+  NumLayers: 2
+  NumFFN: 2
   Dropout: 0.2
   Initialization: None
 

main.py

@@ -120,14 +120,14 @@ if __name__ == "__main__":
         wandb.watch(network, log_freq = log_freq)
 
     for t in range(cfg["Training"]["Epochs"]):
-        logging.info("Epoch {}".format(t))
+        print("Epoch {}".format(t))
         train(args, network, train_loader, f_loss, optimizer, device, log_interval)
 
         val_loss = test.test(network, valid_loader, f_loss, device)
 
         network_checkpoint.update(val_loss)
 
-        logging.info(" Validation : Loss : {:.4f}".format(val_loss))
+        print(" Validation : Loss : {:.4f}".format(val_loss))
         if not args.no_wandb:
             wandb.log({"val_loss": val_loss})
 

test.py

@@ -35,6 +35,10 @@ def test(model, loader, f_loss, device):
 
             # Compute the forward pass, i.e. the scores for each input image
             outputs = model(inputs)
+            print("Validation inputs :")
+            print(inputs)
+            print("Validation outputs :")
+            print(outputs)
 
             # We accumulate the exact number of processed samples
             N += inputs.shape[0]

wandb/debug-internal.log

-run-20230203_215737-zxchwf43/logs/debug-internal.log
+run-20230203_233455-ujner2cx/logs/debug-internal.log
\ No newline at end of file

wandb/debug.log

-run-20230203_215737-zxchwf43/logs/debug.log
+run-20230203_233455-ujner2cx/logs/debug.log
\ No newline at end of file

wandb/latest-run

-run-20230203_215737-zxchwf43
+run-20230203_233455-ujner2cx
\ No newline at end of file