Training loss is not changing at all while training LSTM

nlp
1

I am trying to solve a text classification problem. My training data has input as a sequence of 80 numbers in which each represent a word and target value is just a number between 1 and 3.
I pass it through this model:

class Model(nn.Module):
    def __init__(self, tokenize_vocab_count):
        super().__init__()
        self.embd = nn.Embedding(tokenize_vocab_count+1, 300)
        self.embd_dropout = nn.Dropout(0.3)
        self.LSTM = nn.LSTM(input_size=300, hidden_size=100, dropout=0.3, batch_first=True)
        self.lin1 = nn.Linear(100, 1024)
        self.lin2 = nn.Linear(1024, 512)
        self.lin_dropout = nn.Dropout(0.8)
        self.lin3 = nn.Linear(512, 3)
    
    def forward(self, inp):
        inp = self.embd_dropout(self.embd(inp))
        inp, (h_t, h_o) = self.LSTM(inp)
        h_t = F.relu(self.lin_dropout(self.lin1(h_t)))
        h_t = F.relu(self.lin_dropout(self.lin2(h_t)))
        out = F.softmax(self.lin3(h_t))
        return out

My training loop is as follows:

model = Model(tokenizer_obj.count+1).to('cuda')

optimizer = optim.AdamW(model.parameters(), lr=1e-2)
loss_fn = nn.CrossEntropyLoss()

EPOCH = 10

for epoch in range(0, EPOCH):
     for feature, target in tqdm(author_dataloader):
         train_loss = loss_fn(model(feature.to('cuda')).view(-1,  3), target.to('cuda'))
         optimizer.zero_grad()
         train_loss.backward()
         optimizer.step()
      print(f"epoch: {epoch + 1}\tTrain Loss : {train_loss}")

I printed out the feature and target dimension and it is as follows:

torch.Size([64, 80]) torch.Size([64])

Here 64 is the batch_size.
I am not doing any validation as of now.
When I train I am getting a constant loss value and no change

/home/koushik/Software/miniconda3/envs/fastai/lib/python3.7/site-packages/torch/nn/modules/rnn.py:50: UserWarning: dropout option adds dropout after all but last recurrent layer, so non-zero dropout expects num_layers greater than 1, but got dropout=0.3 and num_layers=1
  "num_layers={}".format(dropout, num_layers))
  0%|                                                                                                                                                 | 0/306 [00:00<?, ?it/s]/media/koushik/Backup Plus/Code/Machine Deep Learning/NLP/src/Deep Learning/model.py:20: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.
  out = F.softmax(self.lin3(h_t))
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 306/306 [00:03<00:00, 89.36it/s]
epoch: 1        Train Loss : 1.0986120700836182
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 306/306 [00:03<00:00, 89.97it/s]
epoch: 2        Train Loss : 1.0986120700836182
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 306/306 [00:03<00:00, 89.35it/s]
epoch: 3        Train Loss : 1.0986120700836182
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 306/306 [00:03<00:00, 89.17it/s]
epoch: 4        Train Loss : 1.0986120700836182
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 306/306 [00:03<00:00, 88.72it/s]
epoch: 5        Train Loss : 1.0986120700836182
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 306/306 [00:03<00:00, 87.75it/s]
epoch: 6        Train Loss : 1.0986120700836182
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 306/306 [00:03<00:00, 85.67it/s]
epoch: 7        Train Loss : 1.0986120700836182
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 306/306 [00:03<00:00, 85.40it/s]
epoch: 8        Train Loss : 1.0986120700836182
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 306/306 [00:03<00:00, 84.49it/s]
epoch: 9        Train Loss : 1.0986120700836182
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 306/306 [00:03<00:00, 84.21it/s]
epoch: 10       Train Loss : 1.0986120700836182

Can anyone please help

2

nn.CrossEntropyLoss uses F.log_softmax and nn.NLLLoss internally, so it expects raw logits as the model outputs.
Remove the F.softmax at the end of your model and pass the output of self.lin(3) directly to the criterion.
Let me know, if that helps. :slight_smile:

1 Like
3

Yes, model training now. Thank you! :grinning:

4

I am having another problem now. I am training the model but the loss is going up and down repeatedly. It’s not decreasing or converging. I tried to decrease the learning rate but it didn’t work.
Here’s the loss plot at lr = 1e-3 for 30 epochs:-
Figure_1
Here’s the loss plot at lr = 1e-6 for 30 epochs:-
Figure_2
Here’s the loss plot at lr = 1e-9 for 30 epochs:-
Figure_3