Simple char level RNN, can't overfit simple training example


I’m new to the PyTorch framework (coming from Theano and Tensorflow mainly) which I really enjoy to use.
I’ve followed the introduction tutorial and read the classifying Classifying Names with a Character-Level RNN one.
I now try to adapt it to a char level LSTM model in order to gain some practical experience with the framework.
Basically I feed in the model sequences of char indices and give as target to the model the same sequence but shifted by one in the future.
However I can’t overfit a simple training example and I don’t see what I did wrong.
If someone can spot my mistake it would be very helpful.
Here is my code:

class LSTMTxtGen(nn.Module):
def __init__(self, hidden_dim, n_layer, vocab_size):
    super(LSTMTxtGen, self).__init__()
    self.n_layer = n_layer
    self.hidden_dim = hidden_dim
    self.vocab_size = vocab_size
    self.lstm = nn.LSTM(vocab_size, hidden_dim, n_layer, batch_first=True)

    # The linear layer that maps from hidden state space to tag space
    #self.hidden = self.init_hidden()

def init_hidden(self, batch_size):
    # Before we've done anything, we dont have any hidden state.
    # Refer to the Pytorch documentation to see exactly
    # why they have this dimensionality.
    # The axes semantics are (num_layers, minibatch_size, hidden_dim)
    return (autograd.Variable(torch.zeros(self.n_layer, batch_size, 
            autograd.Variable(torch.zeros(self.n_layer, batch_size, 

def forward(self, seqs):
    self.hidden = self.init_hidden(seqs.size()[0])
    lstm_out, self.hidden = self.lstm(seqs, self.hidden)
    lstm_out = lstm_out.contiguous().view(-1, self.hidden_dim)
    lstm_out = nn.Linear(lstm_out.size(1), self.vocab_size)(lstm_out)
    return lstm_out

model = LSTMTxtGen (
hidden_dim = 50,
n_layer = 3,
vocab_size = 44,


criterion = nn.CrossEntropyLoss()
optimizer = optim.Adamax(model.parameters())

G = Data.batch_generator(5,100)
batch_per_epoch, to_idx, to_char = next(G)
X, Y = next(G)
for epoch in range(10):
    losses = []
    for batch_count in range(batch_per_epoch):
        #mode.hidden = model.init_hidden()
        #X, Y = next(G)
        X = autograd.Variable(torch.from_numpy(X))
        Y = autograd.Variable(torch.from_numpy(Y))
       preds = model(X)
       loss = criterion(preds.view(-1, model.vocab_size), Y.view(-1))
       if (batch_count % 20 == 0):
           print('Loss: ', losses[-1])    

The loss keep oscillating and no improvement is made.

where is the learning rate?

You are creating a new output layer each time you call the forward.

    lstm_out = nn.Linear(lstm_out.size(1), self.vocab_size)(lstm_out)

and as a result, nothing is learned across iterations.
It should be something like:

    lstm_out = self.output_layer(lstm_out)

where self.output_layer = nn.Linear(...) should be defined in __init__()


Thanks guys !
It solved my problem