'generator' object is not subscriptable

nlp
1

I’m working on a CNN LSTM model. While attempting to pass input through the LSTM, I get the following error:
TypeError: 'generator' object is not subscriptable

The entire problem stems from keeping the hidden state initialization outside the dataloader loop. As soon as I move it in, it works, which would mean that my hidden state would be reset every batch. This is clearly undesirable.

for loader in loaders:
            hidden = model.hidden()         # Fails
            for X, y in loader:
                hidden = model.hidden()     # Works
                # Reshape data
                y_em, y_seg = y[0][1:], y[0][:1]
                x_l, x_c, x_r = X[0].permute(1, 0, 2, 3), X[1], X[2].permute(1, 0, 2, 3)

Here’s the entire fit function for reference:

def fit(model, loaders, run, rm, criterion, optimizer, num_epochs, device='cpu'):
    for epoch in range(num_epochs):
        rm.begin_epoch()
        for loader in loaders:
            hidden = model.hidden()
            for X, y in loader:
                # Reshape data
                y_em, y_seg = y[0][1:], y[0][:1]
                x_l, x_c, x_r = X[0].permute(1, 0, 2, 3), X[1], X[2].permute(1, 0, 2, 3)

                # Run model
                pred_em, pred_seg, hidden = model(x_l, x_c, x_r, hidden)

                # Convert segment prediction to single output
                pred_seg = torch.tensor(
                    [torch.argmax(pred_seg, dim=0)]
                ).float().to(device)

                # Calculate loss
                loss_em = criterion(y_em, pred_em).view([1])
                loss_seg = criterion(y_seg, pred_seg).view([1])

                total_loss = sum(loss_em, loss_seg)

                # Perform backprop
                total_loss.backward(retain_graph=True)
                optimizer.step()

                hidden = (x.detach() for x in hidden)

                # Track metrics
                rm.track_metrics(
                    loss_seg,
                    loss_em,
                    pred_seg,
                    y_seg,
                    pred_em,
                    y_em
                )
        rm.end_epoch()

And this is the code of the model:

class CNN_LSTM(nn.Module):
    def __init__(self, context_size):
        super(CNN_LSTM, self).__init__()
        
        # Context encoder
        self.context_encoder_lr = ContextEncoder()
        self.context_encoder_c  = ContextEncoder()
        
        # Hidden state generation
        self.lstm = nn.LSTM(72, 36)
        
        # Attention generation
        self.attention = nn.Linear(36, 36)
        
        # Output layers
        self.emotions = nn.Linear((2 * context_size + 1) * 36, 13)
        self.segment = nn.Linear((2 * context_size + 1) * 36, 2)
    
    def hidden(self):
        """Reset hidden state with (h_n, c_n)"""
        return (torch.zeros(1,1,36).to("cuda"), torch.zeros(1,1,36).to("cuda"))
        
    def forward(self, x_left, x_center, x_right, hidden):
        # Encode context
        c_l = self.context_encoder_lr(x_left)
        c_c = self.context_encoder_c(x_center)
        c_r = self.context_encoder_lr(x_right)
        
        # Stack contexts and unsqueeze batch size
        c = torch.cat([c_l, c_c, c_r])
        c = c.unsqueeze(1)
        
        # Pass context through LSTM
        x, hidden = self.lstm(c, hidden)
        
        # Generate attention
        attn = self.attention(x)
        attn = F.softmax(attn, dim=1)
        
        # Apply attention
        x = x * attn
        
        # Flatten layers
        x = x.view(-1)
        
        # Feed to output layers
        out_emotions = self.emotions(x)
        out_segment = self.segment(x)
        
        return F.sigmoid(out_emotions), F.softmax(out_segment), hidden```