Requesting help with padding/packing lstm for simple classification task

I’m running into a knowledge block.

I want to make a simple binary classifiyer. I took a deep dive into padded/packed sequences and think I understand them pretty well. After perusing around and thinking about it, I came to the conclusion that I should be grabbing the final non-padded hidden state of each sequence, so that’s what I tried below:


class TextClassificationModel(nn.Module):

    def __init__(self, vocab_size, embed_dim, hidden_size, num_class):
        super(TextClassificationModel, self).__init__()
        self.embedding = nn.Embedding(vocab_size, embed_dim, padding_idx=0)
        self.lstm = nn.LSTM(embed_dim, hidden_size, batch_first=True)
        self.fc1 = nn.Linear(hidden_size, num_class)

    def forward(self, padded_seq, lengths):
        # embedding layer
        embedded_padded = self.embedding(padded_seq)
        packed_output = pack_padded_sequence(embedded_padded, lengths, batch_first=True)

        # lstm layer
        output, _ = self.lstm(packed_output)
        padded_output, lengths = pad_packed_sequence(output, batch_first=True)

        # get hidden state of final non-padded sequence element:
        h_n = []
        for seq, length in zip(padded_output, lengths):
            h_n.append(seq[length - 1, :])
        lstm_out = torch.stack(h_n)
        # linear layers
        out = self.fc1(lstm_out)
        return out

This morning, I ported my notebook over to an IDE and ran the debugger and confirmed that h_n is indeed the final hidden state of each sequence, not including padding.

So everything runs/trains without error but my loss never decreases when I use batch size > 1.

With batch_size = 8:

With `batch_size = 1:

The reason I’m thinking I’ve done something wrong is that this performance is much worse than a simple embedding layer followed by a single linear layer:

self.embedding = nn.EmbeddingBag(vocab_size, embed_dim, sparse=True)
self.fc = nn.Linear(embed_dim, num_class)


I thought I’d add the dataset, collate function and training code. This isn’t getting many hits, and I don’t know if that’s because my question is uninteresting or because I did’t provide enough data. So her’es more code:


class TextAndLabelsFromPandas(Dataset):
    def __init__(self, dataframe):
        self.dataframe = dataframe
        self.n = 0
    def __getitem__(self, idx):
        text = pd.Series(self.dataframe.iloc[idx, :-1], dtype="string")
        label = self.dataframe.iloc[idx, -1]
        return label, " ".join(text.tolist())
    def __len__(self):
        return len(self.dataframe)
    def __iter__(self):
        return self;
    def __next__(self):
        if self.n < len(self):
            result = self[self.n]
            self.n += 1
            return result
            raise StopIteration

Collate Function:

text_pipeline = lambda x: vocab(tokenizer(x))

def collate_batch(batch):
    label_list, text_list = [], []
    # Tokenize batch
    for (label, text) in batch:
        processed_text = torch.tensor(text_pipeline(text), dtype=torch.int64)
    # get lengths of individual sequences and sort by length 
    text_list.sort(key=lambda x: len(x), reverse=True)
    lengths = [len(sequence) for sequence in text_list]

    # convert to tensor and pad
    label_tensor = torch.tensor(label_list, dtype=torch.float64)
    padded_input_tensor = pad_sequence(text_list, batch_first=True, padding_value=vocab["<pad>"])
    return,, lengths

Training Protocol:

def train_one_epoch(model, opt, criterion, lr, trainloader):
    running_tl = 0
    for (label, sequence, lengths) in trainloader:
        label = label.reshape(label.size()[0], 1)
        output = model(sequence, lengths)
        loss = criterion(output, label)

        running_tl += loss.item()
    return running_tl
def validate_one_epoch(model, opt, criterion, lr, validloader):
    running_vl = 0
    with torch.no_grad():
        for (label, sequence, lengths) in validloader:
            label = label.reshape(label.shape[0], 1)
            output = model(sequence, lengths)
            loss = criterion(output, label)
            running_vl += loss.item()
    return running_vl

def train_model(model, opt, criterion, epochs, trainload, testload=None, lr=1e-3):
    avg_tl_per_epoch = []
    avg_vl_per_epoch = []
    for e in trange(epochs):
        running_tl = train_one_epoch(model, opt, criterion, lr, trainload)
        avg_tl_per_epoch.append(running_tl / len(trainload))
        if testload:
            running_vl = validate_one_epoch(model, opt, criterion, lr, validloader)
            avg_vl_per_epoch.append(running_vl / len(testload))
    return avg_tl_per_epoch, avg_vl_per_epoch

I wanted to add one more thing I’ve tried. I replaced the forward part of the model with the following:

    def forward(self, padded_seq, lengths):
        # embedding layer
        x = self.embedding(padded_seq)
        x = pack_padded_sequence(x, lengths, batch_first=True)

        # lstm layer
        x, hidden = self.lstm(x)
        x, lengths = pad_packed_sequence(x, batch_first=True)
        # linear layer1
        out = self.dropout_1(x)
        out = self.fc1(out)   
        out = self.sigmoid(out)
        # mask outputs to those connected with the final hidden state
        return self.mask(out, lengths)
    def mask(self, tensor, lengths):
        idx = torch.tile(torch.arange(max(lengths)), (tensor.shape[0], 1))
        mask = idx == torch.unsqueeze(lengths - 1, axis = 1)
        return tensor[mask]

This resulted in the now all too familiar loss curves:

One thing it did help with is that it let me increase the batch size without getting those crazy loss curves. So that means I’m getting closer.

So in case someone else ever has this same question, this is what a month’s worth of research has led me to believe:

LSTM’s are prone to overfitting unless you have a lot of data. This Kaggle competition data includes ~7600 tweets, and I just don’t think that’s enough.

What leads me to believe this is that I could perfectly overfit 20 samples, 10 with positive and 10 with negative labels. But when training on the whole dataset, I had two options: Overfit or Underfit.

  1. Simplify the model to the extreme. With the following parameters, the model did not overfit but severely underperformed (barely better than a flip of the coin).
    • Embedding size: 10
    • Lstm hidden dimension: 8
    • Added learning rate decay with a factor of 0.1 and patience of 5 epochs.

  1. A slight improvement was to train a nn.EmbeddingBag layer as in my original post, then transfer the trained weights over to the LSTM’s nn.Embedding layer.

  2. Adding any more complexity to the model resulted in overfitting after ~5-30 epochs depending on the amount added.

What I will try next is to make the model bidirectional, but I have no idea how to mask the loss of a bidirectional LSTM. I guess that’s a different question.