[resolved] GRU for sentiment classification

(Florian) #1

I think that I’m missing something obvious so I’m sorry about taking up forum space. I’m trying to get familiar with PyTorch by recreating the Keras examples(https://github.com/fchollet/keras/tree/master/examples) because they are all really simple. The first one that I tried to tackle was the bidirectional RNN for IMDB sentiment analysis. My network just doesn’t seem to be training whereas the same structure in Keras achieves 80% accuracy in two epochs.

My output:

Epoch 1/5000 
Training Loss: 0.669 | Training Acc: 23.2
Time: 2.80514907837

Epoch 2/5000
Training Loss: 0.661 | Training Acc: 23.7
Time: 2.80501890182

Epoch 3/5000
Training Loss: 0.656 | Training Acc: 24.0
Time: 2.80491304398


import time
import numpy as np

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.autograd as autograd
import torch.optim as optim

ITERS = 5000
max_features = 20000
batch_size = 40

print 'Loading data ...'

# training data are index for words for the embedding layer,
x_train = np.load('data/X_train_imdb_idx.npy')
y_train = np.load('data/y_train_imdb_idx.npy')

x_test = np.load('data/X_test_imdb_idx.npy')
y_test = np.load('data/y_test_imdb_idx.npy')

class BIGRU(nn.Module):
    def __init__(self):
        super(BIGRU, self).__init__()

        self.embedding = nn.Embedding(max_features, 128)
        self.gru = nn.GRU(128, 64, num_layers=1, bidirectional=True)
        self.dropout = nn.Dropout(p=0.5)
        self.linear = nn.Linear(128, 1) # input dim is 64*2 because its bidirectional

    def forward(self, x, h):
        x = self.embedding(x)
        x, h = self.gru(x, h)
        x = self.dropout(x[:,-1,:].squeeze()) # just get the last hidden state
        x = F.sigmoid(self.linear(x)) # sigmoid output for binary classification
        return x, h

    def init_hidden(self):
        return autograd.Variable(torch.randn(2, batch_size, 64)).cuda()

model = BIGRU()

print 'Putting model on GPU ... '

loss_fn = nn.BCELoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)

num_batch_epoch = len(x_train) // batch_size

h = model.init_hidden()

print 'Training ...'
for e in range(ITERS):
    print '\n' + 'Epoch {}/{}'.format(e, ITERS)
    print '-' * 10
    start = time.time()

    idx = np.random.permutation(len(x_train))
    iter_loss = 0.
    iter_correct = 0.
    for b in range(len(x_train) // batch_size):
        x_b = torch.from_numpy(x_train[idx[b*batch_size:(b+1)*batch_size]]).long()
        y_b = torch.from_numpy(y_train[idx[b*batch_size:(b+1)*batch_size]]).float()

        x_batch = autograd.Variable(x_b.cuda())#
        y_batch = autograd.Variable(y_b.cuda())#

        y_pred, h = model(x_batch, h)


        loss = loss_fn(y_pred, y_batch)

        #clipped_lr = 0.001 * nn.utils.clip_grad_norm(model.parameters(), 0.25)
        # clip gradients because RNN
        #for p in model.parameters():
        #    p.data.add_(-clipped_lr, p.grad.data)


        trn_preds = torch.round(y_pred.data)
        iter_correct += torch.sum(trn_preds == y_batch.data)
        iter_loss += loss.data[0]

    print 'Training Loss: {:.3} | Training Acc: {:.3}'.format(iter_loss / num_batch_epoch, float(iter_correct) / num_batch_epoch)
    print 'Time: {}'.format(time.time()-start)

print '\n' + 'Evaluating network ...'
model.eval() # changes the behavior of the dropout for making predictions
# evaluate the accuracy
running_corrects = 0
for vb in range(len(x_test) // batch_size):
    x_v_b = torch.from_numpy(x_test[vb*batch_size:(vb+1)*batch_size]).long()
    y_v_b = torch.from_numpy(y_test[vb*batch_size:(vb+1)*batch_size]).float()

    x_v_batch = autograd.Variable(x_v_b.cuda(), volatile=True)
    y_v_batch = autograd.Variable(y_v_b.cuda())
    outputs, h = model(x_v_batch, h)
    preds = torch.round(outputs.data)
    running_corrects += torch.sum(preds == y_v_batch.data)

    if vb % 100 == 0:
        print 'Valid batch:', vb, 'done!'

print 'Percent correct:', float(running_corrects) / len(x_test)

(Florian) #2

I found the answer, I had to set batch_first=True in the GRU layer. My question was close enough to another that it can probably be deleted.

(Lisette García Moya) #3

I am trying to get familiar with PyTorch too.
One question, how do you construct your dataset or where it was donwloaded from? I refer to:

  • X_train_imdb_idx.npy
  • y_train_imdb_idx.npy
  • X_test_imdb_idx.npy
  • y_test_imdb_idx.npy

(nosrpc) #4

I am facing a similar problem. In fact, I tried the exact code given by Florian (with batch_first=True) and I hardly get about 25 % accuracy on the test set, with number of iteration ( epochs) set to 50.

However, the Keras code at the above described link, easily attains 80% + accuracy with just 5 epochs over the training data. The NN architecture between the two seems to be identical, except for the default values for the LSTM and GRU cells in the Keras and Pytorch implementations, such as LSTM’s kernel initialization, recurrenct_activation==‘hard_sigmoid’ … and so on.

Can someone please clarify if I am missing something ?

(nosrpc) #5

Ok… For anyone that might land up here, BCELoss seems to have an issue in PyTorch. Switching to CrossEntropy loss even for a binary classification task, solved my problem.

In summary, if you architecture is right, double check the choice of loss functions and the way the true labels have to be prepared, as expected by the loss function.

(Shengchao Liu) #6

Thanks for pointing this out. But as I tested, the loss values are different when switching from BCELoss to CrossEntropyLoss.