This tutorial is based of Yoon Kim's paper on using convolutional neural networks for sentence sentiment classification.
For this tutorial, we will train a convolutional deep network model on movie review sentences from Rotten Tomatoes labeled with their sentiment. The result will be a model that can classify a sentence based on its sentiment (with 1 being a purely positive sentiment, 0 being a purely negative sentiment and 0.5 being neutral).
Our first step will be to fetch the labeled training data of positive and negative sentiment sentences and process it into sets of vectors that are then randomly split into train and test sets.
import urllib2 import numpy as np import re import itertools from collections import Counter def clean_str(string): """ Tokenization/string cleaning for all datasets except for SST. Original taken from https://github.com/yoonkim/CNN_sentence/blob/master/process_data.py """ string = re.sub(r"[^A-Za-z0-9(),!?\'\`]", " ", string) string = re.sub(r"\'s", " \'s", string) string = re.sub(r"\'ve", " \'ve", string) string = re.sub(r"n\'t", " n\'t", string) string = re.sub(r"\'re", " \'re", string) string = re.sub(r"\'d", " \'d", string) string = re.sub(r"\'ll", " \'ll", string) string = re.sub(r",", " , ", string) string = re.sub(r"!", " ! ", string) string = re.sub(r"\(", " \( ", string) string = re.sub(r"\)", " \) ", string) string = re.sub(r"\?", " \? ", string) string = re.sub(r"\s{2,}", " ", string) return string.strip().lower() def load_data_and_labels(): """ Loads MR polarity data from files, splits the data into words and generates labels. Returns split sentences and labels. """ # Pull sentences with positive sentiment pos_file = urllib2.urlopen('https://raw.githubusercontent.com/yoonkim/CNN_sentence/master/rt-polarity.pos') # Pull sentences with negative sentiment neg_file = urllib2.urlopen('https://raw.githubusercontent.com/yoonkim/CNN_sentence/master/rt-polarity.neg') # Load data from files positive_examples = list(pos_file.readlines()) positive_examples = [s.strip() for s in positive_examples] negative_examples = list(neg_file.readlines()) negative_examples = [s.strip() for s in negative_examples] # Split by words x_text = positive_examples + negative_examples x_text = [clean_str(sent) for sent in x_text] x_text = [s.split(" ") for s in x_text] # Generate labels positive_labels = [1 for _ in positive_examples] negative_labels = [0 for _ in negative_examples] y = np.concatenate([positive_labels, negative_labels], 0) return [x_text, y] def pad_sentences(sentences, padding_word="</s>"): """ Pads all sentences to the same length. The length is defined by the longest sentence. Returns padded sentences. """ sequence_length = max(len(x) for x in sentences) padded_sentences = [] for i in range(len(sentences)): sentence = sentences[i] num_padding = sequence_length - len(sentence) new_sentence = sentence + [padding_word] * num_padding padded_sentences.append(new_sentence) return padded_sentences def build_vocab(sentences): """ Builds a vocabulary mapping from word to index based on the sentences. Returns vocabulary mapping and inverse vocabulary mapping. """ # Build vocabulary word_counts = Counter(itertools.chain(*sentences)) # Mapping from index to word vocabulary_inv = [x[0] for x in word_counts.most_common()] # Mapping from word to index vocabulary = {x: i for i, x in enumerate(vocabulary_inv)} return [vocabulary, vocabulary_inv] def build_input_data(sentences, labels, vocabulary): """ Maps sentences and labels to vectors based on a vocabulary. """ x = np.array([[vocabulary[word] for word in sentence] for sentence in sentences]) y = np.array(labels) return [x, y] """ Loads and preprocessed data for the MR dataset. Returns input vectors, labels, vocabulary, and inverse vocabulary. """ # Load and preprocess data sentences, labels = load_data_and_labels() sentences_padded = pad_sentences(sentences) vocabulary, vocabulary_inv = build_vocab(sentences_padded) x, y = build_input_data(sentences_padded, labels, vocabulary) vocab_size = len(vocabulary) # randomly shuffle data np.random.seed(10) shuffle_indices = np.random.permutation(np.arange(len(y))) x_shuffled = x[shuffle_indices] y_shuffled = y[shuffle_indices] # split train/dev set # there are a total of 10662 labeled examples to train on x_train, x_dev = x_shuffled[:-1000], x_shuffled[-1000:] y_train, y_dev = y_shuffled[:-1000], y_shuffled[-1000:] sentence_size = x_train.shape[1] print 'Train/Dev split: %d/%d' % (len(y_train), len(y_dev)) print 'train shape:', x_train.shape print 'dev shape:', x_dev.shape print 'vocab_size', vocab_size print 'sentence max words', sentence_size
Train/Dev split: 9662/1000 train shape: (9662, 56) dev shape: (1000, 56) vocab_size 18766 sentence max words 56
Now that we prepared the training and test data by loading, vectorizing and shuffling it we can go on to defining the network architecture we want to train with the data.
We will first set up some placeholders for the input and output of the network then define the first layer, an embedding layer, which learns to map word vectors into a lower dimensional vector space where distances between words correspond to how related they are (with respect to sentiment they convey).
import mxnet as mx import sys,os ''' Define batch size and the place holders for network inputs and outputs ''' batch_size = 50 # the size of batches to train network with print 'batch size', batch_size input_x = mx.sym.Variable('data') # placeholder for input data input_y = mx.sym.Variable('softmax_label') # placeholder for output label ''' Define the first network layer (embedding) ''' # create embedding layer to learn representation of words in a lower dimensional subspace (much like word2vec) num_embed = 300 # dimensions to embed words into print 'embedding dimensions', num_embed embed_layer = mx.sym.Embedding(data=input_x, input_dim=vocab_size, output_dim=num_embed, name='vocab_embed') # reshape embedded data for next layer conv_input = mx.sym.Reshape(data=embed_layer, target_shape=(batch_size, 1, sentence_size, num_embed))
batch size 50 embedding dimensions 300
The next layer in the network performs convolutions over the ordered embedded word vectors in a sentence using multiple filter sizes, sliding over 3, 4 or 5 words at a time. This is the equivalent of looking at all 3-grams, 4-grams and 5-grams in a sentence and will allow us to understand how words contribute to sentiment in the context of those around them.
After each convolution, we add a max-pool layer to extract the most significant elements in each convolution and turn them into a feature vector.
Because each convolution+pool filter produces tensors of different shapes we need to create a layer for each of them, and then concatenate the results of these layers into one big feature vector.
# create convolution + (max) pooling layer for each filter operation filter_list=[3, 4, 5] # the size of filters to use print 'convolution filters', filter_list num_filter=100 pooled_outputs = [] for i, filter_size in enumerate(filter_list): convi = mx.sym.Convolution(data=conv_input, kernel=(filter_size, num_embed), num_filter=num_filter) relui = mx.sym.Activation(data=convi, act_type='relu') pooli = mx.sym.Pooling(data=relui, pool_type='max', kernel=(sentence_size - filter_size + 1, 1), stride=(1,1)) pooled_outputs.append(pooli) # combine all pooled outputs total_filters = num_filter * len(filter_list) concat = mx.sym.Concat(*pooled_outputs, dim=1) # reshape for next layer h_pool = mx.sym.Reshape(data=concat, target_shape=(batch_size, total_filters))
convolution filters [3, 4, 5]
Next, we add dropout regularization, which will randomly disable a fraction of neurons in the layer (set to 50% here) to ensure that that model does not overfit. This works by preventing neurons from co-adapting and forcing them to learn individually useful features.
This is necessary for our model because the dataset has a vocabulary of size around 20k and only around 10k examples so since this data set is pretty small we're likely to overfit with a powerful model (like this neural net).
# dropout layer dropout=0.5 print 'dropout probability', dropout if dropout > 0.0: h_drop = mx.sym.Dropout(data=h_pool, p=dropout) else: h_drop = h_pool
dropout probability 0.5
Finally, we add a fully connected layer to add non-linearity to the model. We then classify the resulting output of this layer using a softmax function, yielding a result between 0 (negative sentiment) and 1 (positive).
# fully connected layer num_label=2 cls_weight = mx.sym.Variable('cls_weight') cls_bias = mx.sym.Variable('cls_bias') fc = mx.sym.FullyConnected(data=h_drop, weight=cls_weight, bias=cls_bias, num_hidden=num_label) # softmax output sm = mx.sym.SoftmaxOutput(data=fc, label=input_y, name='softmax') # set CNN pointer to the "back" of the network cnn = sm
Now that we have defined our CNN model we will define the device on our machine that we will train and execute this model on, as well as the datasets to train and test this model with.
If you are running this code be sure that you have a GPU on your machine if your ctx is set to mx.gpu(0) otherwise you can set your ctx to mx.cpu(0) which will run the training much slower
from collections import namedtuple import time import math # Define the structure of our CNN Model (as a named tuple) CNNModel = namedtuple("CNNModel", ['cnn_exec', 'symbol', 'data', 'label', 'param_blocks']) # Define what device to train/test on ctx=mx.gpu(0) # If you have no GPU on your machine change this to # ctx=mx.cpu(0) arg_names = cnn.list_arguments() input_shapes = {} input_shapes['data'] = (batch_size, sentence_size) arg_shape, out_shape, aux_shape = cnn.infer_shape(**input_shapes) arg_arrays = [mx.nd.zeros(s, ctx) for s in arg_shape] args_grad = {} for shape, name in zip(arg_shape, arg_names): if name in ['softmax_label', 'data']: # input, output continue args_grad[name] = mx.nd.zeros(shape, ctx) cnn_exec = cnn.bind(ctx=ctx, args=arg_arrays, args_grad=args_grad, grad_req='add') param_blocks = [] arg_dict = dict(zip(arg_names, cnn_exec.arg_arrays)) initializer=mx.initializer.Uniform(0.1) for i, name in enumerate(arg_names): if name in ['softmax_label', 'data']: # input, output continue initializer(name, arg_dict[name]) param_blocks.append( (i, arg_dict[name], args_grad[name], name) ) out_dict = dict(zip(cnn.list_outputs(), cnn_exec.outputs)) data = cnn_exec.arg_dict['data'] label = cnn_exec.arg_dict['softmax_label'] cnn_model= CNNModel(cnn_exec=cnn_exec, symbol=cnn, data=data, label=label, param_blocks=param_blocks)
We can now execute the training and testing of our network, which in-part mxnet automatically does for us with its forward and backward propagation methods, along with its automatic gradient calculations.
''' Train the cnn_model using back prop ''' optimizer='rmsprop' max_grad_norm=5.0 learning_rate=0.0005 epoch=50 print 'optimizer', optimizer print 'maximum gradient', max_grad_norm print 'learning rate (step size)', learning_rate print 'epochs to train for', epoch # create optimizer opt = mx.optimizer.create(optimizer) opt.lr = learning_rate updater = mx.optimizer.get_updater(opt) # create logging output logs = sys.stderr # For each training epoch for iteration in range(epoch): tic = time.time() num_correct = 0 num_total = 0 # Over each batch of training data for begin in range(0, x_train.shape[0], batch_size): batchX = x_train[begin:begin+batch_size] batchY = y_train[begin:begin+batch_size] if batchX.shape[0] != batch_size: continue cnn_model.data[:] = batchX cnn_model.label[:] = batchY # forward cnn_model.cnn_exec.forward(is_train=True) # backward cnn_model.cnn_exec.backward() # eval on training data num_correct += sum(batchY == np.argmax(cnn_model.cnn_exec.outputs[0].asnumpy(), axis=1)) num_total += len(batchY) # update weights norm = 0 for idx, weight, grad, name in cnn_model.param_blocks: grad /= batch_size l2_norm = mx.nd.norm(grad).asscalar() norm += l2_norm * l2_norm norm = math.sqrt(norm) for idx, weight, grad, name in cnn_model.param_blocks: if norm > max_grad_norm: grad *= (max_grad_norm / norm) updater(idx, grad, weight) # reset gradient to zero grad[:] = 0.0 # Decay learning rate for this epoch to ensure we are not "overshooting" optima if iteration % 50 == 0 and iteration > 0: opt.lr *= 0.5 print >> logs, 'reset learning rate to %g' % opt.lr # End of training loop for this epoch toc = time.time() train_time = toc - tic train_acc = num_correct * 100 / float(num_total) # Saving checkpoint to disk if (iteration + 1) % 10 == 0: prefix = 'cnn' cnn_model.symbol.save('./%s-symbol.json' % prefix) save_dict = {('arg:%s' % k) :v for k, v in cnn_model.cnn_exec.arg_dict.items()} save_dict.update({('aux:%s' % k) : v for k, v in cnn_model.cnn_exec.aux_dict.items()}) param_name = './%s-%04d.params' % (prefix, iteration) mx.nd.save(param_name, save_dict) print >> logs, 'Saved checkpoint to %s' % param_name # Evaluate model after this epoch on dev (test) set num_correct = 0 num_total = 0 # For each test batch for begin in range(0, x_dev.shape[0], batch_size): batchX = x_dev[begin:begin+batch_size] batchY = y_dev[begin:begin+batch_size] if batchX.shape[0] != batch_size: continue cnn_model.data[:] = batchX cnn_model.cnn_exec.forward(is_train=False) num_correct += sum(batchY == np.argmax(cnn_model.cnn_exec.outputs[0].asnumpy(), axis=1)) num_total += len(batchY) dev_acc = num_correct * 100 / float(num_total) print >> logs, 'Iter [%d] Train: Time: %.3fs, Training Accuracy: %.3f \ --- Dev Accuracy thus far: %.3f' % (iteration, train_time, train_acc, dev_acc)
Now that we have gone through the trouble of training the model, we have stored the learned parameters in the .params file in our local directory. We can now load this file whenever we want and predict the sentiment of new sentences by running them through a forward pass of the trained model.