I have written tensorflow code based on:
http://www.wildml.com/2015/12/implementing-a-cnn-for-text-classification-in-tensorflow/
but using precomputed word embeddings from the GoogleNews word2vec 300 dimension model.
I created my own data from the UCML News Aggregator Dataset in which I parsed the content of the news articles and have created my own labels.
Due to the size of the articles I use TF-IDF to filter out the top 120 words per article and embed those into 300 dimensions.
When I run the CNN I created regardless of the hyper parameters it converges to a small general accuracy, around 38%.
Hyper parameters changed:
Various filter sizes:
I've tried a single filter of 1,2,3 Combinations of filters [3,4,5], [1,3,4]
Learning Rate:
I've varied this from very low to very high, very low doesn't converge to 38% but anything between 0.0001 and 0.4 does.
Batch Size:
Tried many ranges between 5 and 100.
Weight and Bias Initialization:
Set stddev of weights between 0.4 and 0.01. Set bias initial values between 0 and 0.1. Tried using the xavier initializer for the conv2d weights.
Dataset Size:
I have only tried on two partial data sets, one with 15 000 training data, and the other on the 5000 test data. In total I have 263 000 data to train on. There is no accuracy difference whether trained and evaluated on the 15 000 training data or by using the 5000 test data as the training data (to save testing time).
I've run successful classifications on the 15 000 / 5000 split using a feed forward network with a BoW input (93% accurate), TF-IDF with SVM (92%), and TF-IDF with Native Bayes (91.5%). So I don't think it is the data.
What does this imply? Is the model just a poor model for this task? Is there an error in my work?
I feel like my do_eval function is incorrect to evaluate the accuracy / loss over an epoch of the data:
def do_eval(data_set,
label_set,
batch_size):
"""
Runs one evaluation against the full epoch of data.
data_set: The set of embeddings to eval
label_set: the set of labels to eval
"""
# And run one epoch of eval.
true_count = 0 # Counts the number of correct predictions.
steps_per_epoch = len(label_set) // batch_size
num_examples = steps_per_epoch * batch_size
totalLoss = 0
# Need to compute eval accuracy
for evalStep in xrange(steps_per_epoch):
input_batch, label_batch = nextBatch(data_set, labels_set, batchSize)
evalAcc, evalLoss = eval_step(input_batch, label_batch)
true_count += evalAcc * batchSize
totalLoss += evalLoss
precision = float(true_count) / num_examples
print(' Num examples: %d Num correct: %d Precision @ 1: %0.04f' % (num_examples, true_count, precision))
print("Eval Loss: " + str(totalLoss))
The entire model is as follows:
class TextCNN(object):
"""
A CNN for text classification
Uses a convolutional, max-pooling and softmax layer.
"""
def __init__(
self, batchSize, numWords, num_classes,
embedding_size, filter_sizes, num_filters):
# Set place holders
self.input_placeholder = tf.placeholder(tf.float32,[batchSize,numWords,embedding_size,1])
self.labels = tf.placeholder(tf.int32, [batchSize,num_classes])
self.pKeep = tf.placeholder(tf.float32)
# Inference
'''
Ready to build conv layers followed by max pooling layers
Each conv layer produces a different shaped output so need to loop over
them and create a layer for each and then merge the results
'''
pooled_outputs = []
for i, filter_size in enumerate(filter_sizes):
with tf.name_scope("conv-maxpool-%s" % filter_size):
# Convolution Layer
filter_shape = [filter_size, embedding_size, 1, num_filters]
# W: Filter matrix
W = tf.Variable(tf.truncated_normal(filter_shape,stddev=0.01), name='W')
b = tf.Variable(tf.constant(0.0,shape=[num_filters]),name="b")
# Valid padding: Narrow convolution (no edge padded so filter slides over everything)
# Output size = (input_size (numWords in this case) + 2 * padding (0 in this case) - filter_size) + 1
conv = tf.nn.conv2d(
self.input_placeholder,
W,
strides=[1, 1, 1, 1],
padding="VALID",
name="conv")
# Apply nonlinearity i.e add the bias to Wx + b
# Where Wx is the conv layer above
# Then run it through the activation function
h = tf.nn.relu(tf.nn.bias_add(conv, b),name='relu')
# Max-pooling over the outputs
# Max-pool to control the output size
# By taking only the best features determined by the filter
# Ksize is the size of the window of the input tensor
pooled = tf.nn.max_pool(
h,
ksize=[1, numWords - filter_size + 1, 1, 1],
strides=[1, 1, 1, 1],
padding='VALID',
name="pool")
# Each pooled outputs a tensor of size
# [batchSize, 1, 1, num_filters] where num_filters represents the
# Number of features we wanted pooled
pooled_outputs.append(pooled)
# Combine all pooled features
num_filters_total = num_filters * len(filter_sizes)
# Concat the pool output along the 3rd (num_filters / feature size) dimension
self.h_pool = tf.concat(pooled_outputs, 3)
# Flatten
self.h_pool_flat = tf.reshape(self.h_pool, [-1, num_filters_total])
# Add drop out to regularize the learning curve / accuracy
with tf.name_scope("dropout"):
self.h_drop = tf.nn.dropout(self.h_pool_flat,self.pKeep)
# Fully connected output layer
with tf.name_scope("output"):
W = tf.Variable(tf.truncated_normal([num_filters_total,num_classes],stddev=0.01),name="W")
b = tf.Variable(tf.constant(0.0,shape=[num_classes]), name='b')
self.logits = tf.nn.xw_plus_b(self.h_drop, W, b, name='logits')
self.predictions = tf.argmax(self.logits, 1, name='predictions')
# Loss
with tf.name_scope("loss"):
losses = tf.nn.softmax_cross_entropy_with_logits(labels=self.labels,logits=self.logits, name="xentropy")
self.loss = tf.reduce_mean(losses)
# Accuracy
with tf.name_scope("accuracy"):
correct_predictions = tf.equal(self.predictions, tf.argmax(self.labels,1))
self.accuracy = tf.reduce_mean(tf.cast(correct_predictions, "float"), name="accuracy")
##################################################################################################################
# Running the training
# Define various parameters for network
batchSize = 100
numWords = 120
embedding_size = 300
num_classes = 4
filter_sizes = [3,4,5] # slide over a the number of words, i.e 3 words, 4 words etc...
num_filters = 126
maxSteps = 5000
initial_learning_rate = 0.001
dropoutRate = 1
data_set = np.load("/home/kevin/Documents/NSERC_2017/articles/classifyDataSet/TestSmaller_CNN_inputMat_0.npy")
labels_set = np.load("Test_NN_target_smaller.npy")
with tf.Graph().as_default():
sess = tf.Session()
with sess.as_default():
cnn = TextCNN(batchSize=batchSize,
numWords=numWords,
num_classes=num_classes,
num_filters=num_filters,
embedding_size=embedding_size,
filter_sizes=filter_sizes)
# Define training operation
# Pick an optimizer, set it's learning rate, and tell it what to minimize
global_step = tf.Variable(0,name='global_step', trainable=False)
optimizer = tf.train.AdamOptimizer(initial_learning_rate)
grads_and_vars = optimizer.compute_gradients(cnn.loss)
train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step)
# Summaries to save for tensor board
# Set directory
out_dir = "/home/kevin/Documents/NSERC_2017/articles/classifyDataSet/tf_logs/CNN_Embedding/"
# Loss and accuracy summaries
loss_summary = tf.summary.scalar("loss",cnn.loss)
acc_summary = tf.summary.scalar("accuracy", cnn.accuracy)
# Train summaries
train_summary_op = tf.summary.merge([loss_summary,acc_summary])
train_summary_dir = out_dir + "train/"
train_summary_writer = tf.summary.FileWriter(train_summary_dir, sess.graph)
# Test summaries
test_summary_op = tf.summary.merge([loss_summary, acc_summary])
test_summary_dir = out_dir + "test/"
test_summary_write = tf.summary.FileWriter(test_summary_dir, sess.graph)
# Init all variables
init = tf.global_variables_initializer()
sess.run(init)
############################################################################################
def train_step(input_data, labels_data):
'''
Single training step
:param input_data: input
:param labels_data: labels to train to
'''
feed_dict = {
cnn.input_placeholder: input_data,
cnn.labels: labels_data,
cnn.pKeep: dropoutRate
}
_, step, summaries, loss, accuracy = sess.run(
[train_op, global_step, train_summary_op, cnn.loss, cnn.accuracy],
feed_dict=feed_dict)
train_summary_writer.add_summary(summaries, step)
###############################################################################################
def eval_step(input_data, labels_data, writer=None):
"""
Evaluates model on a test set
Single step
"""
feed_dict = {
cnn.input_placeholder: input_data,
cnn.labels: labels_data,
cnn.pKeep: 1.0
}
step, summaries, loss, accuracy = sess.run(
[global_step, test_summary_op, cnn.loss, cnn.accuracy],
feed_dict)
if writer:
writer.add_summary(summaries, step)
return accuracy, loss
###############################################################################
def nextBatch(data_set, labels_set, batchSize):
'''
Get the next batch of data
:param data_set: entire training or test data set
:param labels_set: entire training or test label set
:param batchSize: batch size
:return: a batch of the data and it's corresponding labels
'''
# Generate random row indices for the documents
rand_index = np.random.choice(data_set.shape[0], size=batchSize)
# Grab the data to give to the feed dicts
data_batch, labels_batch = data_set[rand_index, :, :], labels_set[rand_index, :]
# Resize for tensorflow
data_batch = data_batch.reshape([data_batch.shape[0],data_batch.shape[1],data_batch.shape[2],1])
return data_batch, labels_batch
################################################################################
def do_eval(data_set,
label_set,
batch_size):
"""
Runs one evaluation against the full epoch of data.
data_set: The set of embeddings to eval
label_set: the set of labels to eval
"""
# And run one epoch of eval.
true_count = 0 # Counts the number of correct predictions.
steps_per_epoch = len(label_set) // batch_size
num_examples = steps_per_epoch * batch_size
totalLoss = 0
# Need to compute eval accuracy
for evalStep in xrange(steps_per_epoch):
input_batch, label_batch = nextBatch(data_set, labels_set, batchSize)
evalAcc, evalLoss = eval_step(input_batch, label_batch)
true_count += evalAcc * batchSize
totalLoss += evalLoss
precision = float(true_count) / num_examples
print(' Num examples: %d Num correct: %d Precision @ 1: %0.04f' % (num_examples, true_count, precision))
print("Eval Loss: " + str(totalLoss))
######################################################################################################
# Training Loop
for step in range(maxSteps):
input_batch, label_batch = nextBatch(data_set,labels_set,batchSize)
train_step(input_batch,label_batch)
# Evaluate over the entire data set on last eval
if step % 100 == 0:
print "On Step : " + str(step) + " of " + str(maxSteps)
do_eval(data_set, labels_set,batchSize)
The embedding is done before the model:
def createInputEmbeddedMatrix(corpusPath, maxWords, svName):
# Create a [docNum, Words per Art, Embedding Size] matrix to fill
genDocsPath = "gen_docs_classifyData_smallerTest_TFIDF.npy"
# corpus = "newsCorpus_word2vec_All_Corpus.mm"
dictPath = 'news_word2vec_smallerDict.dict'
tf_idf_path = "news_tfIdf_word2vec_All.tfidf_model"
gen_docs = np.load(genDocsPath)
dictionary = gensim.corpora.dictionary.Dictionary.load(dictPath)
tf_idf = gensim.models.tfidfmodel.TfidfModel.load(tf_idf_path)
corpus = corpora.MmCorpus(corpusPath)
numOfDocs = len(corpus)
embedding_size = 300
id2embedding = np.load("smallerID2embedding.npy").item()
# Need to process in batches as takes up a ton of memory
step = 5000
totalSteps = int(np.ceil(numOfDocs / step))
for i in range(totalSteps):
# inputMatrix = scipy.sparse.csr_matrix([step,maxWords,embedding_size])
inputMatrix = np.zeros([step, maxWords, embedding_size])
start = i * step
end = start + step
for docNum in range(start, end):
print "On docNum " + str(docNum) + " of " + str(numOfDocs)
# Extract the top N words
topWords, wordVal = tf_idfTopWords(docNum, gen_docs, dictionary, tf_idf, maxWords)
# doc = corpus[docNum]
# Need to track word dex and doc dex seperate
# Doc dex because of the batch processing
wordDex = 0
docDex = 0
for wordID in wordVal:
inputMatrix[docDex, wordDex, :] = id2embedding[wordID]
wordDex += 1
docDex += 1
# Save the batch of input data
# scipy.sparse.save_npz(svName + "_%d" % i, inputMatrix)
np.save(svName + "_%d.npy" % i, inputMatrix)
#####################################################################################
