How to configure and train the model using Glove and CNN for text classification?

Question

I have worked with text classification using Glove and CNN and found the problem below:

File "c:\programfiles_anaconda\anaconda3\envs\math_stat_class\lib\site-packages\tensorflow\python\framework\ops.py", line 1657, in _create_c_op
    raise ValueError(str(e))

ValueError: Negative dimension size caused by subtracting 5 from 1 for '{{node max_pooling1d_9/MaxPool}} = MaxPool[T=DT_FLOAT, data_format="NHWC", ksize=[1, 5, 1, 1], padding="VALID", strides=[1, 5, 1, 1]](max_pooling1d_9/ExpandDims)' with input shapes: [?,1,1,128].

Glove input

EMBEDDING_DIM = 100
    
embeddings_index = {}
    
f = open(glove_path, encoding='utf-8')  
for line in f:    
    values = line.split()
    word = values[0]
    coefs = np.asarray(values[1:], dtype='float32')
    embeddings_index[word] = coefs
    f.close()
    
print('Found %s word vectors.' % len(embeddings_index))
    
embedding_matrix = np.zeros((len(word_index) + 1, EMBEDDING_DIM))
    
for word, i in word_index.items():
    embedding_vector = embeddings_index.get(word)
    if embedding_vector is not None:
        # words not found in embedding index will be all-zeros.
        embedding_matrix[i] = embedding_vector

Layer input of CNN

# apply embedding matrix into an Embedding layer
# trainable=False to prevent the weights from being updated during training
embedding_layer = Embedding(len(word_index) + 1,
                            EMBEDDING_DIM,
                            weights=[embedding_matrix],
                            input_length=MAX_SEQUENCE_LENGTH,
                            trainable=False)

Training 1D CNN

sequence_input = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='int32')
embedded_sequences = embedding_layer(sequence_input)

x = Conv1D(128, 5, activation='relu')(embedded_sequences)   
print("x shape = ", x)

x = MaxPooling1D(5)(x)  
print("x shape = ", x)
        
x = Conv1D(128, 5, activation='relu')(x)
print("x shape = ", x)
    
#-----This line below produced error-----
x = MaxPooling1D(5)(x) #Error this line
#-----This line above produced error-----
        
print("x shape = ", x)

x = Conv1D(128, 5, activation='relu')(x)
print("x shape = ", x)
    
x = MaxPooling1D(35)(x)  # global max pooling
print("x shape = ", x)
    
x = Flatten()(x)
x = Dense(128, activation='relu')(x)
    
preds = Dense(len(labels_index), activation='softmax')(x)
model = Model(sequence_input, preds)
model.compile(loss='categorical_crossentropy',
              optimizer='rmsprop',
              metrics=['acc'])
    
# Learning
model.fit(X_train, y_train, validation_data=(X_val, y_val),
          epochs=2, batch_size=128)

My ideas

1) Are there some issues/problems with Glove input?

2) Conv1D:

• Change the "kernel_size" from 5 to a New Value.

3) MaxPooling1D:

• Change pool_size from 5 to a New Value.
• Specify other parameters: strides, padding and so on.

4) I currently use keras on tensorflow 2.20 and python 3.6

• Do I need to upgrade tensorflow and python?

However, I could not figure out any better way to do. May I have your suggestions?

Answer 1

Two things that come to my mind: your max-pooling layers are reducing the size of the input to the next convolutional layers every time and eventually the size is too small to run another max-pooling operation. Try running

 tf.print(model.summary) 

after each max-pooling operation and you will quickly find out that your tensor cannot be further reduced. You can then consider using a different pool_size in your max-pooling layers.

The second thing I notice (I am not sure if it is intentional), but MaxPooling1D != Global Max Pooling . Keras supports both operations . Take a look at the documentation.

On a side note, sentence classification with CNNs was widely popularized by the work of Yoon Kim. In his work, he shows that global max-pooling operations perform much better than striding max-pooling operations in sentence classification (when using word embeddings, as you are doing).