I want to do multi label classification (20 distinct output labels), based on vectorized words using TfidfVectorizer. I have set of 39974 rows each one containing 2739 items (zeros or ones).
I would like to classify this data using Keras model which will contain 1 hidden layer (~20 nodes with activation='relu') and output layer equal 20 possible output values (with activation='softmax' to choose best fit).
Here's my code so far:
model = Sequential()
model.add(Dense(units=20, activation='relu', input_shape=tfidf_matrix.shape))
model.add(Dense(units=20, activation='softmax'))
model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
model.fit(tfidf_matrix, train_data['cuisine_id'], epochs=10)
But got error:
ValueError: Error when checking input: expected dense_1_input to have 3 dimensions, but got array with shape (39774, 2739)
How can I specify this NN to fit using this matrix?
The number of rows (number of training samples) is not the part of the input shape of the network because the training process feeds the network one sample per batch (or, more precisely, batch_size samples per batch).
So in your case, input shape of the network is (2739, ) and the right code should be like this:
model = Sequential()
# the shape of one training example is
input_shape = tfidf_matrix[0].shape
model.add(Dense(units=20, activation='relu', input_shape=input_shape))
model.add(Dense(units=20, activation='softmax'))
model.compile(optimizer='rmsprop', loss='categorical_crossentropy',
metrics=['accuracy'])
model.fit(tfidf_matrix, train_data['cuisine_id'], epochs=10)