tensorflow/keras lstm input shape

Question

I am trying to learn the keras functional API through the tutorials from Keras, and when I try to modify the example, I seem to get a shape mismatch. The only difference between the tutorial code and the one below is that I remove the embedding layer since mine is a regression problem.

Firstly, I am aware that LSTM expects 3 dimensions. In my example, I have:

TRAIN_BATCH_SIZE=32
MODEL_INPUT_BATCH_SIZE=128

headline_data = np.random.uniform(low=1, high=9000, size=(MODEL_INPUT_BATCH_SIZE, 100)).astype(np.float32)
additional_data = np.random.uniform(low=1, high=9000, size=(MODEL_INPUT_BATCH_SIZE, 5)).astype(np.float32)
labels = np.random.randint(0, 1 + 1, size=(MODEL_INPUT_BATCH_SIZE, 1))

main_input = Input(shape=(100,), dtype='float32', name='main_input')

lstm_out = LSTM(32)(main_input)

auxiliary_output = Dense(1, activation='sigmoid', name='aux_output')(lstm_out)

auxiliary_input = Input(shape=(5,), name='aux_input')
x = keras.layers.concatenate([lstm_out, auxiliary_input])

# We stack a deep densely-connected network on top
x = Dense(64, activation='relu')(x)
x = Dense(64, activation='relu')(x)
x = Dense(64, activation='relu')(x)

# And finally we add the main logistic regression layer
main_output = Dense(1, activation='sigmoid', name='main_output')(x)

# This defines a model with two inputs and two outputs:
model = Model(inputs=[main_input, auxiliary_input], outputs=[main_output, auxiliary_output])

model.compile(optimizer='rmsprop',
                loss={'main_output': 'binary_crossentropy', 'aux_output': 'binary_crossentropy'},
                loss_weights={'main_output': 1., 'aux_output': 0.2})

# And trained it via:
model.fit({'main_input': headline_data, 'aux_input': additional_data},
                {'main_output': labels, 'aux_output': labels},
                epochs=2, batch_size=TRAIN_BATCH_SIZE)

When I run the above, I get:

ValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=2

So, I tried changing my input shape like so:

main_input = Input(shape=(100,1), dtype='float64', name='main_input')

and when I run this, I get:

ValueError: Error when checking input: expected main_input to have 3 dimensions, but got array with shape (128, 100)

I am perplexed and lost as to where the error is coming from. Would really appreciate some guidance on this.

EDIT

I have also tried setting:

headline_data = np.expand_dims(headline_data, axis=2)

and then used,

main_input = Input(shape=headline_data.shape, dtype='float64', name='main_input')

then, I get:

ValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=4

seems really strange!

Answer 1

ValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=2

Your problem is with the shape of your data.

headline_data = np.random.uniform(low=1, high=9000, size=(MODEL_INPUT_BATCH_SIZE, 100))
headline_data.shape

returns

(128,100)

However this should have three dimensions.

Without doublechecking you probably need to do something like:

headline_data.reshape(128,1,100)

Have a look at this post, it should clear everything up.

Link

* UPDATE *

Do the following:

headling_data = healdine_data.reshape(128,1,100)
main_input = Input(shape=(1,100), dtype='float32', name='main_input')

I tested it and it works, so let me know if it doesnt for you =)

---- Complete Code: ----

import numpy as np

from tensorflow import keras
from tensorflow.keras import Model
from tensorflow.keras.layers import Input, LSTM, Dense


TRAIN_BATCH_SIZE=32
MODEL_INPUT_BATCH_SIZE=128

headline_data = np.random.uniform(low=1, high=9000, size=(MODEL_INPUT_BATCH_SIZE, 100)).astype(np.float32)
headline_data.shape
lstm_data = headline_data.reshape(MODEL_INPUT_BATCH_SIZE,1,100)
additional_data = np.random.uniform(low=1, high=9000, size=(MODEL_INPUT_BATCH_SIZE, 5)).astype(np.float32)
labels = np.random.randint(0, 1 + 1, size=(MODEL_INPUT_BATCH_SIZE, 1))

main_input = Input(shape=(1,100), dtype='float32', name='main_input')

lstm_out = LSTM(32)(main_input)

auxiliary_output = Dense(1, activation='sigmoid', name='aux_output')(lstm_out)

auxiliary_input = Input(shape=(5,), name='aux_input')
x = keras.layers.concatenate([lstm_out, auxiliary_input])

# We stack a deep densely-connected network on top
x = Dense(64, activation='relu')(x)
x = Dense(64, activation='relu')(x)
x = Dense(64, activation='relu')(x)

# And finally we add the main logistic regression layer
main_output = Dense(1, activation='sigmoid', name='main_output')(x)

# This defines a model with two inputs and two outputs:
model = Model(inputs=[main_input, auxiliary_input], outputs=[main_output, auxiliary_output])

model.compile(optimizer='rmsprop',
                loss={'main_output': 'binary_crossentropy', 'aux_output': 'binary_crossentropy'},
                loss_weights={'main_output': 1., 'aux_output': 0.2})

# And trained it via:
model.fit({'main_input': lstm_data, 'aux_input': additional_data},
                {'main_output': labels, 'aux_output': labels},
                epochs=1000, batch_size=TRAIN_BATCH_SIZE)