Keras: Using model output as the input for another: When feeding symbolic tensors to a model, we expect thetensors to have a static batch size - keras

Question

I have the following two models, where model_A is trained first, then the output of model_A is used to train the model_C:
import keras
from keras.layers import Input, Dense
from keras.models import Model
inputs = Input(shape=(12,))
# ---------------------------------------
# model_A
x = Dense(64, activation='relu')(inputs)
x = Dense(64, activation='relu')(x)
predictions_A = Dense(3, activation='softmax')(x)
model_A = Model(inputs=inputs, outputs=predictions_A)
model_A.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
model_A.fit(my_data_x, axis = 1), pd.get_dummies(my_data['target_cate'],prefix=['cate_']))
#----------------------------------------
input_C_out_A = Input(shape=(3,))
# Concatenating the two input layers
concat = keras.layers.concatenate([inputs, input_C_out_A])
x1 = Dense(64, activation='relu')(concat)
x1 = Dense(64, activation='relu')(x1)
predictions_C= Dense(1, activation='sigmoid')(x1)
model_C = Model(inputs=[inputs, input_C_out_A], outputs=predictions_C)
model_C.compile(loss='mean_squared_error', optimizer='adam')
model_C.fit([my_data_x,predictions_A], my_data['target_numeric'])
model_A training seems to be fine, but then I got the following errors when training the model_C:
Epoch 1/1
374667/374667 [==============================] - 11s 30us/step - loss: 0.3157 - acc: 0.9119
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-78-8df7b1dec93f> in <module>
28 model_C = Model(inputs=[inputs, input_C_out_A], outputs=predictions_C)
29 model_C.compile(loss='mean_squared_error', optimizer='adam')
---> 30 model_C.fit([my_data_x,predictions_A], my_data['target_numeric'])
~/workspace/git/tensorplay/venv/lib/python3.7/site-packages/keras/engine/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)
950 sample_weight=sample_weight,
951 class_weight=class_weight,
--> 952 batch_size=batch_size)
953 # Prepare validation data.
954 do_validation = False
~/workspace/git/tensorplay/venv/lib/python3.7/site-packages/keras/engine/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, check_array_lengths, batch_size)
749 feed_input_shapes,
750 check_batch_axis=False, # Don't enforce the batch size.
--> 751 exception_prefix='input')
752
753 if y is not None:
~/workspace/git/tensorplay/venv/lib/python3.7/site-packages/keras/engine/training_utils.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)
90 data = data.values if data.__class__.__name__ == 'DataFrame' else data
91 data = [data]
---> 92 data = [standardize_single_array(x) for x in data]
93
94 if len(data) != len(names):
~/workspace/git/tensorplay/venv/lib/python3.7/site-packages/keras/engine/training_utils.py in <listcomp>(.0)
90 data = data.values if data.__class__.__name__ == 'DataFrame' else data
91 data = [data]
---> 92 data = [standardize_single_array(x) for x in data]
93
94 if len(data) != len(names):
~/workspace/git/tensorplay/venv/lib/python3.7/site-packages/keras/engine/training_utils.py in standardize_single_array(x)
23 'When feeding symbolic tensors to a model, we expect the'
24 'tensors to have a static batch size. '
---> 25 'Got tensor with shape: %s' % str(shape))
26 return x
27 elif x.ndim == 1:
ValueError: When feeding symbolic tensors to a model, we expect thetensors to have a static batch size. Got tensor with shape: (None, 3)
Any idea what I missed? Thanks!

Answer 1

The model_C works by:
Giving the input data to model_A,
Getting output of model_A, and
Feeding that, along with the original inputs, to the first Dense layer.
So implement just what you said (and try to keep models separate, i.e. each with its own input/output layers):
input_C = Input(shape=(12,))
out_A = model_A(input_C) # get the output of model_A
concat = keras.layers.concatenate([input_C, out_A])
x1 = Dense(64, activation='relu')(concat)
x1 = Dense(64, activation='relu')(x1)
predictions_C= Dense(1, activation='sigmoid')(x1)
model_C = Model(inputs=input_C, outputs=predictions_C)
model_C.compile(loss='mean_squared_error', optimizer='adam')
model_C.fit(my_data_x, my_data['target_numeric'])
If you don't want the model_A to be trained when training model_C (i.e. if you have already trained model_A and don't want its weights to be changed), just set model_A.trainable = False before compiling model_C.

Answer 2

It makes no sense to put the output of a model (a symbolic tensor) into model.fit, since there is no input data there. You should first obtain predictions from model A and then use them to fit model C:
pred_a = model_A.predict(my_data_x)
model_C.fit([my_data_x, pred_a], my_data['target_numeric'])