Say you have e.g. a boolean tf.placeholder, and you want to feed it when you call Model.fit. How would you do it? Below is some runnable dummy code that illustrates the problem.
import tensorflow as tf
from tensorflow.keras.layers import Dense, Input, Flatten
from tensorflow.keras.models import Model
# A boolean value that should have some effect of something
do_stuff = tf.placeholder(tf.bool)
# If do_stuff is true, return tf.ones else tf.zeros, and a 1 or 0 label
if_dostuff = lambda: [tf.ones((5, 5)), tf.constant(1)]
if_not_dostuff = lambda: [tf.zeros((5, 5)), tf.constant(0)]
X, Y_true = tf.cond(do_stuff, if_dostuff, if_not_dostuff)
# Make some dummy labels
# Do some random model operation
X_input = Input(shape=(5, 5))
layer_mod = Flatten()(X_input)
layer_mod = Dense(1)(layer_mod)
out_model = Model(inputs=[X_input], outputs=[layer_mod])
# Compile model
out_model.compile(
optimizer=tf.keras.optimizers.Adam(),
loss=tf.keras.metrics.binary_crossentropy
)
### Other ops with other models and summaries etc. ###
out_model.fit(...) # What do I do at this point?
Keep in mind that the boolean is just to keep things simple. In reality I have strings that are iterator handles that needs to be fed (based on what dataset I want to train).
How can I do keras' amazing model.fit interface with this sort of layout?
An alternative would be what I ask in this question
