问题
I want to use tf.keras.TimeDistributed() layer with the tf.hub inception_v3 CNN model from the latest TensorFLow V2 version (tf-nightly-gpu-2.0-preview). The output is shown below. It seemst that tf.keras.TimeDistributed() is not fully implemented to work with tf.hub models. Somehow, the shape of the input layer cannot be computed. My question: Is there a workaround this problem?
tf.keras.TimeDistributed with regular tf.keras.layer works fine. I just would like to apply the CNN model to each time step.
Model
import tensorflow as tf
import tensorflow_hub as hub
from tensorflow.keras import layers, Model
model_url = "https://tfhub.dev/google/tf2-
preview/inception_v3/feature_vector/3"
feature_layer = hub.KerasLayer(model_url,
input_shape = (299, 299, 3),
output_shape = [2048],
trainable = False)
video = layers.Input(shape = (None, 299, 299, 3))
encoded_frames = layers.TimeDistributed(feature_layer)(video)
model = Model(inputs = video, outputs = encoded_frames)
Expected output
tf.keras model
Error messages
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/keras/engine/base_layer.py", line 489, in compute_output_shape raise NotImplementedError NotImplementedError
回答1:
In Tensorflow 2 it is possible to use custom layers in combination with the TimeDistributed layer. The error is thrown because it can't compute the output shape (see here).
So in your case you should be able to subclass KerasLayer and implement compute_output_shape manually.
回答2:
Wrapper Layers like TimeDistributed require a layer instance to be passed. If you build the model out of custom layers, you'll need to at least wrap them in tf.keras.layers.Lambda. This might not be possible in your case of models from hub.KerasLayer, so you might consider the solutions posted here:
TimeDistributed of a KerasLayer in Tensorflow 2.0
来源：https://stackoverflow.com/questions/56173992/keras-layers-timedistributed-with-hub-keraslayer-notimplementederror