I have a Keras model that has in input shape of [None, 500, 500, 3] and output shape of [None, 1]. Now, I want to make a wrapper model that has an input shape of [None, 48, 500, 500, 3], and output shape of [None, 48].
To do that, the naive way is to iterate 48 times on the second axis and apply the first model, then use Keras' Concatenate layer to get the desired shape.
model_outputs = []
for i in range(inputs.shape[1]):
im_block = inputs[:, i]
model_outputs += [self.model(im_block)]
return Concatenate()(model_outputs)
However, this makes the graph quite complicated. So I would like to do the following instead:
[None, 48, 500, 500, 3]
-> [None*48, 500, 500, 3]
(apply the model)
-> [None*48, 1]
-> [None, 48, 1]
My attempt at that is
outputs = tf.reshape(inputs, (inputs[0] * inputs[1], *inputs[2:]))
outputs = self.model(outputs)
outputs = tf.reshape(outputs, (inputs[0], inputs[1]))
return outputs
But this gives me
TypeError: Cannot iterate over a tensor with unknown first dimension.
Is there a way to do that ?
This should work:
inp = tf.reshape(inp, (-1, 500, 500, 3))
res = model(inp)
res = tf.reshape(res, (-1, 48))
You're right, I just forgot
.shapeand was using the tensor's value, which don't exist..