Re-training a Hub module as a Siamese network

[catalla]

87

When re-training a hub module (universal sentence encoder) as a Siamese network, are gradients guaranteed to propagate correctly even if the module is run twice shown above?

We cannot specify 'reuse=True' so, it seems unclear whether applying the module a second time will erase the state after the first time.

m = hub.Module("https://tfhub.dev/google/universal-sentence-encoder/2", trainable=True)
embedding1 = m(query_1)
embedding2 = m(query_2)

loss = tf.square(embedding1 - embedding2) # Some loss function here
optimizer = tf.train.GradientDescentOptimizer().minimize(loss)
sess.run(optimizer)

[andresusanopinto]

Hi @catalla,

Yes the gradients will propagate via both call() into the same variables as the variables of a module are created during the constructor and they are shared across call().

From the "Applying a module" section from the user-guide:

"""If this involves variables with trained weights, these are shared between all applications."""

Let us know if you have ideas on how to improve that documentation to make it more clear.

[catalla]

"""If this involves variables with trained weights, these are shared between all applications."""

This makes the behavior clear for applying a model with trained weights. However, it could be clarifying to mention that the same applies for gradients when re-training a hub module with trainable weights.

Thanks for the response!

[Killthebug]

I'm trying to duplicate the method proposed by @catalla. Though when I try train the model, it prompts me with the error : ValueError: No gradients provided for any variable, check your graph for ops that do not support gradients... I've set trainable=True. Is there any element I'm missing or should account for (I've always developed in PyTorch and am just experimenting with Tensorflow, so please pardon any novice mistakes on my end.)