Extracting embeddings from layers

[anjishnu]

Embeddings obtained from training a discriminative NN towards a specific task can be extremely useful on related tasks (e.g. Transfer learning). We can extract a lot of potentially useful embeddings by looking at the weights of a layer of the model. By looking at the documentation http://keras.io/models/ it doesn't seem like Keras supports the abstraction of extracting weight values from individual layers. Seems like it would be relatively easy to implement.

[Smerity]

From the skipgram word embeddings example:

# recover the embedding weights trained with skipgram:
weights = model.layers[0].get_weights()[0]

If instead you're looking to extract the hidden layer representation of a given input, refer to #41

[alyato]

hi @Smerity ,i use the graph model,refer to #41
like this,

model = graph()
model.add_input(name='input0',input_shape=())
model.add_node(Convolution2D(),name='c1',input='input0')
.......

And i want to see the output of the c1,Then

getFeatureMap = theano.function(model.inputs['input0'].input,model.nodes['c1'].get_output(train=False),
allow_input_downcast=True)

But it show me that
TypeError: list indices must be integers, not str

Do you give me some advices? Thanks.

[philipperemy]

More generally you can visualise the output/activations of every layer of your model. I wrote an example with MNIST to show how here:

https://github.com/philipperemy/keras-visualize-activations

So far it's the least painful I've seen.