How to get the output of Conv layer and FC layer? about keras HOT 11 CLOSED

Hey, I think I can answer your questions.

1. In order to visualize the feature maps, just use the get_weights method of each layer and plot it using matplotlib. I have a working implementation of the same, and hope I'll be able to generalize it soon and create a PR.
2. For this, the best method that works for me is to create 2 models. Train the 1st model with your labels. The 2nd model is identical to the 1st except, it does not contain the last (or all fully connected) layer (don't forget to flatten). Using get_weights method above, get the weights of the 1st model and using set_weights assign it to the 2nd model. Then use predict_proba to get the feature vectors.

Hope it helps.

from keras.

As pointed out by pranv, all this is possible : )
To retrieve the output of any layer, you can also check out the answers here: #41

from keras.

Less headache for me was to just "deepcopy" the model and then just .pop() the last few layers and their parameters

from keras.

Thanks!
I try the method ｀save_weights｀after training the model:

model.save_weights("./model.hdf5")

And then create a new model of same architecture, init with the saved weights, and use it to predict the test_data:

model_load = create_model()
model_load.load_weights('./model.hdf5')
test_label = model_load.predict_classes(test_data)

But it shows error:

test_label = model_load.predict_classes(test_data)
 File "build/bdist.linux-x86_64/egg/keras/models.py", line 174, in predict_classes
 File "build/bdist.linux-x86_64/egg/keras/models.py", line 160, in predict_proba
AttributeError: 'Sequential' object has no attribute '_predict'

from keras.

did you also recompile the model?

from keras.

I added a functionality to the model so it can predict output for a given layer index.
I changed the 'compile' method and added a function named predict_layer. If you are interested i can post the code here so you add it

from keras.

I would be interested I've been fucking around with it for 5 hours with no luck

from keras.

Check the code i posted here. It works well but I would change the design if i had time ...
#456

from keras.

hi @wepe ,I use the graph model.
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.

from keras.

hi @pranv ,I use the graph model.
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.

from keras.

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 less painful I've seen.

from keras.