Visualizing learned filters at each layer

[Raunaq Vohra]

Hi, I was wondering whether there's a way to visualize the filters my model has learned? I have the saved weights (from model.save_weights(), after model.evaluate()) and I also dump them at the end of my code, like so:

all_weights = []
for layer in model.layers:
   w = layer.get_weights()
   all_weights.append(w)

all_weights = np.array(all_weights)
np.save('all_weights.npy', all_weights)

How can I use them?

Thanks!

[François Chollet]

These weights are regular numpy arrays, which you can plot using Matplotlib for instance (or other libraries).

You can take a look at: https://github.com/julienr/ipynb_playground/blob/master/keras/convmnist/keras_cnn_mnist.ipynb

--
You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/e9c2a856-331f-4a32-8b1d-3b84ec8627b2%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Raunaq Vohra]

Thanks François, exactly what I was looking for.

[csi...@altius.org]

So I think that the example here is no longer valid.

I used a 1D convolutional layer for my first layer:

```

    model = Sequential()

    model.add(Conv1D(3, 5, input_shape=(None, 4)))

```

then if I do

```

filters = model.layers[0].get_weights()[0]

np.shape(filters)

```

I get a shape of `(5, 4, 3)`! So the first argument is no longer the number of filters, rather filter i is `filters[:, :, i]`. Is that right?

Thanks.

[pr...@cogent.co.jp]

Have a look at this project: https://github.com/philipperemy/keras-visualize-activations

[abhig...@gmail.com]

On a similar note , in a given layer for number of filters how to extract / visualize the filter which has the maximum magnitude or strongest activation map in the given filter layer .  

[Ted Yu]

https://blog.keras.io/how-convolutional-neural-networks-see-the-world.html

https://jacobgil.github.io/deeplearning/filter-visualizations

--
You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.

To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/304df4e9-7f0a-4315-9927-edd237ab46f3%40googlegroups.com.