keract 2.4.0

Keras Activations and Gradients

Keract: Keras Activations + Gradients

pip install keract

You have just found a (easy) way to get the activations (outputs) and gradients for each layer of your Keras model (LSTM, conv nets...).

API

• get_activations
• get_gradients_of_trainable_weights
• get_gradients_of_activations

Get activations (outputs of each layer)

from keract import get_activations
get_activations(model, x)

Inputs are:

• model is a keras.models.Model object.
• x is a numpy array to feed to the model as input. In the case of multi-input, x is of type List. We use the Keras convention (as used in predict, fit...).

The output is a dictionary containing the activations for each layer of model for the input x:

{
  'conv2d_1/Relu:0': np.array(...),
  'conv2d_2/Relu:0': np.array(...),
  ...,
  'dense_2/Softmax:0': np.array(...)
}

The key is the name of the layer and the value is the corresponding output of the layer for the given input x.

Get gradients of weights

• model is a keras.models.Model object.
• x Input data (numpy array). Keras convention.
• y: Labels (numpy array). Keras convention.

from keract import get_gradients_of_trainable_weights
get_gradients_of_trainable_weights(model, x, y)

The output is a dictionary mapping each trainable weight to the values of its gradients (regarding x and y).

Get gradients of activations

• model is a keras.models.Model object.
• x Input data (numpy array). Keras convention.
• y: Labels (numpy array). Keras convention.

from keract import get_gradients_of_activations
get_gradients_of_activations(model, x, y)

The output is a dictionary mapping each layer to the values of its gradients (regarding x and y).

Examples

Examples are provided for:

• keras.models.Sequential - mnist.py
• keras.models.Model - multi_inputs.py
• Recurrent networks - recurrent.py

In the case of MNIST with LeNet, we are able to fetch the activations for a batch of size 128:

conv2d_1/Relu:0
(128, 26, 26, 32)

conv2d_2/Relu:0
(128, 24, 24, 64)

max_pooling2d_1/MaxPool:0
(128, 12, 12, 64)

dropout_1/cond/Merge:0
(128, 12, 12, 64)

flatten_1/Reshape:0
(128, 9216)

dense_1/Relu:0
(128, 128)

dropout_2/cond/Merge:0
(128, 128)

dense_2/Softmax:0
(128, 10)

We can visualise the activations. Here's another example using VGG16:

cd examples
python vgg16.py

A cat.

Outputs of the first convolutional layer of VGG16.

Also, we can visualise the heatmaps of the activations:

cd examples
python heat_map.py

Repo views (since 2018/10/31)