How to apply Guided BackProp in Tensorflow 2.0?

Question

I am starting with Tensorflow 2.0 and trying to implement Guided BackProp to display Saliency Map. I started by computing the loss between y_pred and y_true of an image, then find gradients of all layers due to this loss.

with tf.GradientTape() as tape:
    logits = model(tf.cast(image_batch_val, dtype=tf.float32))
    print('`logits` has type {0}'.format(type(logits)))
    xentropy = tf.nn.softmax_cross_entropy_with_logits(labels=tf.cast(tf.one_hot(1-label_batch_val, depth=2), dtype=tf.int32), logits=logits)
    reduced = tf.reduce_mean(xentropy)
    grads = tape.gradient(reduced, model.trainable_variables)

However, I don't know what to do with gradients in order to obtain the Guided Propagation.

This is my model. I created it using Keras layers:

image_input = Input((input_size, input_size, 3))

conv_0 = Conv2D(32, (3, 3), padding='SAME')(image_input)
conv_0_bn = BatchNormalization()(conv_0)
conv_0_act = Activation('relu')(conv_0_bn)
conv_0_pool = MaxPool2D((2, 2))(conv_0_act)

conv_1 = Conv2D(64, (3, 3), padding='SAME')(conv_0_pool)
conv_1_bn = BatchNormalization()(conv_1)
conv_1_act = Activation('relu')(conv_1_bn)
conv_1_pool = MaxPool2D((2, 2))(conv_1_act)

conv_2 = Conv2D(64, (3, 3), padding='SAME')(conv_1_pool)
conv_2_bn = BatchNormalization()(conv_2)
conv_2_act = Activation('relu')(conv_2_bn)
conv_2_pool = MaxPool2D((2, 2))(conv_2_act)

conv_3 = Conv2D(128, (3, 3), padding='SAME')(conv_2_pool)
conv_3_bn = BatchNormalization()(conv_3)
conv_3_act = Activation('relu')(conv_3_bn)

conv_4 = Conv2D(128, (3, 3), padding='SAME')(conv_3_act)
conv_4_bn = BatchNormalization()(conv_4)
conv_4_act = Activation('relu')(conv_4_bn)
conv_4_pool = MaxPool2D((2, 2))(conv_4_act)

conv_5 = Conv2D(128, (3, 3), padding='SAME')(conv_4_pool)
conv_5_bn = BatchNormalization()(conv_5)
conv_5_act = Activation('relu')(conv_5_bn)

conv_6 = Conv2D(128, (3, 3), padding='SAME')(conv_5_act)
conv_6_bn = BatchNormalization()(conv_6)
conv_6_act = Activation('relu')(conv_6_bn)

flat = Flatten()(conv_6_act)

fc_0 = Dense(64, activation='relu')(flat)
fc_0_bn = BatchNormalization()(fc_0)

fc_1 = Dense(32, activation='relu')(fc_0_bn)
fc_1_drop = Dropout(0.5)(fc_1)

output = Dense(2, activation='softmax')(fc_1_drop)

model = models.Model(inputs=image_input, outputs=output)

I am glad to provide more code if needed.

Answer 1

First of all, you have to change the computation of the gradient through a ReLU, i.e.

Here a graphic example from the paper.

This formula can be implemented with the following code:

@tf.RegisterGradient("GuidedRelu")
def _GuidedReluGrad(op, grad):
   gate_f = tf.cast(op.outputs[0] > 0, "float32") #for f^l > 0
   gate_R = tf.cast(grad > 0, "float32") #for R^l+1 > 0
   return gate_f * gate_R * grad

Now you have to override the original TF implementation of ReLU with:

with tf.compat.v1.get_default_graph().gradient_override_map({'Relu': 'GuidedRelu'}):
   #put here the code for computing the gradient

After computing the gradient, you can visualize the result. However, one last remark. You compute a visualization for a single class. This means, you take the activation of a choosen neuron and set all the activations of the other neurons to zero for the input of Guided BackProp.