saliency is NaN for VGG16 like model with BatchNorm

[sverneka]

saliency = saliency_fn([preprocessed_input, 0]) this gives me NaN for valid inputs. Below is my modified VGG16 network with BatchNorm.

def VGG16(input_shape=(224,224,3),classes=10):
input_img = Input(shape=input_shape)
# Block 1
x=Conv2D(64, (3, 3), activation='relu', padding='same', input_shape= input_shape, name='block1_conv1', kernel_initializer='glorot_normal')(input_img)
x=Conv2D(64, (3, 3), activation='relu', padding='same', name='block1_conv2', kernel_initializer='glorot_normal')(x)
x=MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x)
x=BatchNormalization()(x)
# Block 2
x=Conv2D(128, (3, 3), activation='relu', padding='same', name='block2_conv1', kernel_initializer='glorot_normal')(x)
x=Conv2D(128, (3, 3), activation='relu', padding='same', name='block2_conv2', kernel_initializer='glorot_normal')(x)
x=MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool')(x)
x=BatchNormalization()(x)
# Block 3
x=Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv1', kernel_initializer='glorot_normal')(x)
x=Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv2', kernel_initializer='glorot_normal')(x)
x=Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv3', kernel_initializer='glorot_normal')(x)
x=MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool')(x)
x=BatchNormalization()(x)
# Block 4
x=Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv1', kernel_initializer='glorot_normal')(x)
x=Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv2', kernel_initializer='glorot_normal')(x)
x=Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv3', kernel_initializer='glorot_normal')(x)
x=MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool')(x)
x=BatchNormalization()(x)
# Block 5
x=Conv2D(512, (3, 3), activation='relu', padding='same', name='block5_conv1', kernel_initializer='glorot_normal')(x)
x=Conv2D(512, (3, 3), activation='relu', padding='same', name='block5_conv2', kernel_initializer='glorot_normal')(x)
x=Conv2D(512, (3, 3), activation='relu', padding='same', name='block5_conv3', kernel_initializer='glorot_normal')(x)
x=MaxPooling2D((2, 2), strides=(2, 2), name='block5_pool')(x)
# Classification block
x=Flatten(name='flatten')(x)
x=Dense(256, activation='relu', name='fc1', kernel_initializer='glorot_normal')(x)
#model.add(Dropout(0.5))
x=Dense(128, activation='relu', name='fc2', kernel_initializer='glorot_normal')(x)
#model.add(Dropout(0.5))
x=Dense(classes, activation='softmax', name='predictions', kernel_initializer='glorot_normal')(x)
model = Model(input_img, x)
model.summary()
#model = Model(inputs, x, name='vgg16')
return model

[sverneka]

There was an issue with the model itself, probably some weights were NaN. The issue is now resolved.