def test_gradients_of_activations(self):
model, x = dummy_model_and_inputs()
# important to leave the compile() call to the user. Gradients need this correct.
model.compile(loss='mse', optimizer='adam')
y = np.random.uniform(size=len(x))
grad_acts = get_gradients_of_activations(model, x, y)
acts = get_activations(model, x)
grad_acts_nested = get_gradients_of_activations(model,
x,
y,
nested=True)
acts_nested = get_activations(model, x, nested=True)
# same support.
self.assertListEqual(list(acts), list(grad_acts))
self.assertListEqual(list(grad_acts['i1'].shape),
list(acts['i1'].shape))
self.assertListEqual(list(grad_acts['model'].shape),
list(acts['model'].shape))
self.assertListEqual(list(grad_acts['block'].shape),
list(acts['block'].shape))
self.assertListEqual(list(grad_acts['fc1'].shape),
list(acts['fc1'].shape))
self.assertListEqual(list(acts_nested), list(grad_acts_nested))
self.assertListEqual(list(grad_acts_nested['i1'].shape),
list(acts_nested['i1'].shape))
self.assertListEqual(list(grad_acts_nested['model/fc1'].shape),
list(acts_nested['model/fc1'].shape))
self.assertListEqual(list(grad_acts_nested['model/fc1'].shape),
list(acts_nested['model/fc1'].shape))
self.assertListEqual(list(grad_acts_nested['block/fc1'].shape),
list(acts_nested['block/fc1'].shape))
self.assertListEqual(list(grad_acts_nested['block/relu'].shape),
list(acts_nested['block/relu'].shape))
self.assertListEqual(list(grad_acts_nested['fc1'].shape),
list(acts_nested['fc1'].shape))
def main():
np.random.seed(123)
inp_a = np.random.uniform(size=(5, 10))
inp_b = np.random.uniform(size=(5, 10))
out_c = np.random.uniform(size=(5, 1))
# Just for visual purposes.
np.set_printoptions(precision=2)
# Activations of all the layers
print('MULTI-INPUT MODEL')
m1 = get_multi_inputs_model()
m1.compile(optimizer='adam', loss='mse')
utils.print_names_and_values(keract.get_activations(m1, [inp_a, inp_b]))
utils.print_names_and_values(
keract.get_gradients_of_trainable_weights(m1, [inp_a, inp_b], out_c))
utils.print_names_and_values(
keract.get_gradients_of_activations(m1, [inp_a, inp_b], out_c))
# Just get the last layer!
utils.print_names_and_values(
keract.get_activations(m1, [inp_a, inp_b], layer_names='last_layer'))
utils.print_names_and_values(
keract.get_gradients_of_activations(m1, [inp_a, inp_b],
out_c,
layer_names='last_layer'))
print('')
print('SINGLE-INPUT MODEL')
m2 = get_single_inputs_model()
m2.compile(optimizer='adam', loss='mse')
utils.print_names_and_values(keract.get_activations(m2, inp_a))
utils.print_names_and_values(
keract.get_gradients_of_trainable_weights(m2, inp_a, out_c))
utils.print_names_and_values(
keract.get_gradients_of_activations(m2, inp_a, out_c))
from data import MNIST
if __name__ == '__main__':
# gradients requires no eager execution.
tf.compat.v1.disable_eager_execution()
# Check for GPUs and set them to dynamically grow memory as needed
# Avoids OOM from tensorflow greedily allocating GPU memory
utils.gpu_dynamic_mem_growth()
x_train, y_train, _, _ = MNIST.get_mnist_data()
# (60000, 28, 28, 1) to ((60000, 28, 28)
# LSTM has (batch, time_steps, input_dim)
x_train = x_train.squeeze()
model = Sequential()
model.add(LSTM(16, input_shape=(28, 28)))
model.add(Dense(MNIST.num_classes, activation='softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
utils.print_names_and_shapes(keract.get_activations(model, x_train[:128]))
utils.print_names_and_shapes(
keract.get_gradients_of_trainable_weights(model, x_train[:128],
y_train[:128]))
utils.print_names_and_shapes(
keract.get_gradients_of_activations(model, x_train[:128],
y_train[:128]))
inp_a = np.random.uniform(size=(5, 10))
inp_b = np.random.uniform(size=(5, 10))
out_c = np.random.uniform(size=(5, 1))
# Just for visual purposes.
np.set_printoptions(precision=2)
# Activations of all the layers
print('MULTI-INPUT MODEL')
m1 = get_multi_inputs_model()
m1.compile(optimizer='adam', loss='mse')
utils.print_names_and_values(keract.get_activations(m1, [inp_a, inp_b]))
utils.print_names_and_values(
keract.get_gradients_of_trainable_weights(m1, [inp_a, inp_b], out_c))
utils.print_names_and_values(
keract.get_gradients_of_activations(m1, [inp_a, inp_b], out_c))
# Just get the last layer!
utils.print_names_and_values(
keract.get_activations(m1, [inp_a, inp_b], layer_name='last_layer'))
utils.print_names_and_values(
keract.get_gradients_of_activations(m1, [inp_a, inp_b],
out_c,
layer_name='last_layer'))
print('')
print('SINGLE-INPUT MODEL')
m2 = get_single_inputs_model()
m2.compile(optimizer='adam', loss='mse')
utils.print_names_and_values(keract.get_activations(m2, inp_a))
utils.print_names_and_values(
print(model.summary())
x_train, y_train, x_test, y_test = MNIST.get_mnist_data()
# checking that the accuracy is the same as before 99% at the first epoch.
# test_loss, test_acc = model.evaluate(x_test, y_test, verbose=0, batch_size=128)
# print('')
# assert test_acc > 0.98
utils.print_names_and_shapes(
keract.get_activations(model, x_test[0:200])) # with 200 samples.
utils.print_names_and_shapes(
keract.get_gradients_of_trainable_weights(model, x_train[0:10],
y_train[0:10]))
utils.print_names_and_shapes(
keract.get_gradients_of_activations(model, x_train[0:10],
y_train[0:10]))
a = keract.get_activations(model, x_test[0:1]) # with just one sample.
keract.display_activations(a, directory='mnist_activations', save=True)
# import numpy as np
# import matplotlib.pyplot as plt
# plt.imshow(np.squeeze(x_test[0:1]), interpolation='None', cmap='gray')
else:
x_train, y_train, x_test, y_test = MNIST.get_mnist_data()
model = Sequential()
model.add(
Conv2D(32,
kernel_size=(3, 3),
activation='relu',
def get_gradients(self, x, y=None, layer_name=None):
gradients = keract.get_gradients_of_activations(self.model,
x,
y,
layer_names=layer_name)
return np.squeeze(list(gradients.values())[0])
patience=15,
verbose=1,
restore_best_weights=True)
# model.fit(
# x=x_train,
# y=y_train,
# verbose=1,
# batch_size=batch_size,
# epochs=epochs,
# callbacks=[plateau_callback, es_callback],
# validation_data=(x_test, y_test))
# model.save_weights(filepath=data_model_path)
model.load_weights(filepath=data_model_path)
# score = model.evaluate(
# x_test,
# y_test,
# verbose=0,
# batch_size=batch_size)
# print("Test performance: ", score)
grads = keract.get_gradients_of_activations(model,
x_test[[12]],
y_test[[12]],
layer_name='heatmap1')
keract.display_heatmaps(grads, x_test[12] * 255.0)
activations = keract.get_activations(model, x_test[[12]])
keract.display_activations(activations)
yhat = model.predict(img)
label = decode_predictions(yhat)
label = label[0][0]
print('{} ({})'.format(label[1], label[2] * 100))
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
activations = keract.get_activations(model, img, layer_name='block5_conv3')
first = activations.get('block5_conv3')
# keract.display_activations(activations)
# keract.display_heatmaps(activations, input_image=image)
grad_trainable_weights = keract.get_gradients_of_activations(model, img, yhat, layer_name='block5_conv3')
print(grad_trainable_weights['block5_conv3'].shape)
grad_trainable_weights = tf.convert_to_tensor(grad_trainable_weights['block5_conv3'])
pooled_grads = K.mean(grad_trainable_weights, axis=(0, 1, 2))
# 我们计算相类输出值关于特征图的梯度。然后，我们沿着除了通道维度之外的轴对梯度进行池化操作。最后，我们用计算出的梯度值对输出特征图加权。
heatmap = tf.reduce_mean(tf.multiply(pooled_grads, first[0]), axis=-1)
# 然后，我们沿着通道维度对加权的特征图求均值，从而得到大小为 14*14 的热力图。最后，我们对热力图进行归一化处理，以使其值在 0 和 1 之间。
heatmap = np.maximum(heatmap, 0)
heatmap /= np.max(heatmap)
print(heatmap.shape)
def gradients_of_activations(self, st=0, en=None, indices=None, layer_name=None) -> dict:
test_x, test_y, test_label = self.fetch_data(start=st, ende=en, shuffle=False,
cache_data=False,
indices=indices)
return keract.get_gradients_of_activations(self.k_model, test_x, test_label, layer_names=layer_name)
