def test_gradients_of_trainable_weights(self):
model, x = dummy_model_and_inputs()
model.compile(loss='mse', optimizer='adam')
y = np.random.uniform(size=len(x))
grad_trainable_weights = get_gradients_of_trainable_weights(
model, x, y)
self.assertListEqual(list(grad_trainable_weights),
['fc1/kernel:0', 'fc1/bias:0'])
w = grad_trainable_weights['fc1/kernel:0']
b = grad_trainable_weights['fc1/bias:0']
self.assertListEqual(list(w.shape), [10, 1]) # Dense.w
self.assertListEqual(list(b.shape), [
1,
]) # Dense.b
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]))
if __name__ == '__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_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')
loss='categorical_crossentropy',
metrics=['accuracy'])
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(
def gradients_of_weights(self, st=0, en=None, indices=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_trainable_weights(self.k_model, test_x, test_label)