def test_time_distributed(self):
input_time = Input(shape=(None, ), name='input_time')
emb = Embedding(10, 4, name='embedding')(input_time)
lstm = LSTM(4, name='lstm')(emb)
output_time = Dense(1, name='output_time')(lstm)
model_time = Model(input_time, output_time)
print(model_time.summary())
# define full model
input = Input(shape=(None, None), name='input')
timesteps = TimeDistributed(model_time, name='timedistributed')(input)
output = Dense(1)(timesteps)
model = Model(input, output)
# create example
x = np.array([[[1, 2, 1]], [[3, 4, 5]]])
# get activations
acts = keract.get_activations(model,
x,
layer_names=['timedistributed'],
nested=True)
self.assertTrue('timedistributed' in acts)
def test_custom_loss(self):
x = np.random.rand(100).reshape(100, 1)
inputs = tf.keras.layers.Input(shape=(1, ))
d = tf.keras.layers.Dense(11)(inputs)
d = tf.keras.layers.Dense(1)(d)
model = tf.keras.Model(inputs=inputs, outputs=d)
def loss(y_true, y_pred):
return tf.reduce_mean((y_true - y_pred)**2)
optimizer = tf.keras.optimizers.Adam()
model.add_loss(loss(inputs, d))
model.compile(optimizer)
keract.get_activations(model, x)
def test_model_in_model(self):
np.random.seed(123)
inputs = np.random.uniform(size=(8, 32, 32, 1))
nested_model = create_network_with_one_subnet()
# will get the activations of every layer, EXCLUDING subnet layers.
acts_not_nested = keract.get_activations(nested_model, inputs)
self.assertTrue('subnet' in acts_not_nested)
self.assertTrue('subnet/conv2d' not in acts_not_nested)
# will get the activations of every layer, including subnet.
acts_nested = keract.get_activations(nested_model, inputs, nested=True)
self.assertTrue('subnet' not in acts_nested)
self.assertTrue('subnet/conv2d' in acts_nested)
self.assertTrue('subnet/max_pooling2d' in acts_nested)
self.assertTrue('subnet/conv2d_1' in acts_nested)
self.assertTrue('subnet/max_pooling2d_1' in acts_nested)
def test_multi_inputs_multi_outputs(self):
inp_a = np.random.uniform(size=(5, 10))
inp_b = np.random.uniform(size=(5, 10))
# out_d = np.random.uniform(size=(5, 1))
# out_e = np.random.uniform(size=(5, 1))
m1 = get_multi_outputs_model()
m1.compile(optimizer='adam', loss='mse')
# m1.fit(x=[inp_a, inp_b], y=[out_d, out_e])
acts = keract.get_activations(m1, [inp_a, inp_b])
self.assertListEqual(list(acts['i1'].shape), [5, 10])
self.assertListEqual(list(acts['i2'].shape), [5, 10])
self.assertListEqual(list(acts['add'].shape), [5, 10])
self.assertListEqual(list(acts['o1'].shape), [5, 1])
self.assertListEqual(list(acts['o2'].shape), [5, 2])
def test_get_activations_from_multi_outputs(self):
# define model
inputs = tf.keras.Input(shape=(None, ), name='input')
emb = tf.keras.layers.Embedding(100, 4, name='embeddding')(inputs)
lstm = tf.keras.layers.LSTM(4,
return_sequences=True,
return_state=True,
name='lstm')
lstm_outputs, state_h, state_c = lstm(emb)
outputs = tf.keras.layers.Dense(1, name='final_dense')(state_h)
model = tf.keras.Model(inputs, outputs)
model.summary()
# create example
x = np.ones(shape=(16, 10))
# get activations
act = keract.get_activations(model, x, layer_names='lstm')['lstm']
self.assertEqual(len(act), 3)
self.assertListEqual(list(act[0].shape), [16, 10, 4])
self.assertListEqual(list(act[1].shape), [16, 4])
self.assertListEqual(list(act[2].shape), [16, 4])
from tensorflow.keras.preprocessing.image import img_to_array
from Models import cifar10_modelfn as cf
# Check for GPUs and set them to dynamically grow memory as needed
# Avoids OOM from tensorflow greedily allocating GPU memory
model = cf.load_weights()
# url = 'https://upload.wikimedia.org/wikipedia/commons/thumb/1/14/Gatto_' \
# 'europeo4.jpg/250px-Gatto_europeo4.jpg'
# response = requests.get(url)
# image = Image.open(BytesIO(response.content))
# image = image.crop((0, 0, 224, 224))
image = Image.open('Visualizations/dog10.png')
image = img_to_array(image)
image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))
image = preprocess_input(image)
# yhat = model.predict(image)
# 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, image)
first = activations.get('block1_conv1')
keract.display_activations(activations, save=True)