def test_dropout(self):
inputs = lib.Input((5, 5))
net = lib.layers.Dropout(0.5)
net.eval()
outputs = net(inputs)
self.assertListEqual(list(lib.engine.int_shape(outputs)), [1, 5, 5])
with self.cached_session() as sess:
sess.run(tf.global_variables_initializer())
print(sess.run(outputs))
def test_rnn(self):
inputs = lib.Input(batch_input_shape=(2, 10, 28))
rnn = lib.layers.RNN(10)
outputs = rnn(inputs)
node = getattr(outputs, '_anchor')[0]
self.assertEqual(node.layer.name, rnn.name)
self.assertListEqual(list(lib.engine.int_shape(outputs)), [2, 10])
# writer = tf.summary.FileWriter('D:/GeekGank/workspace/graph/model_graph', tf.get_default_graph())
# writer.close()
with self.cached_session() as sess:
sess.run(tf.global_variables_initializer())
print(sess.run(outputs))
def test_mobilenet_v2_forward(self):
start = time.time()
inputs = lib.Input((224, 224, 3))
with lib.arg_scope(lib.research.resnet_beta_arg_scope()):
model = lib.research.MobileNetV2(3,
num_classes=1001,
multiplier=0.25)
outputs = model(inputs)
print(time.time() - start)
self.assertListEqual(outputs.get_shape().as_list(), [1, 1001])
writer = tf.summary.FileWriter(
"D:/GeekGank/workspace/graph/model_graph", tf.get_default_graph())
writer.close()
def test_batch_normal(self):
inputs = lib.Input(input_shape=(7, 7, 3), batch_size=2)
outputs = lib.layers.Conv2D(3, 3, padding='VALID')(inputs)
bn = lib.layers.BatchNorm()
# bn.train()
outputs = bn(outputs)
writer = tf.summary.FileWriter('D:/GeekGank/workspace/graph/model_graph', tf.get_default_graph())
writer.close()
print(bn.trainable, True)
self.assertListEqual(list(lib.engine.int_shape(outputs)), [2, 5, 5, 3])
with self.cached_session() as sess:
sess.run(tf.global_variables_initializer())
print(sess.run(outputs))
def test_network_weights(self):
inputs = lib.Input((224, 224, 3), batch_size=2)
with lib.arg_scope(lib.research.resnet_beta_arg_scope()):
net = lib.research.ResNet_V1_50_beta(num_classes=1001,
global_pool=True)
# net = lib.research.MobileNetV2(3, 10)
net.train()
_ = net(inputs)
weights = set()
for w in net.weights:
if w in weights:
print('repeat', w)
else:
weights.add(w)
print(w)
def create_model(input_shape):
with tf.name_scope('model') as scope:
model_inputs = lib.Input(batch_input_shape=input_shape)
net = lib.research.MobileNetV2(3, base_only=True)(model_inputs)
classes = lib.layers.Conv2D(out_channels=10, kernel_size=(1, 13))(net)
pattern = lib.layers.Dense(num_units=24)(lib.layers.flatten(classes))
pattern = lib.layers.reshape(pattern, (-1, 1, 1, 24))
pattern = lib.layers.tile(pattern,
[1, 1, lib.engine.int_shape(net)[2], 1])
net = lib.layers.concat(classes, pattern, axis=3)
net = lib.layers.Conv2D(out_channels=10, kernel_size=1)(net)
model_outputs = lib.layers.squeeze(net, axis=1)
return lib.engine.Network(inputs=model_inputs,
outputs=model_outputs,
name=scope)
def test_graph_model_estimator_style_fit(self):
(train_images,
train_labels), (test_images,
test_labels) = keras.datasets.mnist.load_data()
# x_train = np.random.random((1000, 20))
# y_train = np.random.randint(2, size=(1000, 1))
# x_test = np.random.random((100, 20))
# y_test = np.random.randint(2, size=(100, 1))
train_images = train_images / 255.0
test_images = test_images / 255.0
class MNIST(lib.training.Model):
def __init__(self, *args, **kwargs):
super(MNIST, self).__init__(*args, **kwargs)
self.flatten = lib.layers.Flatten()
self.fc1 = lib.layers.Dense(units=128, activation='relu')
self.fc2 = lib.layers.Dense(units=10)
def forward(self, inputs):
x = self.flatten(inputs)
x = self.fc1(x)
x = self.fc2(x)
return x
def model_fn(_model, x, y):
loss = lib.training.SparseCategoricalCrossEntropy(
from_logits=True)(y, _model.outputs[0])
metric = lib.training.SparseCategoricalAccuracy()(
y, _model.outputs[0])
return lib.training.EstimatorSpec(outputs=_model.outputs,
loss=loss,
metrics=[metric])
inputs = lib.Input(input_shape=(28, 28))
outputs = MNIST()(inputs)
model = lib.training.Model(inputs=inputs,
outputs=outputs,
name='mnist/')
model.train()
model.compile(model_fn=model_fn,
optimizer=tf.train.AdamOptimizer(),
checkpoint_dir='./test_ckpt',
session_cfg=dict(per_process_gpu_memory_fraction=0.4))
model.fit(train_images,
train_labels,
test_images,
test_labels,
epochs=10)
def test_resnet(self):
start = time.time()
inputs = lib.Input((224, 224, 3))
with lib.arg_scope(lib.research.resnet_beta_arg_scope()):
model = lib.research.ResNet_V1_101_beta((224, 224, 3),
is_training=False,
global_pool=True,
num_classes=1001)
outputs = model(inputs)
print(time.time() - start)
node = getattr(outputs, '_anchor')[0]
self.assertEqual(node.layer.name, 'resnet_v1_101/')
self.assertListEqual(outputs.get_shape().as_list(), [1, 1001])
writer = tf.summary.FileWriter(
"D:/GeekGank/workspace/graph/model_graph", tf.get_default_graph())
writer.close()
def test_residual_block(self):
start = time.time()
inputs = lib.Input((54, 54, 64), batch_size=2)
net = lib.research.Bottleneck(128 * 4,
128,
stride=2,
downsample=lib.layers.Conv2D(
128 * 4,
kernel_size=1,
strides=2,
use_bias=False))
outputs = net(inputs)
print(time.time() - start)
print(outputs)
self.assertListEqual(outputs.get_shape().as_list(),
[2, 27, 27, 128 * 4])
writer = tf.summary.FileWriter(
"D:/GeekGank/workspace/graph/model_graph", tf.get_default_graph())
writer.close()
def _ResNetV1Beta(blocks,
num_classes=None,
is_training=None,
global_pool=False,
base_only=False,
root_block=None,
output_stride=None,
extract_blocks=None,
input_shape=None,
input_tensor=None,
name=None):
if extract_blocks is not None:
if input_shape is None:
if input_tensor is None:
raise ValueError("You must provide either `input_shape` or"
" a traceable `input_tensor` to build graph")
with lib.hooks.ExtractHook(extract_blocks, prefix=name) as hook:
network = ResNetV1Beta(blocks=blocks,
num_classes=num_classes,
is_training=is_training,
global_pool=global_pool,
base_only=base_only,
output_stride=output_stride,
name=name)
inputs = lib.Input(input_tensor=input_tensor,
input_shape=input_shape)
_ = network(inputs)
outputs = list(hook.get_endpoints().values())
return lib.Network(inputs=inputs, outputs=outputs, name=name + '/')
else:
return ResNetV1Beta(blocks=blocks,
num_classes=num_classes,
is_training=is_training,
global_pool=global_pool,
base_only=base_only,
root_block=root_block,
output_stride=output_stride,
name=name)
def test_lambda(self):
inputs = [lib.Input((2,)), lib.Input((2,))]
outputs = lib.layers.Lambda(tf.add, name='add')(*inputs)
self.assertListEqual(list(lib.engine.int_shape(outputs)), [1, 2])
with self.cached_session() as sess:
print(sess.run(outputs))
def test_convolution(self):
inputs = lib.Input(input_shape=(7, 7, 3), batch_size=2)
outputs = lib.layers.Conv2D(3, 3, strides=2)(inputs)
print(outputs)
