def kernel(self, inputs):
x = inputs[self.KEYS.TENSOR.HITS]
x = flatten(x)
m = identity
models = []
for i in range(3):
models += [Dense(self.config(self.KEYS.CONFIG.NB_UNITS)
[i], info='dense_{}'.format(i)),
ReLU,
DropOut()]
# models.append(DropOut())
models.append(
Dense(self.config(self.KEYS.CONFIG.MAX_NB_HITS), info='dense_end'))
seq = self.graphs.get('seq', Stack(info='stack', models=models))
return seq(x)
def test_base_model_inheritance_ver1(clean_config):
x = tf.ones([32, 64, 64, 3], dtype=tf.float32)
m1 = Conv2D('conv', 32, 3)
m2 = Dense('dense', 160)
m3 = Stack([m1, m2])
res1 = m3(x)
assert shape(res1) == [32, 64, 64, 160]
def kernel(self, inputs):
x = inputs[self.KEYS.TENSOR.HITS]
x = flatten(x)
m = identity
models = []
for i in range(len(self.config(self.KEYS.CONFIG.NB_UNITS))):
models += [Dense(self.config(self.KEYS.CONFIG.NB_UNITS)
[i], info='dense_{}'.format(i)),
ReLU,
DropOut()]
models.append(
Dense(self.config(self.KEYS.CONFIG.MAX_NB_HITS), info='dense_end'))
if self.graphs.get(self.KEYS.GRAPH.SEQUENTIAL) is None:
self.graphs[self.KEYS.GRAPH.SEQUENTIAL] = Sequential(
info='stack', models=models)
return self.graphs[self.KEYS.GRAPH.SEQUENTIAL](x)
def test_base_model_inheritance_ver3(clean_config):
x = tf.ones([32, 64, 64, 160], dtype=tf.float32)
m1 = Conv2D('conv1', 32, 3)
m7 = Conv2D('conv2', 3, 3)
m8 = Dense('dense', 128)
m6 = Inception('inception', m7, m7, [m1, m8])
res3 = m6(x)
assert shape(res3) == [32, 64, 64, 3]
def test_base_model_inheritance_ver2(clean_config):
x = tf.ones([32, 64, 64, 3], dtype=tf.float32)
m4 = Conv2D('conv', 160, 3)
m9 = Dense('dense', 160)
x = m9(x)
m5 = Residual('res', m4)
res2 = m5(x)
assert shape(res2) == [32, 64, 64, 160]
def test_Inception_basic(clean_config):
x = tf.constant(np.ones([32, 64, 64, 3], np.float32))
m1 = Conv2D('conv3', 64, 3)
m2 = Dense('dense1', 128)
paths = [Conv2D('conv_{}'.format(i), 32 + 32 * i, 3) for i in range(3)]
m = Inception('inception1', m1, m2, paths)
y = m(x)
assert shape(y) == [32, 64, 64, 128]
def test_merge_basic(clean_config):
m1 = Dense('d1', 256)
m2 = Conv2D('conv1', 64, 3)
m3 = Conv2D('conv2', 128, 3)
concate = partial(concat, axis=3)
m = Merge(merger=concate, models=[m2, m3])
x = tf.ones([32, 64, 64, 3], dtype=tf.float32)
res = m(x)
res = m1(res)
assert shape(res) == [32, 64, 64, 256]
def test_merge_parameter(clean_config):
m1 = Dense('d1', 256)
m2 = Conv2D('conv1', 64, 3)
m3 = Conv2D('conv2', 128, 3)
concate = partial(concat, axis=3)
m = Merge(merger=concate, models=[m2, m3])
x = tf.ones([32, 64, 64, 3], dtype=tf.float32)
res = m(x)
res = m1(res)
assert m.parameters[0].get_shape() == (3, 3, 3, 64)
def test_Inception_parameter(clean_config):
x = tf.constant(np.ones([32, 64, 64, 3], np.float32))
m1 = Conv2D('conv4', 64, 3)
m2 = Dense('dense2', 128)
paths = [Conv2D('conv_{}'.format(i), 32 + 32 * i, 3) for i in range(3)]
m = Inception('inception2', m1, m2, paths)
y = m(x)
assert m.parameters[0].get_shape() == (3, 3, 3, 64)
assert m.parameters[1].get_shape() == (64, )
assert m.parameters[2].get_shape() == (3, 3, 64, 32)
assert m.parameters[3].get_shape() == (32, )