def test_printer(self, b, w, h, c):
layer = Input_layer(input_shape=(b, w, h, c))
print(layer)
layer.input_shape = (3.14, w, h, c)
with pytest.raises(ValueError):
print(layer)
def __init__(self, batch, input_shape=None, train=None):
'''
'''
self.batch = batch
self.train = train
if input_shape is not None:
try:
self.w, self.h, self.c = input_shape
except:
raise ValueError(
'Network model : incorrect input_shape. Expected a 3D array (width, height, channel). Given {}'
.format(input_shape))
self._net = [
Input_layer(input_shape=(self.batch, self.w, self.h, self.c))
]
else:
self._net = []
self.metrics = None
self._fitted = False
def test_constructor(self, b, w, h, c):
input_shape = choice([(b, w, h, c), (b, w, h), b, None])
if input_shape != (b, w, h, c):
with pytest.raises(ValueError):
layer = Input_layer(input_shape=input_shape)
else:
layer = Input_layer(input_shape=input_shape)
layer.input_shape == (b, w, h, c)
assert layer.output == None
assert layer.delta == None
assert layer.out_shape == (b, w, h, c)
def test_backward(self, b, w, h, c):
inpt = np.random.uniform(low=-1, high=1.,
size=(b, w, h, c)).astype(float)
tf_input = tf.Variable(inpt)
# numpynet model init
layer = Input_layer(input_shape=inpt.shape)
# Keras Model init
model = tf.keras.layers.InputLayer(input_shape=(w, h, c))
# FORWARD
# Tensorflow Forward and backward
with tf.GradientTape() as tape:
preds = model(tf_input)
grads = tape.gradient(preds, tf_input)
forward_out_keras = preds.numpy()
delta_keras = grads.numpy()
# layer forward
layer.forward(inpt=inpt)
forward_out_numpynet = layer.output
# Forward check (Shape and Values)
assert forward_out_keras.shape == forward_out_numpynet.shape
np.testing.assert_allclose(forward_out_keras, forward_out_numpynet)
# BACKWARD
# layer delta init.
layer.delta = np.ones(shape=inpt.shape, dtype=float)
# Global delta init.
delta = np.empty(shape=inpt.shape, dtype=float)
# layer Backward
layer.backward(delta=delta)
# Check dimension and delta
assert delta_keras.shape == delta.shape
np.testing.assert_allclose(delta_keras, delta)
delta = np.zeros(shape=(1, 2, 3, 4), dtype=float)
with pytest.raises(ValueError):
layer.backward(delta)
def load(self, cfg_filename, weights=None):
'''
Load network model from config file in INI fmt
'''
model = net_config(cfg_filename)
self.batch = model.get('net1', 'batch', 1)
self.w = model.get('net1', 'width', 416)
self.h = model.get('net1', 'height', 416)
self.c = model.get('net1', 'channels', 3)
# TODO: add other network parameters
input_shape = (self.batch, self.w, self.h, self.c)
self._net = [ Input_layer(input_shape=input_shape) ]
print('layer filters size input output')
for i, layer in enumerate(model):
layer_t = re.split(r'\d+', layer)[0]
params = dict(model.get_params(layer))
layer_params = {}
for k, v in params.items():
try:
val = eval(v)
except NameError:
val = v
except:
raise DataVariableError('Type variable not recognized! Possible variables are only [int, float, string, vector<float>].')
layer_params[k] = val
if layer_t == 'shortcut':
_from = model.get(layer, 'from', 0)
self._net.append( self.LAYERS[layer_t](input_shape=input_shape, **layer_params)([self._net[-1], self._net[_from]]) )
elif layer_t == 'route':
_layers = model.get(layer, 'layers', [])
self._net.append( self.LAYERS[layer_t](input_shape=input_shape, **layer_params)(self._net[_layers]) )
else:
self._net.append( self.LAYERS[layer_t](input_shape=input_shape, **layer_params)(self._net[-1]) )
print('{:>4d} {}'.format(i, self._net[-1]), flush=True, end='\n')
#if model.get(layer, 'batch_normalize', 0): # wrong because it add a new layer and so the shortcut is broken
# self._net.append( BatchNorm_layer()(self._net[-1]) )
# print('{:>4d} {}'.format(i, self._net[-1]), flush=True, end='\n')
return self
if weights is not None:
self.load_weights(weights)
def load(self, cfg_filename, weights=None):
'''
Load network model from config file in INI fmt
'''
model = net_config(cfg_filename)
self.batch = model.get('net0', 'batch', 1)
self.w = model.get('net0', 'width', 416)
self.h = model.get('net0', 'height', 416)
self.c = model.get('net0', 'channels', 3)
# TODO: add other network parameters
input_shape = (self.batch, self.w, self.h, self.c)
self._net = [Input_layer(input_shape=input_shape)]
print(
'layer filters size input output'
)
for i, layer in enumerate(model):
layer_t = re.split(r'\d+', layer)[0]
params = model.get_params(layer)
if layer_t == 'shortcut':
_from = model.get(layer, 'from', 0)
self._net.append(self.LAYERS[layer_t](
input_shape=input_shape,
**params)([self._net[-1], self._net[_from]]))
elif layer_t == 'route':
_layers = model.get(layer, 'layers', [])
self._net.append(self.LAYERS[layer_t](input_shape=input_shape,
**params)(
self._net[_layers]))
else:
self._net.append(self.LAYERS[layer_t](input_shape=input_shape,
**params)(self._net[-1]))
input_shape = self._net[-1].out_shape
print('{:>4d} {}'.format(i, self._net[-1]), end='\n') # flush=True
sys.stdout.flush() # compatibility with pythonn 2.7
# if model.get(layer, 'batch_normalize', 0): # wrong because it add a new layer and so the shortcut is broken
# self._net.append( BatchNorm_layer()(self._net[-1]) )
# print('{:>4d} {}'.format(i, self._net[-1]), flush=True, end='\n')
if weights is not None:
self.load_weights(weights)
return self
def test_input_layer(batch, w, h, c):
'''
Tests:
if the forward and the backward of Numpy_net are consistent with keras.
to be:
'''
inpt = np.random.uniform(low=-1, high=1., size=(batch, w, h, c))
# numpynet model init
numpynet = Input_layer(input_shape=inpt.shape)
# Keras Model init
inp = Input(shape=(w, h, c), batch_shape=(batch, w, h, c))
x = Activation(activation='linear')(inp)
model = Model(inputs=[inp], outputs=x)
# FORWARD
# Keras Forward
forward_out_keras = model.predict(inpt)
# numpynet forwrd
numpynet.forward(inpt)
forward_out_numpynet = numpynet.output
# Forward check (Shape and Values)
assert forward_out_keras.shape == forward_out_numpynet.shape
assert np.allclose(forward_out_keras, forward_out_numpynet)
# BACKWARD
# Gradient computation (Analytical)
grad = K.gradients(model.output, [model.input])
# Define a function to compute the gradient numerically
func = K.function(model.inputs + [model.output], grad)
# Keras delta
keras_delta = func([inpt])[0] # It returns a list with one array inside.
# numpynet delta init.
numpynet.delta = np.ones(shape=inpt.shape, dtype=float)
# Global delta init.
delta = np.empty(shape=inpt.shape, dtype=float)
# numpynet Backward
numpynet.backward(delta)
# Check dimension and delta
assert keras_delta.shape == delta.shape
assert np.allclose(keras_delta, delta)
def test_forward(self, b, w, h, c):
inpt = np.random.uniform(low=-1, high=1.,
size=(b, w, h, c)).astype(float)
# numpynet model init
layer = Input_layer(input_shape=inpt.shape)
# Keras Model init
model = tf.keras.layers.InputLayer(input_shape=(w, h, c))
# FORWARD
# Keras Forward
forward_out_keras = model(inpt)
# numpynet forwrd
layer.forward(inpt=inpt)
forward_out_numpynet = layer.output
# Forward check (Shape and Values)
assert forward_out_keras.shape == forward_out_numpynet.shape
np.testing.assert_allclose(forward_out_keras, forward_out_numpynet)