def test_raise_conv_multi_src(self):
with self.assertRaises(DLPyError):
inputs = Input(3, 512, 512, scale=1.0 / 255, name='input1')
conv1 = Conv2d(8, 3, act='relu')(inputs)
conv2 = Conv2d(8, 3, act='relu')(inputs)
conv3 = Conv2d(8, 3)([conv1, conv2])
output1 = OutputLayer(name='output1')(conv3)
model = Model(self.s, inputs=inputs, outputs=output1)
model.compile()
def test_multi_src_layer_fc(self):
inputs = Input(1, 28, 28, scale=1.0 / 255, name='InputLayer_1')
fc1 = Dense(n=128, name='dense1')(inputs)
fc2 = Dense(n=64, name='dense2')(fc1)
fc3 = Dense(n=64, name='dense3')([fc1, fc2])
output1 = OutputLayer(n=10, name='OutputLayer_1')(fc3)
model = Model(self.s, inputs=inputs, outputs=output1)
model.compile()
self.assertTrue(model.count_params() == 117386)
def test_submodel_as_inputs_network(self):
inputs1 = Input(1, 53, 53, scale=1.0 / 255, name='InputLayer_1')
# inputs2 = InputLayer(1, 28, 28, scale = 1.0 / 255, name = 'InputLayer_2')
dense1 = Dense(10, name='d1')(inputs1)
dense2 = Dense(12, name='d2')(dense1)
dense3 = Dense(n=128, name='d3')(dense2)
model_dense = Model(self.s, inputs=inputs1, outputs=dense3)
inputs2 = Input(1, 53, 53, scale=1.0 / 255, name='InputLayer_2')
# inputs2 = InputLayer(1, 28, 28, scale = 1.0 / 255, name = 'InputLayer_2')
dense4 = Dense(10, name='d4')(inputs2)
dense5 = Dense(12, name='d5')(dense4)
dense6 = Dense(n=128, name='d6')(dense5)
model_dense2 = Model(self.s, inputs=inputs2, outputs=dense6)
inputs3 = Input(1, 53, 53, scale=1.0 / 255, name='InputLayer_3')
inputs4 = Input(1, 53, 53, scale=1.0 / 255, name='InputLayer_4')
out3 = model_dense(inputs3)
out4 = model_dense2(inputs4)
out5 = [out4[0], out3[0]]
concat1 = Dense(20, name='d7')(out5)
output1 = OutputLayer(n=10, name='OutputLayer_1')(concat1)
model = Model(self.s, inputs=[inputs3, inputs4], outputs=output1)
model.compile()
model.print_summary()
self.assertTrue(model.count_params() == 62262)
def test_load_reshape_detection(self):
if self.data_dir is None:
unittest.TestCase.skipTest(self, "DLPY_DATA_DIR is not set in the environment variables")
yolo_model = Model(self.s)
yolo_model.load(self.data_dir + 'YOLOV2_MULTISIZE.sashdat')
model_df = self.s.fetch(table = dict(name = yolo_model.model_name,
where = '_DLKey0_ eq "detection1" or _DLKey0_ eq "reshape1"'), to = 50).Fetch
anchors_5 = model_df['_DLNumVal_'][model_df['_DLKey1_'] == 'detectionopts.anchors.8'].tolist()[0]
self.assertAlmostEqual(anchors_5, 1.0907, 4)
depth = model_df['_DLNumVal_'][model_df['_DLKey1_'] == 'reshapeopts.depth'].tolist()[0]
self.assertEqual(depth, 256)
def test_submodel_multiple_inputs_network(self):
inputs1 = Input(1, 53, 53, scale=1.0 / 255, name='InputLayer_1')
# inputs2 = InputLayer(1, 28, 28, scale = 1.0 / 255, name = 'InputLayer_2')
dense1 = Dense(10, name='d1')(inputs1)
dense2 = Dense(12, name='d2')(dense1)
dense3 = Dense(n=128, name='d3')(dense2)
# model_dense = Model(self.s, inputs=inputs1, outputs= dense3)
inputs2 = Input(1, 53, 53, scale=1.0 / 255, name='InputLayer_2')
# inputs2 = InputLayer(1, 28, 28, scale = 1.0 / 255, name = 'InputLayer_2')
dense4 = Dense(10, name='d4')(inputs2)
dense5 = Dense(12, name='d5')(dense4)
dense6 = Dense(n=128, name='d6')(dense5)
dense7 = Dense(n=190, name='cat1')([dense3, dense6])
model_dense = Model(self.s,
inputs=[inputs1, inputs2],
outputs=[dense7])
input3 = Input(1, 53, 53, scale=1.0 / 255, name='InputLayer_3')
input4 = Input(1, 53, 53, scale=1.0 / 255, name='InputLayer_4')
out = model_dense([input3, input4])
dense7 = Dense(20, name='d7')(out)
output1 = OutputLayer(n=10, name='OutputLayer_1')(dense7)
model = Model(self.s, inputs=[input3, input4], outputs=output1)
model.compile()
model.print_summary()
self.assertTrue(model.count_params() == 87822)
def test_extract_clustering1(self):
model = Sequential(self.s, model_table = 'Simple_CNN')
model.add(InputLayer(3, 48, 96, scale = 1.0 / 255, random_mutation = 'none'))
model.add(Split(16))
model.add(Clustering(n_clusters=10))
model_extracted = Model.from_table(self.s.CASTable(model.model_table['name']))
model_extracted.compile()
def test_mobilenetv2(self):
try:
import onnx
from dlpy.model_conversion.onnx_transforms import (Transformer, OpTypePattern,
ConstToInitializer,
InitReshape, InitUnsqueeze,
FuseMulAddBN)
from dlpy.model_conversion.onnx_graph import OnnxGraph
from onnx import helper, numpy_helper
except:
unittest.TestCase.skipTest(self, 'onnx package not found')
from dlpy.model import Model
if self.data_dir_local is None:
unittest.TestCase.skipTest(self, "DLPY_DATA_DIR_LOCAL is not set in the environment variables")
path = os.path.join(self.data_dir_local, 'mobilenetv2-1.0.onnx')
onnx_model = onnx.load_model(path)
model1 = Model.from_onnx_model(self.s,
onnx_model,
output_model_table='mobilenetv2',
offsets=255*[0.485, 0.456, 0.406],
norm_stds=255*[0.229, 0.224, 0.225])
def test_model29(self):
# test specifying output layer in Model.from_onnx_model
try:
import onnx
except:
unittest.TestCase.skipTest(self, "onnx not found in the libraries")
if self.data_dir_local is None:
unittest.TestCase.skipTest(
self, "DLPY_DATA_DIR_LOCAL is not set in "
"the environment variables")
m = onnx.load(os.path.join(self.data_dir_local, 'Simple_CNN1.onnx'))
output_layer = OutputLayer(n=100)
model1 = Model.from_onnx_model(conn=self.s,
onnx_model=m,
offsets=[
1,
1,
1,
],
scale=2,
std='std',
output_layer=output_layer)
self.assertTrue(model1.layers[-1].config['n'] == 100)
def test_extract_embeddingloss1(self):
model = Sequential(self.s, model_table = 'Simple_CNN')
model.add(InputLayer(3, 48, 96, scale = 1.0 / 255, random_mutation = 'none'))
model.add(Dense(16))
model.add(EmbeddingLoss(margin=10))
model_extracted = Model.from_table(self.s.CASTable(model.model_table['name']))
model_extracted.compile()
self.assertTrue(model_extracted.layers[-1].config['margin'] == 10)
def load_acoustic_model(self, acoustic_model_path):
"""
Load the RNN acoustic model.
Parameters
----------
acoustic_model_path : string
Specifies the absolute server-side path of the acoustic model file.
Please make sure the weights file and the weights attribute file are placed under the same directory.
"""
self.acoustic_model = Model(self.conn)
self.acoustic_model.from_sashdat(self.conn, path=acoustic_model_path)
if self.acoustic_model.model_table is None:
raise DLPyError("Failed to load the acoustic model.")
if self.acoustic_model.model_weights is None:
raise DLPyError("Failed to load the acoustic model weights.")
def test_resnet(self):
def conv_block(x, filters, size, stride = 1, mode = 'same', act = True):
x = Conv2d(filters, size, size, act = 'identity', include_bias = False, stride = stride)(x)
x = BN(act = 'relu' if act else 'identity')(x)
return x
def res_block(ip, nf = 64):
x = conv_block(ip, nf, 3, 1)
x = conv_block(x, nf, 3, 1, act = False)
return Res()([x, ip])
inp = Input(1, 32, 32, scale = 1.0 / 255, name = 'InputLayer_1')
x = conv_block(inp, 64, 9, 1)
for i in range(1): x = res_block(x)
x = Conv2d(1, 9, 9, act = 'tanh')(x)
output = OutputLayer(n = 100)(x)
model = Model(self.s, inputs = inp, outputs = output)
model.compile()
def test_extract_segmentation1(self):
model = Sequential(self.s, model_table = 'Simple_CNN')
model.add(InputLayer(3, 48, 96, scale = 1.0 / 255, random_mutation = 'none'))
model.add(Dense(16))
model.add(Segmentation(act='AUTO', target_scale=10))
model_extracted = Model.from_table(self.s.CASTable(model.model_table['name']))
model_extracted.compile()
self.assertTrue(model_extracted.layers[-1].config['act'] == 'AUTO')
self.assertTrue(model_extracted.layers[-1].config['target_scale'] == 10)
def test_without_variable(self):
input1 = Input(n_channels=1, width=28, height=28)
conv1 = Conv2d(2)(Conv2d(2)(input1))
output1 = OutputLayer(n=2)(conv1)
model1 = Model(conn=self.s, inputs=[input1], outputs=[output1])
model1.compile()
model1.print_summary()
self.assertTrue(model1.count_params() == 3196)
def __init__(self, conn, layers=None, model_table=None):
Model.__init__(self, conn, model_table=model_table)
if layers is None:
self.layers = []
self.layers_dict = {}
elif type(layers) is list or type(layers) is set or type(
layers) is tuple:
self.layers = layers
for layer in self.layers:
if layer.name is not None:
self.layers_dict[layer.name] = layer
if len(layers) > 0 and isinstance(
layers[-1], Layer) and layers[-1].can_be_last_layer:
self.compile()
else:
raise DLPyError('layers has to be a list of layer(s).')
else:
raise DLPyError('layers has to be a list of layer(s).')
def test_non_list_src_layers(self):
inputs = Input(1, 28, 28, scale=1.0 / 255, name='InputLayer_1')
fc1 = Dense(n=128)(inputs)
fc2 = Dense(n=64)(fc1)
output1 = OutputLayer(n=10, name='OutputLayer_1')([fc1, fc2])
model = Model(self.s, inputs=inputs, outputs=output1)
model.compile()
model.print_summary()
self.assertTrue(model.count_params() == 109834)
def test_imagescaler1(self):
# test import model with imagescaler
try:
import onnx
from onnx import helper, TensorProto
except:
unittest.TestCase.skipTest(self, 'onnx not found')
if self.data_dir_local is None:
unittest.TestCase.skipTest(
self, 'DLPY_DATA_DIR_LOCAL is not set in '
'the environment variables')
import numpy as np
n1 = helper.make_node('ImageScaler', ['X'], ['X1'],
bias=[0., 0., 0.],
scale=1.)
n2 = helper.make_node('Conv',
inputs=['X1', 'W1'],
outputs=['X2'],
kernel_shape=[3, 3],
pads=[0, 0, 0, 0])
n3 = helper.make_node('MatMul', inputs=['X2', 'W2'], outputs=['X3'])
W1 = np.ones((3, 3, 3)).astype(np.float32)
W2 = np.ones((9, 2)).astype(np.float32)
graph_def = helper.make_graph(
[n1, n2, n3],
name='test',
inputs=[
helper.make_tensor_value_info('X', TensorProto.FLOAT,
[1, 3, 10, 10]),
helper.make_tensor_value_info('W1', TensorProto.FLOAT,
[3, 3, 3]),
helper.make_tensor_value_info('W2', TensorProto.FLOAT, [9, 2])
],
outputs=[
helper.make_tensor_value_info('X3', TensorProto.FLOAT, [1, 2])
],
initializer=[
helper.make_tensor('W1', TensorProto.FLOAT, [3, 3, 3],
W1.flatten().astype(np.float32)),
helper.make_tensor('W2', TensorProto.FLOAT, [9, 2],
W2.flatten().astype(np.float32))
])
onnx_model = helper.make_model(graph_def)
model1 = Model.from_onnx_model(self.s, onnx_model)
l1 = model1.layers[0]
self.assertTrue(l1.type == 'input')
self.assertTrue(l1.config['offsets'] == [0., 0., 0.])
self.assertTrue(l1.config['scale'] == 1.)
def test_model6(self):
try:
import onnx
except:
unittest.TestCase.skipTest(self, "onnx not found in the libraries")
m = onnx.load(
os.path.join(os.path.dirname(__file__), 'datasources',
'model.onnx'))
model1 = Model.from_onnx_model(self.s, m)
model1.print_summary()
def test_outputs(self):
with self.assertRaises(DLPyError):
input1 = InputLayer(n_channels=1, width=28, height=28)
conv1 = Conv2d(2)(input1)
model1 = Model(conn=self.s, inputs=[input1], outputs=[conv1])
model1.compile()
with self.assertRaises(DLPyError):
input1 = InputLayer(n_channels=1, width=28, height=28)
conv1 = Conv2d(2)(input1)
conv2 = BN()(input1)
output1 = OutputLayer(2)(conv1)
model1 = Model(conn=self.s, inputs=[input1], outputs=[conv2, output1])
model1.compile()
def test_model28(self):
try:
import onnx
except:
unittest.TestCase.skipTest(self, "onnx not found in the libraries")
if self.data_dir_local is None:
unittest.TestCase.skipTest(self, "DLPY_DATA_DIR_LOCAL is not set in "
"the environment variables")
m = onnx.load(os.path.join(self.data_dir_local, 'pytorch_net2.onnx'))
model1 = Model.from_onnx_model(self.s, m, offsets=[1, 1, 1,], scale=2, std='std')
model1.print_summary()
def test_concat(self):
input1 = Input(n_channels=1, width=28, height=28)
input2 = Input(n_channels=3, width=28, height=28)
conv1 = Conv2d(2)(input1)
conv2 = Conv2d(2)(input1)
conv3 = Conv2d(2)(input2)
output2 = OutputLayer(n=2)(conv3)
concat1 = Concat()([conv1, conv2, conv3])
output1 = OutputLayer(n=2)(concat1)
model1 = Model(conn=self.s, inputs=[input1, input2], outputs=[output1, output2])
model1.compile()
model1.print_summary()
self.assertTrue(model1.count_params() == 12644)
def test_submodel_in_middle_network(self):
inputs1 = Input(1, 53, 53, scale=1.0 / 255, name='InputLayer_1')
inputs2 = Input(1, 28, 28, scale=1.0 / 255, name='InputLayer_2')
dense1 = Dense(10, name='d1')(inputs2)
dense2 = Dense(12, name='d2')(dense1)
dense3 = Dense(n=128, name='d3')(dense2)
model_dense = Model(self.s, inputs=inputs2, outputs=dense3)
fore = model_dense(inputs1)
concat1 = Dense(20, name='d7')(fore)
output1 = OutputLayer(n=10, name='OutputLayer_1')(concat1)
model = Model(self.s, inputs=[inputs1], outputs=output1)
model.compile()
model.print_summary()
self.assertTrue(model.count_params() == 32516)
def test_submodel_as_input_network(self):
inputs1 = Input(1, 53, 53, scale=1.0 / 255, name='InputLayer_1')
# inputs2 = InputLayer(1, 28, 28, scale = 1.0 / 255, name = 'InputLayer_2')
dense1 = Dense(10, name='dense1')(inputs1)
dense2 = Dense(12, name='dense2')(dense1)
dense3 = Dense(n=128, name='dense3')(dense2)
model_dense = Model(self.s, inputs=inputs1, outputs=dense3)
inputs2 = Input(1, 53, 53, scale=1.0 / 255, name='InputLayer_2')
out_submodel = model_dense(inputs2)
concat1 = Dense(20, 'concat1')(out_submodel)
output1 = OutputLayer(n=10, name='OutputLayer_1')(concat1)
model = Model(self.s, inputs=[inputs2], outputs=output1)
model.compile()
model.print_summary()
self.assertTrue(model.count_params() == 32516)
def test_given_name(self):
inputs = Input(3, 512, 512, scale=1.0 / 255, name='input1')
conv1 = Conv2d(8, 3, act='relu')(inputs)
conv2 = Conv2d(8, 3, act='relu')(inputs)
merge3 = Concat()([conv1, conv2])
conv3 = Conv2d(3, 3, act='relu')(merge3)
output1 = OutputLayer(name='output1')(conv3)
model = Model(self.s, inputs=inputs, outputs=output1)
model.compile()
self.assertTrue(inputs._op.name == 'input1')
self.assertTrue(output1._op.name == 'output1')
model.print_summary()
def test_residual_output_shape0(self):
inLayer = Input(n_channels = 3, width = 32, height = 128,
name = 'input1', random_mutation = 'random', random_flip = 'HV')
conv1 = Conv2d(32, 3, 3, name = 'conv1', act = 'relu', init = 'msra')([inLayer])
conv2 = Conv2d(32, 5, 5, name = 'conv2', act = 'relu', init = 'msra')([inLayer])
fc1 = Dense(3, name = 'fc1')([conv1])
fc2 = Dense(3, name = 'fc2')([conv2])
res = Res(name = 'res1')([fc1, fc2])
outLayer = OutputLayer(n = 3, name = 'output')([res])
model = Model(self.s, inLayer, outLayer)
model.compile()
self.assertEqual(model.summary['Output Size'].values[-2], 3)
model.print_summary()
def test_model10(self):
try:
import onnx
except:
unittest.TestCase.skipTest(self, "onnx not found in the libraries")
m = onnx.load(
os.path.join(os.path.dirname(__file__), 'datasources',
'pytorch_net2.onnx'))
model1 = Model.from_onnx_model(self.s,
m,
offsets=[
1,
1,
1,
],
scale=2,
std='std')
model1.print_summary()
def test_stop_layers(self):
from dlpy.sequential import Sequential
model1 = Sequential(self.s)
model1.add(InputLayer(3, 224, 224))
model1.add(Conv2d(8, 7))
model1.add(Pooling(2))
model1.add(Conv2d(8, 7))
model1.add(Pooling(2))
model1.add(Dense(16))
model1.add(OutputLayer(act = 'softmax', n = 2))
inputlayer = model1.layers[0]
inp = Input(**inputlayer.config)
func_model = model1.to_functional_model(stop_layers = [model1.layers[-1]])
x = func_model(inp)
out = Keypoints(n=10)(x)
func_model_keypoints = Model(self.s, inp, out)
func_model_keypoints.compile()
func_model_keypoints.print_summary()
def MobileNetV2_ONNX(conn,
model_file,
n_classes=1000,
width=224,
height=224,
offsets=(255 * 0.406, 255 * 0.456, 255 * 0.485),
norm_stds=(255 * 0.225, 255 * 0.224, 255 * 0.229),
random_flip=None,
random_crop=None,
random_mutation=None,
include_top=False):
"""
Generates a deep learning model with the MobileNetV2_ONNX architecture.
The model architecture and pre-trained weights is generated from MobileNetV2 ONNX trained on ImageNet dataset.
The model file and the weights file can be downloaded from https://support.sas.com/documentation/prod-p/vdmml/zip/.
To learn more information about the model and pre-processing.
Please go to the websites: https://github.com/onnx/models/tree/master/vision/classification/mobilenet.
Parameters
----------
conn : CAS
Specifies the CAS connection object.
model_file : string
Specifies the absolute server-side path of the model table file.
The model table file can be downloaded from https://support.sas.com/documentation/prod-p/vdmml/zip/.
n_classes : int, optional
Specifies the number of classes.
Default: 1000
width : int, optional
Specifies the width of the input layer.
Default: 224
height : int, optional
Specifies the height of the input layer.
Default: 224
offsets : double or iter-of-doubles, optional
Specifies an offset for each channel in the input data. The final input
data is set after applying scaling and subtracting the specified offsets.
The channel order is BGR.
Default: (255*0.406, 255*0.456, 255*0.485)
norm_stds : double or iter-of-doubles, optional
Specifies a standard deviation for each channel in the input data.
The final input data is normalized with specified means and standard deviations.
The channel order is BGR.
Default: (255*0.225, 255*0.224, 255*0.229)
random_flip : string, optional
Specifies how to flip the data in the input layer when image data is
used. Approximately half of the input data is subject to flipping.
Valid Values: 'h', 'hv', 'v', 'none'
random_crop : string, optional
Specifies how to crop the data in the input layer when image data is
used. Images are cropped to the values that are specified in the width
and height parameters. Only the images with one or both dimensions
that are larger than those sizes are cropped.
Valid Values: 'none', 'unique', 'randomresized', 'resizethencrop'
random_mutation : string, optional
Specifies how to apply data augmentations/mutations to the data in the input layer.
Valid Values: 'none', 'random'
include_top : bool, optional
Specifies whether to include pre-trained weights of the top layers (i.e., the FC layers)
Default: False
"""
parameters = locals()
input_parameters = get_layer_options(input_layer_options, parameters)
# load model and model weights
model = Model.from_sashdat(conn, path=model_file)
# check if a user points to a correct model.
if model.summary.shape[0] != 120:
raise DLPyError(
"The model file doesn't point to a valid MobileNetV2_ONNX model. "
"Please check the SASHDAT file.")
# extract input layer config
model_table_df = conn.CASTable(**model.model_table).to_frame()
input_layer_df = model_table_df[model_table_df['_DLLayerID_'] == 0]
input_layer = extract_input_layer(input_layer_df)
input_layer_config = input_layer.config
# update input layer config
input_layer_config.update(input_parameters)
# update the layer list
model.layers[0] = InputLayer(**input_layer_config,
name=model.layers[0].name)
# warning if model weights doesn't exist
if not conn.tableexists(model.model_weights.name).exists:
weights_file_path = os.path.join(os.path.dirname(model_file),
model.model_name + '_weights.sashdat')
print('WARNING: Model weights is not attached '
'since system cannot find a weights file located at {}'.format(
weights_file_path))
if include_top:
if n_classes != 1000:
raise DLPyError(
"If include_top is enabled, n_classes has to be 1000.")
else:
# since the output layer is non fully connected layer,
# we need to modify the convolution right before the output. The number of filter is set to n_classes.
conv_layer_df = model_table_df[model_table_df['_DLLayerID_'] == 118]
conv_layer = extract_conv_layer(conv_layer_df)
conv_layer_config = conv_layer.config
# update input layer config
conv_layer_config.update({'n_filters': n_classes})
# update the layer list
model.layers[-2] = Conv2d(**conv_layer_config,
name=model.layers[-2].name,
src_layers=model.layers[-3])
# overwrite n_classes in output layer
out_layer_df = model_table_df[model_table_df['_DLLayerID_'] == 119]
out_layer = extract_output_layer(out_layer_df)
out_layer_config = out_layer.config
# update input layer config
out_layer_config.update({'n': n_classes})
# update the layer list
model.layers[-1] = OutputLayer(**out_layer_config,
name=model.layers[-1].name,
src_layers=model.layers[-2])
# remove top weights
model.model_weights.append_where('_LayerID_<118')
model._retrieve_('table.partition',
table=model.model_weights,
casout=dict(replace=True,
name=model.model_weights.name))
model.set_weights(model.model_weights.name)
# recompile the whole network according to the new layer list
model.compile()
return model
def MobileNetV2(conn,
model_table='MobileNetV2',
n_classes=1000,
n_channels=3,
width=224,
height=224,
norm_stds=(255 * 0.229, 255 * 0.224, 255 * 0.225),
offsets=(255 * 0.485, 255 * 0.456, 255 * 0.406),
random_flip=None,
random_crop=None,
random_mutation=None,
alpha=1):
'''
Generates a deep learning model with the MobileNetV2 architecture.
The implementation is revised based on
https://github.com/keras-team/keras-applications/blob/master/keras_applications/mobilenet_v2.py
Parameters
----------
conn : CAS
Specifies the CAS connection object.
model_table : string or dict or CAS table, optional
Specifies the CAS table to store the deep learning model.
n_classes : int, optional
Specifies the number of classes. If None is assigned, the model will
automatically detect the number of classes based on the training set.
Default: 1000
n_channels : int, optional
Specifies the number of the channels (i.e., depth) of the input layer.
Default: 3
width : int, optional
Specifies the width of the input layer.
Default: 224
height : int, optional
Specifies the height of the input layer.
Default: 224
norm_stds : double or iter-of-doubles, optional
Specifies a standard deviation for each channel in the input data.
The final input data is normalized with specified means and standard deviations.
Default: (255 * 0.229, 255 * 0.224, 255 * 0.225)
offsets : double or iter-of-doubles, optional
Specifies an offset for each channel in the input data. The final input
data is set after applying scaling and subtracting the specified offsets.
Default: (255*0.485, 255*0.456, 255*0.406)
random_flip : string, optional
Specifies how to flip the data in the input layer when image data is
used. Approximately half of the input data is subject to flipping.
Valid Values: 'h', 'hv', 'v', 'none'
random_crop : string, optional
Specifies how to crop the data in the input layer when image data is
used. Images are cropped to the values that are specified in the width
and height parameters. Only the images with one or both dimensions
that are larger than those sizes are cropped.
Valid Values: 'none', 'unique', 'randomresized', 'resizethencrop'
random_mutation : string, optional
Specifies how to apply data augmentations/mutations to the data in the input layer.
Valid Values: 'none', 'random'
alpha : int, optional
Specifies the width multiplier in the MobileNet paper
Default: 1
alpha : int, optional
Returns
-------
:class:`Model`
References
----------
https://arxiv.org/abs/1801.04381
'''
def _make_divisible(v, divisor, min_value=None):
# make number of channel divisible
if min_value is None:
min_value = divisor
new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if new_v < 0.9 * v:
new_v += divisor
return new_v
def _inverted_res_block(inputs, in_channels, expansion, stride, alpha,
filters, block_id):
"""
Inverted Residual Block
Parameters
----------
inputs:
Input tensor
in_channels:
Specifies the number of input tensor's channel
expansion:
expansion factor always applied to the input size.
stride:
the strides of the convolution
alpha:
width multiplier.
filters:
the dimensionality of the output space.
block_id:
block id used for naming layers
"""
pointwise_conv_filters = int(filters * alpha)
pointwise_filters = _make_divisible(pointwise_conv_filters, 8)
x = inputs
prefix = 'block_{}_'.format(block_id)
n_groups = in_channels
if block_id:
# Expand
n_groups = expansion * in_channels
x = Conv2d(expansion * in_channels,
1,
include_bias=False,
act='identity',
name=prefix + 'expand')(x)
x = BN(name=prefix + 'expand_BN', act='identity')(x)
else:
prefix = 'expanded_conv_'
# Depthwise
x = GroupConv2d(n_groups,
n_groups,
3,
stride=stride,
act='identity',
include_bias=False,
name=prefix + 'depthwise')(x)
x = BN(name=prefix + 'depthwise_BN', act='relu')(x)
# Project
x = Conv2d(pointwise_filters,
1,
include_bias=False,
act='identity',
name=prefix + 'project')(x)
x = BN(name=prefix + 'project_BN',
act='identity')(x) # identity activation on narrow tensor
if in_channels == pointwise_filters and stride == 1:
return Res(name=prefix + 'add')([inputs, x]), pointwise_filters
return x, pointwise_filters
parameters = locals()
input_parameters = get_layer_options(input_layer_options, parameters)
inp = Input(**input_parameters, name='data')
# compared with mobilenetv1, v2 introduces inverted residual structure.
# and Non-linearities in narrow layers are removed.
# inverted residual block does three convolutins: first is 1*1 convolution, second is depthwise convolution,
# third is 1*1 convolution but without any non-linearity
first_block_filters = _make_divisible(32 * alpha, 8)
x = Conv2d(first_block_filters,
3,
stride=2,
include_bias=False,
name='Conv1',
act='identity')(inp)
x = BN(name='bn_Conv1', act='relu')(x)
x, n_channels = _inverted_res_block(x,
first_block_filters,
filters=16,
alpha=alpha,
stride=1,
expansion=1,
block_id=0)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=24,
alpha=alpha,
stride=2,
expansion=6,
block_id=1)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=24,
alpha=alpha,
stride=1,
expansion=6,
block_id=2)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=32,
alpha=alpha,
stride=2,
expansion=6,
block_id=3)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=32,
alpha=alpha,
stride=1,
expansion=6,
block_id=4)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=32,
alpha=alpha,
stride=1,
expansion=6,
block_id=5)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=64,
alpha=alpha,
stride=2,
expansion=6,
block_id=6)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=64,
alpha=alpha,
stride=1,
expansion=6,
block_id=7)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=64,
alpha=alpha,
stride=1,
expansion=6,
block_id=8)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=64,
alpha=alpha,
stride=1,
expansion=6,
block_id=9)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=96,
alpha=alpha,
stride=1,
expansion=6,
block_id=10)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=96,
alpha=alpha,
stride=1,
expansion=6,
block_id=11)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=96,
alpha=alpha,
stride=1,
expansion=6,
block_id=12)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=160,
alpha=alpha,
stride=2,
expansion=6,
block_id=13)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=160,
alpha=alpha,
stride=1,
expansion=6,
block_id=14)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=160,
alpha=alpha,
stride=1,
expansion=6,
block_id=15)
x, n_channels = _inverted_res_block(x,
n_channels,
filters=320,
alpha=alpha,
stride=1,
expansion=6,
block_id=16)
# no alpha applied to last conv as stated in the paper:
# if the width multiplier is greater than 1 we increase the number of output channels
if alpha > 1.0:
last_block_filters = _make_divisible(1280 * alpha, 8)
else:
last_block_filters = 1280
x = Conv2d(last_block_filters,
1,
include_bias=False,
name='Conv_1',
act='identity')(x)
x = BN(name='Conv_1_bn', act='relu')(x)
x = GlobalAveragePooling2D(name="Global_avg_pool")(x)
x = OutputLayer(n=n_classes)(x)
model = Model(conn, inp, x, model_table)
model.compile()
return model
def MobileNetV1(conn,
model_table='MobileNetV1',
n_classes=1000,
n_channels=3,
width=224,
height=224,
random_flip=None,
random_crop=None,
random_mutation=None,
norm_stds=(255 * 0.229, 255 * 0.224, 255 * 0.225),
offsets=(255 * 0.485, 255 * 0.456, 255 * 0.406),
alpha=1,
depth_multiplier=1):
'''
Generates a deep learning model with the MobileNetV1 architecture.
The implementation is revised based on
https://github.com/keras-team/keras-applications/blob/master/keras_applications/mobilenet.py
Parameters
----------
conn : CAS
Specifies the CAS connection object.
model_table : string or dict or CAS table, optional
Specifies the CAS table to store the deep learning model.
n_classes : int, optional
Specifies the number of classes. If None is assigned, the model will
automatically detect the number of classes based on the training set.
Default: 1000
n_channels : int, optional
Specifies the number of the channels (i.e., depth) of the input layer.
Default: 3
width : int, optional
Specifies the width of the input layer.
Default: 32
height : int, optional
Specifies the height of the input layer.
Default: 32
random_flip : string, optional
Specifies how to flip the data in the input layer when image data is
used. Approximately half of the input data is subject to flipping.
Valid Values: 'h', 'hv', 'v', 'none'
random_crop : string, optional
Specifies how to crop the data in the input layer when image data is
used. Images are cropped to the values that are specified in the width
and height parameters. Only the images with one or both dimensions
that are larger than those sizes are cropped.
Valid Values: 'none', 'unique', 'randomresized', 'resizethencrop'
random_mutation : string, optional
Specifies how to apply data augmentations/mutations to the data in the input layer.
Valid Values: 'none', 'random'
norm_stds : double or iter-of-doubles, optional
Specifies a standard deviation for each channel in the input data.
The final input data is normalized with specified means and standard deviations.
Default: (255*0.229, 255*0.224, 255*0.225)
offsets : double or iter-of-doubles, optional
Specifies an offset for each channel in the input data. The final input
data is set after applying scaling and subtracting the specified offsets.
Default: (255*0.485, 255*0.456, 255*0.406)
alpha : int, optional
Specifies the width multiplier in the MobileNet paper
Default: 1
depth_multiplier : int, optional
Specifies the number of depthwise convolution output channels for each input channel.
Default: 1
Returns
-------
:class:`Model`
References
----------
https://arxiv.org/pdf/1605.07146.pdf
'''
def _conv_block(inputs, filters, alpha, kernel=3, stride=1):
"""
Adds an initial convolution layer (with batch normalization
inputs:
Input tensor
filters:
the dimensionality of the output space
alpha: controls the width of the network.
- If `alpha` < 1.0, proportionally decreases the number
of filters in each layer.
- If `alpha` > 1.0, proportionally increases the number
of filters in each layer.
- If `alpha` = 1, default number of filters from the paper
are used at each layer.
kernel:
specifying the width and height of the 2D convolution window.
strides:
the strides of the convolution
"""
filters = int(filters * alpha)
x = Conv2d(filters,
kernel,
act='identity',
include_bias=False,
stride=stride,
name='conv1')(inputs)
x = BN(name='conv1_bn', act='relu')(x)
return x, filters
def _depthwise_conv_block(inputs,
n_groups,
pointwise_conv_filters,
alpha,
depth_multiplier=1,
stride=1,
block_id=1):
"""Adds a depthwise convolution block.
inputs:
Input tensor
n_groups : int
number of groups
pointwise_conv_filters:
the dimensionality of the output space
alpha: controls the width of the network.
- If `alpha` < 1.0, proportionally decreases the number
of filters in each layer.
- If `alpha` > 1.0, proportionally increases the number
of filters in each layer.
- If `alpha` = 1, default number of filters from the paper
are used at each layer.
depth_multiplier:
The number of depthwise convolution output channels
strides: An integer or tuple/list of 2 integers,
specifying the strides of the convolution
block_id: Integer, a unique identification designating
the block number.
"""
pointwise_conv_filters = int(pointwise_conv_filters * alpha)
x = GroupConv2d(n_groups * depth_multiplier,
n_groups,
3,
stride=stride,
act='identity',
include_bias=False,
name='conv_dw_%d' % block_id)(inputs)
x = BN(name='conv_dw_%d_bn' % block_id, act='relu')(x)
x = Conv2d(pointwise_conv_filters,
1,
act='identity',
include_bias=False,
stride=1,
name='conv_pw_%d' % block_id)(x)
x = BN(name='conv_pw_%d_bn' % block_id, act='relu')(x)
return x, pointwise_conv_filters
parameters = locals()
input_parameters = get_layer_options(input_layer_options, parameters)
inp = Input(**input_parameters, name='data')
# the model down-sampled for 5 times by performing stride=2 convolution on
# conv_dw_1, conv_dw_2, conv_dw_4, conv_dw_6, conv_dw_12
# for each block, we use depthwise convolution with kernel=3 and point-wise convolution to save computation
x, depth = _conv_block(inp, 32, alpha, stride=2)
x, depth = _depthwise_conv_block(x,
depth,
64,
alpha,
depth_multiplier,
block_id=1)
x, depth = _depthwise_conv_block(x,
depth,
128,
alpha,
depth_multiplier,
stride=2,
block_id=2)
x, depth = _depthwise_conv_block(x,
depth,
128,
alpha,
depth_multiplier,
block_id=3)
x, depth = _depthwise_conv_block(x,
depth,
256,
alpha,
depth_multiplier,
stride=2,
block_id=4)
x, depth = _depthwise_conv_block(x,
depth,
256,
alpha,
depth_multiplier,
block_id=5)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
stride=2,
block_id=6)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
block_id=7)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
block_id=8)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
block_id=9)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
block_id=10)
x, depth = _depthwise_conv_block(x,
depth,
512,
alpha,
depth_multiplier,
block_id=11)
x, depth = _depthwise_conv_block(x,
depth,
1024,
alpha,
depth_multiplier,
stride=2,
block_id=12)
x, depth = _depthwise_conv_block(x,
depth,
1024,
alpha,
depth_multiplier,
block_id=13)
x = GlobalAveragePooling2D(name="Global_avg_pool")(x)
x = OutputLayer(n=n_classes)(x)
model = Model(conn, inp, x, model_table)
model.compile()
return model
def test_inputs(self):
conv1 = Conv2d(2)
with self.assertRaises(ValueError):
output1 = OutputLayer(2)(conv1)
Model(conn=self.s, inputs=[conv1], outputs=[output1])
