def get_network(type, placeholder_input, sess_for_load=None):
if type == 'mobilenet':
net = MobilenetNetwork({'image': placeholder_input}, conv_width=0.75, conv_width2=0.50)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_accurate':
net = MobilenetNetwork({'image': placeholder_input}, conv_width=1.00)
pretrain_path = 'pretrained/mobilenet_v1_1.0_224_2017_06_14/mobilenet_v1_1.0_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_fast':
net = MobilenetNetwork({'image': placeholder_input}, conv_width=0.50)
pretrain_path = 'pretrained/mobilenet_v1_0.50_224_2017_06_14/mobilenet_v1_0.50_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'cmu':
net = CmuNetwork({'image': placeholder_input})
pretrain_path = 'numpy/openpose_coco.npy'
last_layer = 'Mconv7_stage6_L{aux}'
else:
raise Exception('Invalid Mode.')
if sess_for_load is not None:
if type == 'cmu':
net.load('./models/numpy/openpose_coco.npy', sess_for_load)
else:
ckpts = {
'mobilenet': os.path.join(_get_base_path(), 'pretrained/mobilenet_v1_0.50_224_2017_06_14/mobilenet_v1_0.50_224.ckpt'),
'mobilenet_fast': os.path.join(_get_base_path(), 'pretrained/mobilenet_fast/-163000'),
'mobilenet_accurate': os.path.join(_get_base_path(), 'pretrained/mobilenet_accurate/model-170000')
}
# loader = tf.train.import_meta_graph(ckpts[type] + '.meta')
loader = tf.train.Saver()
loader.restore(sess_for_load, ckpts[type])
return net, os.path.join(_get_base_path(), pretrain_path), last_layer
def get_variables(model_path, height , width):
input_node = tf.placeholder(tf.float32, shape=(1, height, width, 3), name='image')
net = MobilenetNetwork({'image': input_node}, trainable=False, conv_width=0.75, conv_width2=0.50)
saver = tf.train.Saver(max_to_keep=100)
config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=False)
with tf.Session(config=config) as sess:
saver.restore(sess, model_path)
variables = tf.global_variables()
variables = [(v.name, v.eval(session=sess).copy(order='C')) for v in variables]
return variables
def get_network(type, placeholder_input, sess_for_load=None, trainable=False):
if type == 'mobilenet':
net = MobilenetNetwork({'image': placeholder_input},
trainable=trainable,
conv_width=0.75,
conv_width2=0.50)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_accurate':
net = MobilenetNetwork({'image': placeholder_input},
trainable=trainable,
conv_width=1.00)
pretrain_path = 'pretrained/mobilenet_v1_1.0_224_2017_06_14/mobilenet_v1_1.0_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_fast':
net = MobilenetNetwork({'image': placeholder_input},
trainable=trainable,
conv_width=0.50)
pretrain_path = 'pretrained/mobilenet_v1_0.50_224_2017_06_14/mobilenet_v1_0.50_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'cmu':
net = CmuNetwork({'image': placeholder_input})
pretrain_path = 'numpy/openpose_coco.npy'
last_layer = 'Mconv7_stage6_L{aux}'
else:
raise Exception('Invalid Mode.')
if sess_for_load is not None:
if type == 'cmu':
net.load('./models/numpy/openpose_coco.npy', sess_for_load)
else:
s = '%dx%d' % (placeholder_input.shape[2],
placeholder_input.shape[1])
ckpts = {
'mobilenet': 'trained/mobilenet_%s/model-release' % s,
'mobilenet_fast': 'trained/mobilenet_fast/model-163000',
'mobilenet_accurate': 'trained/mobilenet_accurate/model-170000'
}
loader = tf.train.Saver()
loader.restore(sess_for_load,
os.path.join(_get_base_path(), ckpts[type]))
return net, os.path.join(_get_base_path(), pretrain_path), last_layer
def get_network(type, placeholder_input, sess_for_load=None, trainable=True):
if type == 'mobilenet':
net = MobilenetNetwork({'image': placeholder_input}, conv_width=0.75, conv_width2=1.00, trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_fast':
net = MobilenetNetwork({'image': placeholder_input}, conv_width=0.5, conv_width2=0.5, trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_accurate':
net = MobilenetNetwork({'image': placeholder_input}, conv_width=1.00, conv_width2=1.00, trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_thin':
net = MobilenetNetworkThin({'image': placeholder_input}, conv_width=0.75, conv_width2=0.50, trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_1.0_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'cmu':
net = CmuNetwork({'image': placeholder_input}, trainable=trainable)
pretrain_path = 'numpy/openpose_coco.npy'
last_layer = 'Mconv7_stage6_L{aux}'
elif type == 'vgg':
net = CmuNetwork({'image': placeholder_input}, trainable=trainable)
pretrain_path = 'numpy/openpose_vgg16.npy'
last_layer = 'Mconv7_stage6_L{aux}'
else:
raise Exception('Invalid Mode.')
if sess_for_load is not None:
if type == 'cmu' or type == 'vgg':
net.load(os.path.join(_get_base_path(), pretrain_path), sess_for_load)
else:
s = '%dx%d' % (placeholder_input.shape[2], placeholder_input.shape[1])
ckpts = {
'mobilenet': 'trained/mobilenet_%s/model-246038' % s,
'mobilenet_thin': 'trained/mobilenet_thin_%s/model-449003' % s,
'mobilenet_fast': 'trained/mobilenet_fast_%s/model-189000' % s,
'mobilenet_accurate': 'trained/mobilenet_accurate/model-170000'
}
loader = tf.train.Saver()
loader.restore(sess_for_load, os.path.join(_get_base_path(), ckpts[type]))
return net, os.path.join(_get_base_path(), pretrain_path), last_layer
def get_network(type, placeholder_input, sess_for_load=None, trainable=True):
if type == 'mobilenet':
net = MobilenetNetwork({'image': placeholder_input},
conv_width=0.75,
conv_width2=1.00,
trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_fast':
net = MobilenetNetwork({'image': placeholder_input},
conv_width=0.5,
conv_width2=0.5,
trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_accurate':
net = MobilenetNetwork({'image': placeholder_input},
conv_width=1.00,
conv_width2=1.00,
trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_1.0_224_2017_06_14/mobilenet_v1_1.0_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_thin':
net = MobilenetNetworkThin({'image': placeholder_input},
conv_width=0.75,
conv_width2=0.50,
trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'cmu':
net = CmuNetwork({'image': placeholder_input}, trainable=trainable)
pretrain_path = 'numpy/openpose_coco.npy'
last_layer = 'Mconv7_stage6_L{aux}'
elif type == 'vgg':
net = CmuNetwork({'image': placeholder_input}, trainable=trainable)
pretrain_path = 'numpy/openpose_vgg16.npy'
last_layer = 'Mconv7_stage6_L{aux}'
else:
raise Exception('Invalid Mode.')
pretrain_path_full = os.path.join(_get_base_path(), pretrain_path)
if sess_for_load is not None:
if type == 'cmu' or type == 'vgg':
if not os.path.isfile(pretrain_path_full):
raise Exception('Model file doesn\'t exist, path=%s' %
pretrain_path_full)
net.load(os.path.join(_get_base_path(), pretrain_path),
sess_for_load)
else:
s = '%dx%d' % (placeholder_input.shape[2],
placeholder_input.shape[1])
ckpts = {
'mobilenet': 'trained/mobilenet_%s/model-246038' % s,
'mobilenet_thin': 'trained/mobilenet_thin_%s/model-449003' % s,
'mobilenet_fast': 'trained/mobilenet_fast_%s/model-189000' % s,
'mobilenet_accurate': 'trained/mobilenet_accurate/model-170000'
}
ckpt_path = os.path.join(_get_base_path(), ckpts[type])
loader = tf.train.Saver()
try:
loader.restore(sess_for_load, ckpt_path)
except Exception as e:
raise Exception('Fail to load model files. \npath=%s\nerr=%s' %
(ckpt_path, str(e)))
return net, pretrain_path_full, last_layer
shape=(None, args.input_height,
args.input_width, 3),
name='image')
with tf.Session(config=config) as sess:
if args.model == 'dsconv':
net = DSConvNetwork({'image': input_node}, trainable=False)
net.load('./models/numpy/fastopenpose_coco_v170729.npy', sess)
last_layer = 'Mconv7_stage{stage}_L{aux}'
elif args.model == 'cmu':
net = CmuNetwork({'image': input_node}, trainable=False)
net.load('./models/numpy/openpose_coco.npy', sess)
last_layer = 'Mconv7_stage{stage}_L{aux}'
elif args.model == 'mobilenet_1.0':
net = MobilenetNetwork({'image': input_node},
trainable=False,
conv_width=1.0)
sess.run(tf.global_variables_initializer()) # TODO
last_layer = 'MConv_Stage{stage}_L{aux}_5'
elif args.model == 'mobilenet_0.50':
net = MobilenetNetwork({'image': input_node},
trainable=False,
conv_width=0.50)
sess.run(tf.global_variables_initializer()) # TODO
last_layer = 'MConv_Stage{stage}_L{aux}_5'
else:
raise Exception('Invalid Mode.')
logging.debug('read image+')
image = cv2.imread(args.imgpath)
image = cv2.resize(image, (args.input_width, args.input_height))
def get_model(sess, height, width):
init = tf.global_variables_initializer()
sess.run(init)
net = MobilenetNetwork({'image': input_node}
, trainable=False, conv_width=0.75, conv_width2=0.50)
K.set_session(sess)
conv_width=0.75
conv_width2=0.50
min_depth = 8
depth = lambda d: max(int(d * conv_width), min_depth)
depth2 = lambda d: max(int(d * conv_width2), min_depth)
image = Input(shape=(width, height, 3),name="image")
x = Conv2D(depth(32),(3,3)
, strides=2
, use_bias=False
, name="Conv2d_0"
, trainable = False
, padding='same'
, kernel_regularizer=l2(0.04)
)(image)
x = BatchNormalization(scale=True, name='Conv2d_0_bn')(x,training=False)
x = Activation('relu', name='Conv2d_0_relu')(x)
x = separable_conv(x,depth(64),(3,3),1,name='Conv2d_1')
x = separable_conv(x,depth(128),(3,3),2,name='Conv2d_2')
o3 = separable_conv(x,depth(128),(3,3),1,name='Conv2d_3')
x = separable_conv(o3,depth(256),(3,3),2,name='Conv2d_4')
x = separable_conv(x,depth(256),(3,3),1,name='Conv2d_5')
x = separable_conv(x,depth(512),(3,3),1,name='Conv2d_6')
o7 = separable_conv(x,depth(512),(3,3),1,name='Conv2d_7')
x = separable_conv(o7,depth(512),(3,3),1,name='Conv2d_8')
x = separable_conv(x,depth(512),(3,3),1,name='Conv2d_9')
x = separable_conv(x,depth(512),(3,3),1,name='Conv2d_10')
o11 = separable_conv(x,depth(512),(3,3),1,name='Conv2d_11')
o3_pool = MaxPooling2D((2, 2),(2, 2),padding='same')(o3)
feat_concat = concatenate([o3_pool,o7,o11], axis=3)
prefix = 'MConv_Stage1'
r1 = separable_conv(feat_concat,depth2(128),(3,3),1,name=prefix + '_L1_1')
r1 = separable_conv(r1,depth2(128),(3,3),1,name=prefix + '_L1_2')
r1 = separable_conv(r1,depth2(128),(3,3),1,name=prefix + '_L1_3')
r1 = separable_conv(r1,depth2(512),(1,1),1,name=prefix + '_L1_4')
r1 = separable_conv(r1,38,(1,1),1,relu=False,name=prefix + '_L1_5')
# concat = Input(shape=(46, 46, 864))
r2 = separable_conv(feat_concat,depth2(128),(3,3),1,name=prefix + '_L2_1')
r2 = separable_conv(r2,depth2(128),(3,3),1,name=prefix + '_L2_2')
r2 = separable_conv(r2,depth2(128),(3,3),1,name=prefix + '_L2_3')
r2 = separable_conv(r2,depth2(512),(1,1),1,name=prefix + '_L2_4')
r2 = separable_conv(r2,19,(1,1),1,relu=False,name=prefix + '_L2_5')
for stage_id in range(5):
prefix = 'MConv_Stage%d' % (stage_id + 2)
cc = concatenate([r1,r2,feat_concat], axis=3)
r1 = separable_conv(cc,depth2(128),(3,3),1,name=prefix + '_L1_1')
r1 = separable_conv(r1,depth2(128),(3,3),1,name=prefix + '_L1_2')
r1 = separable_conv(r1,depth2(128),(3,3),1,name=prefix + '_L1_3')
r1 = separable_conv(r1,depth2(128),(1,1),1,name=prefix + '_L1_4')
r1 = separable_conv(r1,38,(1,1),1,relu=False,name=prefix + '_L1_5')
r2 = separable_conv(cc,depth2(128),(3,3),1,name=prefix + '_L2_1')
r2 = separable_conv(r2,depth2(128),(3,3),1,name=prefix + '_L2_2')
r2 = separable_conv(r2,depth2(128),(3,3),1,name=prefix + '_L2_3')
r2 = separable_conv(r2,depth2(128),(1,1),1,name=prefix + '_L2_4')
r2 = separable_conv(r2,19,(1,1),1,relu=False,name=prefix + '_L2_5')
out = concatenate([r2, r1],axis=3)
print(out)
model = Model(image, out)
layers = getTupleLayer("MobilenetV1","Conv2d_0")
model = setLayer(model,layers)
for (i, layer) in enumerate(global_layers):
# idx = i + 2
n = layer.split("_")
n.pop()
prefix = ""
if n[0] == "Conv2d":
prefix = "MobilenetV1"
if n[0] == "MConv":
prefix = "Openpose"
if prefix != "":
layers = getTupleLayer(prefix,layer)
model = setLayer(model,layers)
if not os.path.exists("output"):
os.mkdir("output")
model.save('output/predict.hd5')
# plot_model(model, to_file='model_shape.png', show_shapes=True)
# img = load_img(args.imgpath, target_size=(args.input_width, args.input_height))
# img = np.expand_dims(img, axis=0)
# print(img.shape)
# prediction = model.predict(img)
# prediction = prediction[0]
# print("#output")
# print(prediction.shape)
# print(prediction[0:1, 0:1, :])
# print(np.mean(prediction))
# np.save('output/prediction.npy', prediction, allow_pickle=False)
return model
val_image = val_image.astype(float)
val_image = val_image * (2.0 / 255.0) - 1.0
# define model for multi-gpu
q_inp_split = tf.split(q_inp, args.gpus)
output_vectmap = []
output_heatmap = []
vectmap_losses = []
heatmap_losses = []
for gpu_id in range(args.gpus):
with tf.device(tf.DeviceSpec(device_type="GPU", device_index=gpu_id)):
with tf.variable_scope(tf.get_variable_scope(),
reuse=(gpu_id > 0)):
if args.model == 'mobilenet_1.0':
net = MobilenetNetwork({'image': q_inp_split[gpu_id]},
conv_width=1.0)
pretrain_path = './models/pretrained/mobilenet_v1_1.0_224_2017_06_14/mobilenet_v1_1.0_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif args.model == 'mobilenet_0.75':
net = MobilenetNetwork({'image': q_inp_split[gpu_id]},
conv_width=0.75)
pretrain_path = './models/pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif args.model == 'mobilenet_0.50':
net = MobilenetNetwork({'image': q_inp_split[gpu_id]},
conv_width=0.50)
pretrain_path = './models/pretrained/mobilenet_v1_0.50_224_2017_06_14/mobilenet_v1_0.50_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif args.model == 'cmu':
net = CmuNetwork({'image': q_inp_split[gpu_id]})
pretrain_path = './models/numpy/openpose_coco.npy'
def get_network(type, placeholder_input, sess_for_load=None, trainable=True):
if type == 'mobilenet':
net = MobilenetNetwork({'image': placeholder_input},
conv_width=0.75,
conv_width2=1.00,
trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_fast':
net = MobilenetNetworkFast({'image': placeholder_input},
conv_width=0.5,
conv_width2=0.5,
trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_0.50_224_2017_06_14/mobilenet_v1_0.50_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_accurate':
net = MobilenetNetworkThin({'image': placeholder_input},
conv_width=1.00,
conv_width2=0.50,
trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_1.0_224_2017_06_14/mobilenet_v1_1.0_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_thin':
net = MobilenetNetworkThin({'image': placeholder_input},
conv_width=0.75,
conv_width2=0.750,
trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_original':
net = MobilenetNetworkOriginal({'image': placeholder_input},
conv_width=0.75,
conv_width2=0.50,
trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v1_0.75_224_2017_06_14/mobilenet_v1_0.75_224.ckpt'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'mobilenet_v2':
net = MobilenetNetworkV2({'image': placeholder_input},
conv_width=1,
conv_width2=0.50,
trainable=trainable)
pretrain_path = 'pretrained/mobilenet_v2/model.ckpt-1450000'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'cmu':
net = CmuNetwork({'image': placeholder_input}, trainable=trainable)
pretrain_path = 'numpy/openpose_coco.npy'
last_layer = 'Mconv7_stage6_L{aux}'
elif type == 'vgg':
net = CmuNetwork({'image': placeholder_input}, trainable=trainable)
pretrain_path = 'numpy/openpose_coco.npy'
last_layer = 'Mconv7_stage6_L{aux}'
elif type == 'resnet32':
net = Resnet32({'image': placeholder_input},
conv_width=0.75,
conv_width2=0.50,
trainable=trainable)
pretrain_path = 'numpy/resnet32.npy'
last_layer = 'MConv_Stage6_L{aux}_5'
elif type == 'vgg16x4':
net = VGG16x4Network({'image': placeholder_input},
conv_width=0.75,
conv_width2=0.75,
trainable=trainable)
pretrain_path = 'numpy/vgg16x4.npy'
last_layer = 'Mconv7_stage6_L{aux}'
else:
raise Exception('Invalid Mode.')
if sess_for_load is not None:
if type == 'cmu' or type == 'vgg':
net.load(os.path.join(_get_base_path(), pretrain_path),
sess_for_load)
else:
s = '%dx%d' % (placeholder_input.shape[2],
placeholder_input.shape[1])
ckpts = {
'mobilenet': 'models/trained/mobilenet_%s/model-53008' % s,
'mobilenet_thin':
'models/trained/mobilenet_thin_432x368/model-160001',
'mobilenet_original':
'models/trained/mobilenet_thin_benchmark/model-388003',
'mobilenet_v2': 'models/trained/mobilenet_v2/model-218000',
'vgg16x4': 'models/trained/vgg16x4_0.75/model-35000',
}
loader = tf.train.Saver()
loader.restore(sess_for_load, os.path.join(base_path, ckpts[type]))
return net, os.path.join(_get_base_path(), pretrain_path), last_layer