def build_model(self, mobile=False, show=False):
if mobile:
self.model = create_hourglass_network(self.num_classes, self.num_stacks, self.inres, self.outres, bottleneck_mobile)
else:
self.model = create_hourglass_network(self.num_classes, self.num_stacks, self.inres, self.outres, bottleneck_block)
# show model summary and layer name
if show :
self.model.summary()
def main():
model = create_hourglass_network(16, 2, (256, 256), (64, 64))
model.summary()
print(len(model.output_layers))
plot_model(model, 'hg_s2.png', show_shapes=True)
for layer in model.output_layers:
print(layer.output_shape)
def main():
model = create_hourglass_network(16, 8, (256, 256), (64, 64),
bottleneck_mobile)
model.summary()
print len(model.output_layers)
#plot_model(model, 'hg_s2.png', show_shapes=True)
for layer in model.output_layers:
print layer.output_shape
def build_model(self, mobile=False, show=True):
if mobile:
self.model = create_hourglass_network(self.num_classes,
self.num_stacks,
self.num_channels,
self.inres, self.outres,
bottleneck_mobile)
else:
self.model = create_hourglass_network(self.num_classes,
self.num_stacks,
self.num_channels,
self.inres, self.outres,
bottleneck_block)
# show model summary and layer name
if show:
self.model.summary()
model_json = self.model.to_json()
with open("HourglassNet.json", "w") as json_file:
json_file.write(model_json)
return self.model
wandb.config.downsample_ratio), # 100ms
'min_peak_height':
0.3,
})
mit_loader_train = MITLoader('train', wandb.config)
mit_loader_val = MITLoader('valid', wandb.config)
training_set = DataGenerator(mit_loader_train, wandb.config.batch_size)
validation_set = DataGenerator(mit_loader_val, wandb.config.batch_size)
print('#batch_train, #batch_valid: {}, {}'.format(len(training_set),
len(validation_set)))
model = create_hourglass_network(
len(wandb.config.labels), wandb.config.number_hourglass_modules,
wandb.config.number_inner_channels,
(int(wandb.config.sampling_rate * wandb.config.length_s), 1),
bottleneck_block, wandb.config.hourglass_module_layers)
model.summary()
model.compile(optimizer='adam',
loss=focal_loss(
length_head_ignore=int(wandb.config.sampling_rate *
wandb.config.head_ignore_s *
wandb.config.downsample_ratio),
length_tail_ignore=int(wandb.config.sampling_rate *
wandb.config.tail_ignore_s *
wandb.config.downsample_ratio)),
run_eagerly=False)
model.fit(
training_set,