def builder(self, *params):
try:
return ModelFromJsonTF(None, json_str=self._json(*params))
except:
str_param = ','.join('{0}={1!r}'.format(k, v)
for k, v in self._args(*params).items())
print("Failed to build model with params: {}".format(str_param))
return None
per_process_gpu_memory_fraction=0.1, #was 0.0
allow_growth=True,
visible_device_list=device[-1] if 'gpu' in device else '')
if hide_device:
gpu_options = K.tf.GPUOptions(
per_process_gpu_memory_fraction=0.0,
allow_growth=True,
)
K.set_session(
K.tf.Session(
config=K.tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False,
gpu_options=gpu_options)))
if args.tf:
model_builder = ModelFromJsonTF(comm,
args.model_json,
device_name=device,
weights=args.model_weights)
print("Process {0} using device {1}".format(
comm.Get_rank(), model_builder.device))
else:
model_builder = ModelFromJson(comm,
args.model_json,
weights=args.model_weights)
print("Process {0} using device {1}".format(
comm.Get_rank(), device))
os.environ[
'THEANO_FLAGS'] = "profile=%s,device=%s,floatX=float32" % (
args.profile, device.replace('gpu', 'cuda'))
# GPU ops need to be executed synchronously in order for profiling to make sense
if args.profile:
os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
def builder(self, *params):
args = dict(zip([p.name for p in self.parameters], params))
model_json = self.model_fn(**args)
return ModelFromJsonTF(None, json_str=model_json)