def get_image_agent(model_path):
log_dir = Path(model_path).parent
config = bzu.load_json(str(log_dir / 'config.json'))
agent_args = config.get('agent_args', dict())
model = image.ImagePolicyModelSS(**config['model_args'])
model.load_state_dict(torch.load(str(model_path)))
model.eval()
agent_args['model'] = model
return image.ImageAgent(**agent_args)
def load_model(model_path):
''' Carica i pesi del modello preaddestrato e li inserisce nella classe PytorchClassifier, che fa da interfaccia tra modello e ART'''
config = bzu.load_json('/'.join([model_path, 'config.json']))
model_args = config['model_args']
model_name = model_args['model']
model_class = image.ImagePolicyModelSS
model = model_class(**config['model_args'])
model.load_state_dict(torch.load('/'.join([model_path, 'model-10.th'])))
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(),lr=0.0001)
classifier = PyTorchClassifier(model=model, loss=criterion,optimizer=optimizer, input_shape=(3, 160, 384),
nb_classes=10,device_type='gpu')
return classifier
def load_model(model_path):
config = bzu.load_json('/'.join([model_path, 'config.json']))
model_args = config['model_args']
model_name = model_args['model']
model_class = image.ImagePolicyModelSS
model = model_class(**config['model_args'])
model.load_state_dict(torch.load('/'.join([model_path, 'model-10.th'])))
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(),lr=0.0001)
classifier = PyTorchClassifier(model=model, loss=criterion,optimizer=optimizer, input_shape=(3, 160, 184),
nb_classes=10,device_type='gpu')
return classifier
def run(model_path,
port,
suite,
big_cam,
seed,
autopilot,
resume,
max_run=10,
show=False):
log_dir = model_path.parent
config = bzu.load_json(str(log_dir / 'config.json'))
total_time = 0.0
for suite_name in get_suites(suite):
tick = time.time()
import pdb
pdb.set_trace()
benchmark_dir = log_dir / 'benchmark' / model_path.stem / (
'%s_seed%d' % (suite_name, seed))
benchmark_dir.mkdir(parents=True, exist_ok=True)
with make_suite(suite_name, port=port, big_cam=big_cam) as env:
agent_maker = _agent_factory_hack(model_path, config, autopilot)
run_benchmark(agent_maker,
env,
benchmark_dir,
seed,
autopilot,
resume,
max_run=max_run,
show=show)
elapsed = time.time() - tick
total_time += elapsed
print('%s: %.3f hours.' % (suite_name, elapsed / 3600))
print('Total time: %.3f hours.' % (total_time / 3600))