###################################################################
# TEST
# plot cost
fig = plt.figure()
plt.semilogy(train_costs)
plt.show(block=False)
fig.savefig('./eval/' + eval_prefix + '/' + eval_prefix + '_cost.png')
fig = plt.figure()
plt.semilogy(val_errs)
plt.show(block=False)
fig.savefig('./eval/' + eval_prefix + '/' + eval_prefix + '_errs.png')
# save results
poseNet.save("./eval/{}/net_{}.pkl".format(eval_prefix, eval_prefix))
# poseNet.load("./eval/{}/net_{}.pkl".format(eval_prefix,eval_prefix))
# add prior to network
cfg = HiddenLayerParams(inputDim=(batchSize, train_gt3D_embed.shape[1]),
outputDim=(batchSize,
numpy.prod(train_gt3D.shape[1:])),
activation=None)
pcalayer = HiddenLayer(rng,
poseNet.layers[-1].output,
cfg,
copyLayer=None,
layerNum=len(poseNet.layers))
pcalayer.W.set_value(pca.components_)
pcalayer.b.set_value(pca.mean_)
poseNet.layers.append(pcalayer)
class Network:
def __init__(self):
self.type = 1 # 1 = Regressor
self.network_model = '0' # 0 = Tiny BB8, 1 = BB8 - VGG arch.
self.batch_size = 128
self.optimizer = 'MOMENTUM'
self.learning_rate = 0.001
self.steps = []
self.scales = []
self.nb_epoch = 300
self.network = None
self.trainer = None
self.network_name = None
self.save_path = './'
self.validation_size = 5000
self.train_set_para = None
self.config = None
self.nb_process = 10
def setup_from_config(self):
if self.config is not None:
with open(self.config, 'r') as f:
config = yaml.load(f)
for key in config.keys():
value = config[key]
print('set {0} to {1}'.format(key, value))
setattr(self, key, value)
def update(self):
self.setup_from_config()
self.print_type()
sys.path.insert(0,
self.train_set_para[:self.train_set_para.rindex('/')])
self.create_training = __import__(
self.train_set_para[self.train_set_para.rindex('/') + 1:],
fromlist=['init', 'pre_create_data', 'create_data', 'get_dim'])
self.create_training.init()
self.regressor()
def regressor(self):
regressorNetType = regressorNetType = int(self.network_model)
batch_size = int(self.batch_size)
learning_rate = float(self.learning_rate)
optimizer = self.optimizer
assert len(self.steps) == len(self.scales)
n_chan, h_in, w_in, output_dim = self.create_training.get_dim()
assert n_chan is not None
assert h_in is not None
assert w_in is not None
assert output_dim is not None
nb_training = int(self.nb_training)
rng = np.random.RandomState(23455)
#theano.config.compute_test_value = 'warn'
theano.config.exception_verbosity = 'high'
regressorNetParams = PoseRegNetParams(type=regressorNetType,
n_chan=n_chan,
w_in=w_in,
h_in=h_in,
batchSize=batch_size,
output_dim=output_dim)
self.network = PoseRegNet(rng, cfgParams=regressorNetParams)
if regressorNetType == 1:
self.network.load_vgg()
print(self.network)
regressorNetTrainingParams = PoseRegNetTrainingParams()
regressorNetTrainingParams.batch_size = batch_size
regressorNetTrainingParams.learning_rate = learning_rate
if self.scales != []:
scales = [float(s) for s in self.scales.split()]
regressorNetTrainingParams.learning_rate_scales = scales
if self.steps != []:
steps = [int(s) for s in self.steps.split()]
regressorNetTrainingParams.learning_rate_steps = steps
regressorNetTrainingParams.optimizer = optimizer
self.trainer = PoseRegNetTrainer(self.network,
regressorNetTrainingParams, rng)
self.trainer.setup_para(nb_training,
n_chan,
h_in,
w_in,
output_dim,
type='regressor')
self.create_training.pre_create_data()
def train(self):
self.trainer.train(n_epochs=int(self.nb_epoch), storeFilters=True)
def train_para(self):
self.trainer.train_para(int(self.nb_epoch),
self.create_training.create_data,
int(self.nb_process))
def save(self):
if self.network_name is None:
self.network_name += "_model"
self.network_name += str(self.network_model)
self.network_name += "_epoch"
self.network_name += str(self.nb_epoch)
self.network.save(join(self.save_path, self.network_name + ".weight"))
f = file(join(self.save_path, self.network_name + ".cfg"), 'wb')
cPickle.dump(self.network.cfgParams,
f,
protocol=cPickle.HIGHEST_PROTOCOL)
f.close()
def print_type(self):
if int(self.type) == 1:
print('*****************************************************')
print('* REGRESSOR *')
print('*****************************************************')
else:
assert False, 'It is not implemented'