def eval_baseline_h36m():
weights_path = sys.args[1]
eval_model = MultiBranchModel(
dim=3, n_joints=17, nb_pose_blocks=1) # TODO retrieve from config
eval_model.load_weights(weights_path, pose_only=True)
# local loading
local_h36m_path = '/home/caleml/datasets/h36m'
local_h36m = Human36M(local_h36m_path,
dataconf=config.human36m_dataconf,
poselayout=pose_format.pa17j3d,
topology='frames')
h36m_val = BatchLoader(local_h36m, ['frame'],
['pose_w', 'pose_uvd', 'afmat', 'camera'],
VALID_MODE,
batch_size=local_h36m.get_length(VALID_MODE),
shuffle=True)
log.printcn(log.OKBLUE, 'Preloading Human3.6M validation samples...')
[x_val], [pw_val, puvd_val, afmat_val, scam_val] = h36m_val[0]
scores = eval_human36m_sc_error(eval_model.model,
x_val,
pw_val,
afmat_val,
puvd_val[:, 0, 2],
scam_val,
batch_size=24)
pprint(scores)
def eval_h36m(exp_folder, pose_only):
# local loading
local_h36m_path = '/home/caleml/datasets/h36m'
local_h36m = Human36M(local_h36m_path,
dataconf=config.human36m_dataconf,
poselayout=pose_format.pa17j3d,
topology='frames')
h36m_val = BatchLoader(local_h36m, ['frame'],
['pose_w', 'pose_uvd', 'afmat', 'camera'],
VALID_MODE,
batch_size=local_h36m.get_length(VALID_MODE),
shuffle=True)
log.printcn(log.OKBLUE, 'Preloading Human3.6M validation samples...')
[x_val], [pw_val, puvd_val, afmat_val, scam_val] = h36m_val[0]
dataset = {
'x_val': x_val,
'pw_val': pw_val,
'puvd_val': puvd_val,
'afmat_val': afmat_val,
'scam_val': scam_val
}
all_scores_baseline = eval_folder(exp_folder, dataset, pose_only=pose_only)
print(all_scores_baseline)
def launch_training(dataset_path, model_folder, n_epochs, batch_size,
pose_blocks):
# loading dataset
h36m_path = dataset_path
h36m = Human36M(h36m_path,
dataconf=config.human36m_dataconf,
poselayout=pose_format.pa17j3d,
topology='frames')
data_tr_h36m = BatchLoader(h36m, ['frame'], ['pose'] * pose_blocks,
TRAIN_MODE,
batch_size=batch_size,
shuffle=True)
# validation
h36m_val = BatchLoader(h36m, ['frame'],
['pose_w', 'pose_uvd', 'afmat', 'camera'],
VALID_MODE,
batch_size=h36m.get_length(VALID_MODE),
shuffle=True)
log.printcn(log.OKBLUE, 'Preloading Human3.6M validation samples...')
[x_val], [pw_val, puvd_val, afmat_val, scam_val] = h36m_val[0]
eval_callback = callbacks.H36MEvalCallback(pose_blocks,
x_val,
pw_val,
afmat_val,
puvd_val[:, 0, 2],
scam_val,
pose_only=True,
logdir=model_folder)
# model
model = MultiBranchModel(dim=3, n_joints=17, nb_pose_blocks=pose_blocks)
model.build_pose_only()
model.add_callback(eval_callback)
model.train(data_tr_h36m,
steps_per_epoch=len(data_tr_h36m),
model_folder=model_folder,
n_epochs=n_epochs)
def launch_training(dataset_path, model_folder, n_epochs, batch_size):
mpii_path = dataset_path
mpii = MpiiSinglePerson(mpii_path, dataconf=config.mpii_dataconf)
data_tr_mpii = BatchLoader(mpii,
['frame'],
['frame', 'pose', 'pose', 'pose', 'pose'],
TRAIN_MODE,
batch_size=16,
shuffle=False)
model = MultiBranchModel(nb_pose_blocks=4)
model.build()
model.train(data_tr_mpii, steps_per_epoch=len(data_tr_mpii), model_folder=model_folder, n_epochs=n_epochs)
def launch_training(dataset_path, model_folder, n_epochs, batch_size, pose_blocks):
h36m_path = dataset_path
h36m = Human36M(h36m_path, dataconf=config.human36m_dataconf, poselayout=pose_format.pa17j3d, topology='frames')
data_tr_h36m = BatchLoader(
h36m,
['frame'],
['frame'] + ['pose'] * pose_blocks + ['action'] * 3,
TRAIN_MODE,
batch_size=batch_size,
shuffle=True)
model = CycleModel(dim=3, n_joints=17, nb_pose_blocks=pose_blocks)
model.build()
model.train(data_tr_h36m, steps_per_epoch=len(data_tr_h36m), model_folder=model_folder, n_epochs=n_epochs)
def launch_training(dataset_path, model_folder, n_epochs):
mpii_path = dataset_path
mpii = MpiiSinglePerson(mpii_path,
dataconf=config.mpii_dataconf,
poselayout=pose_format.pa17j3d)
data_tr_mpii = BatchLoader(mpii, ['frame'], ['frame'],
TRAIN_MODE,
shuffle=False)
model = AppearanceModel()
model.build()
model.train(data_tr_mpii,
steps_per_epoch=len(data_tr_mpii),
model_folder=model_folder,
n_epochs=n_epochs)
def launch(self):
'''
main entrypoint
1. training data
2. validation data
3. callbacks
4. model building
5. launch actual training
'''
h36m = Human36M(self.dataset_path,
dataconf=config.human36m_dataconf,
poselayout=pose_format.pa17j3d,
topology='frames')
# training data
dataset_output = self.get_h36m_outputs()
data_tr_h36m = BatchLoader(h36m, ['frame'],
dataset_output,
TRAIN_MODE,
batch_size=self.batch_size,
shuffle=True)
# validation data
h36m_val = BatchLoader(h36m, ['frame'],
['pose_w', 'pose_uvd', 'afmat', 'camera'],
VALID_MODE,
batch_size=h36m.get_length(VALID_MODE),
shuffle=True)
log.printcn(log.OKBLUE, 'Preloading Human3.6M validation samples...')
[x_val], [pw_val, puvd_val, afmat_val, scam_val] = h36m_val[0]
# assert rootz == puvd_val[:,0,2]
# callbacks
cb_list = list()
eval_callback = callbacks.H36MEvalCallback(self.pose_blocks,
x_val,
pw_val,
afmat_val,
puvd_val[:, 0, 2],
scam_val,
pose_only=self.pose_only,
logdir=self.model_folder)
logs_folder = os.environ[
'HOME'] + '/pe_experiments/tensorboard/' + self.model_folder.split(
'/')[-1]
print('Tensorboard log folder %s' % logs_folder)
tensorboard = TensorBoard(
log_dir=os.path.join(logs_folder, 'tensorboard'))
cb_list.append(tensorboard)
cb_list.append(eval_callback)
cb_list.append(LearningRateScheduler(lr_scheduler))
cb_list.append(callbacks.SaveModel(self.model_folder))
# model
self.build_model()
# train
self.model.train(data_tr_h36m,
steps_per_epoch=len(data_tr_h36m),
model_folder=self.model_folder,
n_epochs=self.n_epochs,
cb_list=cb_list,
n_workers=self.nb_workers)