def predict():
nonlocal output_dict
model.eval()
with torch.no_grad():
input_dict_cuda, label_dict_cuda = rhodin_utils_datasets.nestedDictToDevice((input_dict, label_dict), device=device)
output_dict_cuda = model(input_dict_cuda)
output_dict = rhodin_utils_datasets.nestedDictToDevice(output_dict_cuda, device='cpu')
def main():
swap_bg_treadmill()
return
import latent_space_nn as lsn
dataset_path = '../data/pain_no_pain_x2h_intervals_for_extraction_128_128_2fps/'
config_path = 'configs/config_train_rotation_newCal.py'
job_identifier = 'withRotNewCal'
out_dir = '../scratch/rot'
nth_frame = 10
# dataset_path = '../data/pain_no_pain_x2h_intervals_for_extraction_672_380_0.2fps_crop/'
# config_path = 'configs/config_train_rotation_crop_newCal.py'
# job_identifier = 'withRotCropNewCal'
# out_dir = '../scratch/rotcrop'
# nth_frame = 1
util.mkdir(out_dir)
model_num = 50
batch_size_test = 64
output_to_get = ['img_crop']
input_to_get = ['img_crop']
train_subjects = 'brava/herrera/inkasso/julia/kastanjett/naughty_but_nice/sir_holger'.split(
'/')
test_subject = 'aslan'
all_subjects = train_subjects + [test_subject]
config_dict = lsn.set_up_config_dict(config_path, train_subjects,
[test_subject], job_identifier,
batch_size_test, dataset_path,
input_to_get, output_to_get)
model_path = lsn.get_model_path(config_dict, str(model_num))
config_dict['pretrained_network_path'] = model_path
config_dict['every_nth_frame'] = nth_frame
ignite = lsn.IgniteTestNVS()
model, data_loader, config_dict = ignite.run(config_path, config_dict)
data_iterator = iter(data_loader)
for input_dict, label_dict in data_iterator:
# input_dict['bg_crop'] = torch.max(input_dict['img_crop'])*torch.ones(input_dict['img_crop'].size())
input_dict_cuda, label_dict_cuda = rhodin_utils_datasets.nestedDictToDevice(
(input_dict, label_dict), device=device)
output_dict_cuda = model(input_dict_cuda)
utils_plot_batch.plot_iol(input_dict_cuda, label_dict_cuda,
output_dict_cuda, config_dict, 'train',
os.path.join(out_dir, 'lps_true.jpg'))
input_dict['bg_crop'] = 0 * torch.ones(input_dict['img_crop'].size())
input_dict_cuda, label_dict_cuda = rhodin_utils_datasets.nestedDictToDevice(
(input_dict, label_dict), device=device)
output_dict_cuda = model(input_dict_cuda)
utils_plot_batch.plot_iol(input_dict_cuda, label_dict_cuda,
output_dict_cuda, config_dict, 'train',
os.path.join(out_dir, 'lps_white.jpg'))
break
def predict(self, input_dict, label_dict):
self.model.eval()
with torch.no_grad():
input_dict_cuda, label_dict_cuda = rhodin_utils_datasets.nestedDictToDevice(
(input_dict, label_dict), device=device)
output_dict_cuda = self.model(input_dict_cuda)
output_dict = rhodin_utils_datasets.nestedDictToDevice(
output_dict_cuda, device='cpu')
return output_dict
def _inference(engine, batch):
# now compute error
model.eval()
with torch.no_grad():
x, y = utils_data.nestedDictToDevice(
batch, device=device) # make it work for dict input too
if forward_fn is None:
y_pred = model(x)
else:
y_pred = forward_fn(x)
return y_pred, y
def swap_bg_treadmill():
config_path = 'configs/config_pose_debug.py'
train_subjects = ['HOR', 'LAC', 'LAR', 'LAZ', 'LEA', 'LOR', 'PRA']
test_subjects = ['ART']
job_identifier = 'testing_bg_swap'
batch_size_test = 64
dataset_path = '../data'
config_path = 'configs/config_pose_debug.py'
job_identifier = 'withRotCropNewCal'
output_to_get = ['img_crop']
input_to_get = ['img_crop']
train_subjects_model = 'brava/herrera/inkasso/julia/kastanjett/naughty_but_nice/sir_holger'.split(
'/')
test_subjects_model = ['aslan']
epoch_encdec = str(50)
config_file_model = 'configs/config_train_rotation_newCal.py'
job_identifier_encdec = 'withRotNewCal'
config_dict = set_up_config_dict(
config_path, train_subjects, test_subjects, job_identifier,
batch_size_test, dataset_path, input_to_get, output_to_get,
train_subjects_model, test_subjects_model, config_file_model,
job_identifier_encdec, epoch_encdec)
ignite = IgniteTestNVS()
model, data_loader, config_dict = ignite.run(config_dict)
data_iterator = iter(data_loader)
for input_dict, label_dict in data_iterator:
input_dict['bg_crop'] = torch.ones(input_dict['img_crop'].size())
# torch.max(input_dict['img_crop'])
input_dict_cuda, label_dict_cuda = rhodin_utils_datasets.nestedDictToDevice(
(input_dict, label_dict), device=device)
output_dict_cuda = model(input_dict_cuda)
utils_plot_batch.plot_iol(input_dict_cuda, label_dict_cuda,
output_dict_cuda, config_dict, 'train',
'../scratch/try_it_out_samebg.jpg')
break
def _update(engine, batch):
model.train()
# print ('engine.state.iteration',engine.state.iteration)
if not engine.state.iteration % backward_every:
# print ('zero')
optimizer.zero_grad()
x, y = utils_data.nestedDictToDevice(
batch, device=device) # make it work for dict input too
if forward_fn is None:
y_pred = model(x)
else:
y_pred = forward_fn(x)
loss = loss_fn(y_pred, y) / backward_every
loss.backward()
if not engine.state.iteration % backward_every:
# print ('step')
optimizer.step()
return loss.item(), y_pred