def main():
args = get_args()
config = process_config(args.config)
# create the experiments dirs
create_dirs([
config["callbacks"]["tensorboard_log_dir"],
config["callbacks"]["checkpoint_dir"]
])
# resolve classes
data_loader_cls = factory("data_loaders.{}".format(
config["data_loader"]["name"]))
model_cls = factory("models.{}".format(config["model"]["name"]))
trainer_cls = factory("trainers.{}".format(config["trainer"]["name"]))
# create data generators for the data sets
loader_params = config["data_loader"]
training_generator = data_loader_cls("train",
shuffle=True,
**loader_params)
eval_generator = data_loader_cls("eval", **loader_params)
# train ze model
model_params = config['model']
print(training_generator.feature_dim)
model = model_cls(training_generator.feature_dim,
training_generator.n_classes, **model_params)
trainer = trainer_cls(config["exp"]["name"], model.model,
config["callbacks"], **config["trainer"])
trainer.train(training_generator, eval_generator)
def main():
args = parse_args()
config = process_config(args.config)
# create the experiments dirs
create_dirs([config["callbacks"]["tensorboard_log_dir"], config["callbacks"]["checkpoint_dir"]])
# resolve classes
data_loader_cls = factory("data_loaders.{}".format(config["data_loader"]["name"]))
trainer_cls = factory("trainers.{}".format(config["trainer"]["name"]))
# create data generators for the data sets
loader_params = config["data_loader"]
generators = {
name: data_loader_cls(name, **loader_params)
for name in ['train', 'eval', 'unlabeled']
}
# train ze model(s)
if args.tri_training:
model = list()
for i in range(1, 4):
print('Creating model {}/3'.format(i))
model_params = config['model{}'.format(i)]
model_cls = factory("models.{}".format(config["model{}".format(i)]["name"]))
model.append(model_cls(generators['train'].feature_dim, generators['train'].n_classes, **model_params))
else:
model_params = config['model']
model_cls = factory("models.{}".format(config["model"]["name"]))
model = model_cls(generators['train'].feature_dim, generators['train'].n_classes, **model_params)
trainer = trainer_cls(
config["exp"]["name"],
model,
config["callbacks"],
**config["trainer"]
)
if args.mode == 'train':
trainer.train(
generators['train'],
generators['eval'],
# generators['unlabeled']
# confidence_threshold=0
)
elif args.mode == 'eval':
evaluator = DCASE2019T4Evaluator(
trainer.predict,
generators['eval'],
verbose=True
)
ret = evaluator.evaluate(evaluator.find_class_thresholds())
else:
# mode == 'predict'
for data_type, fns in config['trainer']['prediction'].items():
print('Predicting on {} data and storing at {}.'.format(data_type, fns['predictions_fn']))
trainer.predict(
generators[data_type],
write_fn=path.join('experiments', config['exp']['name'], fns['predictions_fn'])
)
#q_dict = Dictionary.load_from_file('data/question_dictionary.pkl')
#c_dict = Dictionary.load_from_file('data/caption_dictionary.pkl')
q_dict = Dictionary.load_from_file('data/VQAE/question_dictionary.pkl')
c_dict = Dictionary.load_from_file('data/VQAE/explain_dictionary.pkl')
#train_dset = VQAFeatureDataset('train', q_dict, c_dict, args.att_thr)
#eval_dset = VQAFeatureDataset('val', q_dict, c_dict, args.att_thr)
train_dset = VQAEDataset('train', q_dict, c_dict, 'cache/VQAE2')
eval_dset = VQAEDataset('val', q_dict, c_dict, 'cache/VQAE2')
#train_dset = VQAEVQA2Dataset('train', q_dict, c_dict, 'cache')
#eval_dset = VQAEVQA2Dataset('val', q_dict, c_dict, 'cache')
batch_size = args.batch_size
constructor = 'build_%s' % args.model
model = utils.factory(constructor, train_dset, args.num_hid, args.att_dim,
args.decode_dim).cuda()
model_path = os.path.join(args.output, 'model.pth')
model_state = torch.load(model_path)
model.load_state_dict(model_state)
print('Model has {} parameters in total'.format(utils.params_count(model)))
#model = nn.DataParallel(model).cuda()
eval_loader = DataLoader(eval_dset,
batch_size,
shuffle=False,
num_workers=1)
model.train(False)
vqa_score, results = evaluate(model, eval_loader, q_dict, c_dict)
save_obj = {'vqa_score': vqa_score, 'results': results}
def trainval():
pass
def deploy():
pass
def net():
pass
if __name__ == '__main__':
n = factory('unet')
h = cfgs.inShape[0]
w = cfgs.inShape[1]
batch_size = 32
n_filter = 32
n.Data("/home/yihuihe/Ultrasound-Nerve-Segmentation/data/train.txt",
backend='image',
label="aaa",
new_height=h,
new_width=w,
scale=0.01578412369702059,
batch_size=batch_size)
n.Deconvolution('up1', )
n.Data("data/mask.txt",
# by yihui
import caffe
from utils import factory
import cfgs_res as cfgs
def trainval():
pass
def deploy():
pass
def net():
pass
if __name__ == '__main__':
n=factory('unet')
h=cfgs.inShape[0]
w=cfgs.inShape[1]
batch_size=32
n_filter=32
n.Data("/home/yihuihe/Ultrasound-Nerve-Segmentation/data/train.txt",backend='image',label="aaa",new_height=h,new_width=w,scale=0.01578412369702059, batch_size=batch_size)
n.Deconvolution('up1',)
n.Data("data/mask.txt",name='label',mean_file="data/mask_mean.binaryproto",backend='image',new_height=h,new_width=w,batch_size=batch_size, scale=1./255)
n.silence('nothing','aaa')
n.diceLoss('prob','label')
n.totxt('resnet/addition.prototxt')
def trainval():
pass
def deploy():
pass
def net():
pass
if __name__ == "__main__":
n = factory("unet")
h = cfgs.inShape[0]
w = cfgs.inShape[1]
batch_size = 24
omit = 0
layers = 5
if use_lmdb:
n.Data(
"/mnt/data1/yihuihe/ultrasound-nerve/lmdb_train_val/train_data", # mean_file='data/data_mean.binaryproto',
mean_value=[128],
scale=1.0 / 255,
backend="LMDB",
)
else:
n.Data(
"data/val_mask.txt",