def main(args):
logger = Logger(args.output_dir)
args.logger = logger
trainer = Trainer(args)
evaluator = Evaluator(trainer)
for i_epoch in range(0, args.epoch + 1):
# train
log_dict = {
'i_epoch': i_epoch,
'train_losses': [], # per batch
'test_bleus': []
} # per sample
trainer.train_one_epoch(log_dict)
# evaluation and logging
logger.log('%d th epoch' % i_epoch)
evaluator.bleu(log_dict)
evaluator.sample_translation()
log_dict_mean = {
'i_epoch': log_dict['i_epoch'],
'train_loss': np.mean(log_dict['train_losses']),
'test_bleu': np.mean(log_dict['test_bleus'])
}
logger.dump(log_dict_mean)
trainer.save_best(log_dict_mean)
logger.log('-' * 10)
def main(args):
"""
Read config file and run the experiment.
Parameters
----------
config : str
path to config JSON file.
Returns
-------
None
Notes
-----
See configs/example.json
"""
with open(args.config, 'r') as f:
config = json.load(f)
mode = config['mode']
trainer = Trainer(config)
if mode == 'train':
trainer.train()
elif mode == 'test':
trainer.test()
return
def setUp(self):
gameMock = mock.Mock()
memoryMock = mock.Mock()
networkMock = mock.Mock()
competitorMock = mock.Mock()
self.trainer = Trainer(gameMock, memoryMock, networkMock,
competitorMock, 1, 0, 0.5, 0.9, 0.9, 1.1)
def main():
params = vars(
train_parse_args(hyperparameters_default=HYPERPARAMETERS,
paths_default=PATHS))
if not params['disable_cuda'] and torch.cuda.is_available():
params['device'] = torch.device('cuda:0')
else:
params['device'] = torch.device('cpu')
print(torch.cuda.is_available(), params['device'])
loaders = get_data_loaders(imgs_dir=params['imgs_dir'],
labels_filename=params['labels_filename'],
batch_size=params['batch_size'],
n_imgs=params['n_imgs'])
model = ResNet()
#model = SimpleClassifier()
model.cuda()
optimizer = Adam(model.parameters(), lr=3e-4)
criterion = BCELoss()
trainer = Trainer(params, model, optimizer, criterion)
trainer.run(loaders)
def test_XOR_operation(self):
network = Perceptron(input=2, hidden=[20], output=1)
trainer = Trainer(network)
error, epoch = trainer.XOR()
self.assertTrue(error < 0.1)
self.assertTrue(abs(sum(network.activate([0, 0])) - 0) < 0.1)
self.assertTrue(abs(sum(network.activate([0, 1])) - 1) < 0.1)
self.assertTrue(abs(sum(network.activate([1, 0])) - 1) < 0.1)
self.assertTrue(abs(sum(network.activate([1, 1])) - 0) < 0.1)
def main(_):
prepare_dirs(config)
rng = np.random.RandomState(config.random_seed)
tf.set_random_seed(config.random_seed)
model = JMOD2(config)
trainer = Trainer(config, model, rng)
if config.is_train:
trainer.train()
#trainer.resume_training()
else:
trainer.test(showFigure=True)
def test_from_json(self):
return
fn = os.path.join(os.path.dirname(__file__), './test_network.json')
settings = {
'momentum': 0.99,
'epoch': 5000,
'log': 100,
'error': 0.1,
'rate': 0.2,
'cost': Cost.SE
}
with open(fn) as data_file:
network_json = json.load(data_file)
network = Network.from_json(network_json)
trainer = Trainer(network)
start = time.time()
error, epoch = trainer.XOR(settings)
stop = time.time()
print(stop - start)
def run_training(H):
# torch.cuda.is_available = lambda : False
# torch.backends.cudnn.enabled=False
torch.backends.cudnn.deterministic = True
# torch.backends.cudnn.benchmark = True
create_logger(H)
random.seed(H.SEED)
np.random.seed(H.SEED)
torch.manual_seed(H.SEED)
if torch.cuda.is_available():
torch.cuda.manual_seed(H.SEED)
torch.cuda.manual_seed_all(H.SEED)
logger.info("Training start.")
logger.info(repr(H))
train_loader, valid_loader, vocab = create_data_pipelines(H)
logger.info(train_loader.dataset)
logger.info(valid_loader.dataset)
m = Metric([('train_loss', np.inf), ('train_score', np.inf),
('valid_loss', np.inf), ('valid_score', 0), ('train_lr', 0),
('valid_cer', np.inf)])
model = SpeechCNN(len(vocab),
input_size=256,
hidden_size=H.CNN_HIDDEN_SIZE,
dropout=H.CNN_DROPOUT,
initialize=torch_weight_init)
if H.USE_CUDA:
model.cuda()
if H.PRELOAD_MODEL_PATH:
path = os.path.join(H.EXPERIMENT, H.PRELOAD_MODEL_PATH)
state = torch.load(path)
model.load_state_dict(state)
print("Preloaded model: {}".format(path))
criterion = PytorchCTCLoss(vocab)
optimizer = optim.SGD(list(
filter(lambda p: p.requires_grad, model.parameters())),
lr=H.LR,
weight_decay=H.WEIGHT_DECAY,
momentum=H.MOMENTUM,
nesterov=H.NESTEROV)
stopping = Stopping(model, patience=H.STOPPING_PATIENCE)
scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=[H.LR_LAMBDA])
ctc_decoder = CTCGreedyDecoder(vocab)
scorer = Scorer(reduction='sum')
tlogger = TensorboardLogger(root_dir=H.EXPERIMENT,
experiment_dir=H.TIMESTAMP) # PytorchLogger()
checkpoint = Checkpoint(model,
optimizer,
stopping,
m,
root_dir=H.EXPERIMENT,
experiment_dir=H.TIMESTAMP,
restore_from=-1,
interval=H.CHECKPOINT_INTERVAL,
verbose=0)
trainer = Trainer(model, train_loader, optimizer, scheduler, criterion,
ctc_decoder, scorer, H.MAX_GRAD_NORM)
evaluator = Evaluator(model, valid_loader, criterion, ctc_decoder, scorer)
epoch_start = 1
if H.CHECKPOINT_RESTORE:
epoch_start = checkpoint.restore() + 1
train_loader.batch_sampler.shuffle(epoch_start)
epoch = epoch_start
try:
epoch_itr = tlogger.set_itr(range(epoch_start, H.MAX_EPOCHS + 1))
for epoch in epoch_itr:
with DelayedKeyboardInterrupt():
m.train_loss, m.train_score, m.train_lr = trainer(epoch)
m.valid_loss, m.valid_score = evaluator()
if checkpoint:
checkpoint.step(epoch)
stopping_flag = stopping.step(epoch, m.valid_loss,
m.valid_score)
epoch_itr.log_values(m.train_loss, m.train_score, m.train_lr,
m.valid_loss, m.valid_score,
stopping.best_score_epoch,
stopping.best_score)
if stopping_flag:
logger.info(
"Early stopping at epoch: %d, score %f" %
(stopping.best_score_epoch, stopping.best_score))
break
train_loader.batch_sampler.shuffle(epoch)
except KeyboardInterrupt:
logger.info("Training interrupted at: {}".format(epoch))
pass
checkpoint.create(epoch)
model.load_state_dict(stopping.best_score_state)
torch.save(model.state_dict(),
os.path.join(H.EXPERIMENT, H.MODEL_NAME + '.tar'))
logger.info(repr(tlogger))
logger.info(repr(stopping))
logger.info(repr(checkpoint))
logger.info("Training end.")
parser.add_argument('--speed',
type=int,
default=defaultSpeed,
help=speedHelp)
parser.add_argument('inmodel',
type=str,
default=defaultInput,
help=inputHelp)
parser.add_argument('outmodel',
type=str,
default=defaultOutput,
help=outputHelp)
args = parser.parse_args()
# create a trainer
trainer = Trainer(args.inmodel)
# sleep - so keras and tensorflow will allow multiple models...
time.sleep(1)
# create an agent to train
agent = Agent(args.outmodel + '/model-session{}.h5',
maxrecall=args.maxrecall)
# sleep - so keras and tensorflow will allow multiple models...
time.sleep(1)
# Define environment/simulation
sim = Drive(agent.width, agent.height, args.speed)
# train in the environment
def run_training(H):
# torch.cuda.is_available = lambda : False
# torch.backends.cudnn.enabled=False
torch.backends.cudnn.deterministic = True
# torch.backends.cudnn.benchmark = True
create_logger(H)
random.seed(H.SEED)
np.random.seed(H.SEED)
torch.manual_seed(H.SEED)
if torch.cuda.is_available():
torch.cuda.manual_seed(H.SEED)
torch.cuda.manual_seed_all(H.SEED)
logger.info("Training start.")
logger.info(repr(H))
train_loader, valid_loader, vocab = create_data_pipelines(H)
logger.info(train_loader.dataset)
logger.info(valid_loader.dataset)
m = Metric([('train_loss', np.inf), ('train_score', np.inf),
('valid_loss', np.inf), ('valid_score', 0), ('train_lr', 0),
('valid_cer', np.inf)])
model = NeuralSpeechRecognizer(
vocab,
train_loader.dataset.max_seq_length,
rnn_hidden_size=H.RNN_HIDDEN_SIZE,
rnn_num_layers=H.RNN_NUM_LAYERS,
rnn_dropout=H.RNN_DROPOUT,
cnn_dropout=H.CNN_DROPOUT,
teacher_forcing_ratio=H.TEACHER_FORCING_RATIO,
sample_rate=H.AUDIO_SAMPLE_RATE,
window_size=H.SPECT_WINDOW_SIZE,
initialize=torch_weight_init)
if H.USE_CUDA:
model.cuda()
logging.info(model_summary(model, line_length=100))
if H.PRELOAD_MODEL_PATH:
path = os.path.join(H.EXPERIMENT, H.PRELOAD_MODEL_PATH)
state = torch.load(path)
model.load_state_dict(state)
logging.info("Preloaded model: {}".format(path))
criterion = LabelSmoothingLoss(padding_idx=0,
label_smoothing=H.LABEL_SMOOTHING)
sts_decoder = STSDecoder(vocab)
scorer = Scorer()
optimizer = optim.Adam(list(
filter(lambda p: p.requires_grad, model.parameters())),
amsgrad=False,
betas=(0.9, 0.999),
eps=1e-08,
lr=H.LR,
weight_decay=H.WEIGHT_DECAY)
stopping = Stopping(model, patience=H.STOPPING_PATIENCE)
scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=[H.LR_LAMBDA])
tlogger = TensorboardLogger(root_dir=H.EXPERIMENT,
experiment_dir=H.TIMESTAMP) # PytorchLogger()
checkpoint = Checkpoint(model,
optimizer,
stopping,
m,
root_dir=H.EXPERIMENT,
experiment_dir=H.TIMESTAMP,
restore_from=-1,
interval=H.CHECKPOINT_INTERVAL,
verbose=0)
trainer = Trainer(model, train_loader, optimizer, scheduler, criterion,
sts_decoder, scorer, H.MAX_GRAD_NORM)
evaluator = Evaluator(model, valid_loader, criterion, sts_decoder, scorer)
epoch_start = 1
if H.CHECKPOINT_RESTORE:
epoch_start = checkpoint.restore() + 1
train_loader.batch_sampler.shuffle(epoch_start)
epoch = epoch_start
try:
epoch_itr = tlogger.set_itr(range(epoch_start, H.MAX_EPOCHS + 1))
for epoch in epoch_itr:
with DelayedKeyboardInterrupt():
m.train_loss, m.train_score, m.train_lr = trainer(epoch)
m.valid_loss, m.valid_score = evaluator()
if checkpoint:
checkpoint.step(epoch)
stopping_flag = stopping.step(epoch, m.valid_loss,
m.valid_score)
epoch_itr.log_values(m.train_loss, m.train_score, m.train_lr,
m.valid_loss, m.valid_score,
stopping.best_score_epoch,
stopping.best_score)
if stopping_flag:
logger.info(
"Early stopping at epoch: %d, score %f" %
(stopping.best_score_epoch, stopping.best_score))
break
train_loader.batch_sampler.shuffle(epoch)
except KeyboardInterrupt:
logger.info("Training interrupted at: {}".format(epoch))
pass
checkpoint.create(epoch)
model.load_state_dict(stopping.best_score_state)
torch.save(model.state_dict(),
os.path.join(H.EXPERIMENT, H.MODEL_NAME + '.tar'))
logger.info(repr(tlogger))
logger.info(repr(stopping))
logger.info(repr(checkpoint))
logger.info("Training end.")
# train data_sources can not be empty
if ds_train.is_empty:
raise RuntimeError("training dataset can not be empty!")
# evaluation dataset can be empty, yet a warning will be given
ds_eval = DataSource(
_tf_feature_cols=tf_feature_cols, _ds_class="eval", _dsutil=data_utils
)
if ds_eval.is_empty:
ds_eval = DataSource(
_tf_feature_cols=tf_feature_cols,
_ds_class="train",
_dsutil=data_utils,
force_test=True,
)
tf.logging.warn(
"evaluation dataset is empty; will use training dataset for evaluation instead!"
)
_handle = Trainer(
model_fn=resolve_model_type(conf_parser.model_type),
train_data_source=ds_train,
eval_data_source=ds_eval,
hyper_params=conf_parser.model_params,
model_dir=model_dir,
date_time_str=date_time_str,
)
_result = _handle.run()
tf.logging.info(_result)
parser.add_argument('--use-cuda', type=strtobool, default='1')
parser.add_argument('--max-sent-len', type=int, default=30)
parser.add_argument('--data-dir', type=str, default='./')
parser.add_argument('--vocab-size', type=int, default=50000)
parser.add_argument('--embedding-dim', type=int, default=300)
parser.add_argument('--class-n', type=int, default=4)
parser.add_argument('--model', type=str, default='CNN')
parser.add_argument('--epoch', type=int, default=100)
parser.add_argument('--gpu-id', type=int, default=0)
return parser.parse_args()
if __name__ == '__main__':
# args = Args()
args = get_args()
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_id)
print('using GPU id: ', os.environ['CUDA_VISIBLE_DEVICES'])
if args.use_cuda and torch.cuda.is_available():
args.use_cuda = True
else:
args.use_cuda = False
trainer = Trainer(args)
for i_epoch in range(1, args.epoch + 1):
loss = trainer.train_one_epoch(1)
print('%d th epoch: training loss -> %f' % (i_epoch, loss))
loss, acc = trainer.test()
print('test loss: %f, test acc: %f' % (loss, acc))
def main():
memory_bank = Memory(MEMORY_SIZE)
pong_game = Game(GAME_LENGTH, GAME_STEP_TIME)
champion = Network(3,
7,
hidden_layer_size=HIDDEN_LAYER_SIZE,
no_hidden_layers=NO_HIDDEN_LAYERS,
learning_rate=LEARNING_RATE)
competitor = Network(3,
7,
hidden_layer_size=HIDDEN_LAYER_SIZE,
no_hidden_layers=NO_HIDDEN_LAYERS)
trainer = Trainer(pong_game,
memory_bank,
champion,
competitor,
MAX_EPSILON,
MIN_EPSILON,
EPSILON_DECAY,
GAMMA,
RETURNS_DECAY,
WINNERS_GROWTH,
batch_size=BATCH_SIZE)
champion.save_network(DIRECTORY + '/version_' + str(STARTING_VERSION))
for version in range(STARTING_VERSION,
STARTING_VERSION + NUMBER_OF_TRAINING_SESSIONS):
start_time = time.time()
for _ in range(GAMES_PER_TRAINING_SESSION):
print('New game')
trainer.run_game()
trainer.game = Game(GAME_LENGTH, GAME_STEP_TIME)
print("Time taken for training session: ", time.time() - start_time)
champion.save_network(DIRECTORY + '/version_' + str(version + 1))
current_epsilon = trainer.epsilon
current_returns_parameter = trainer.returns_parameter
current_winners_parameter = trainer.winners_parameter
trainer = Trainer(Game(GAME_LENGTH, GAME_STEP_TIME),
memory_bank,
champion,
competitor,
current_epsilon,
MIN_EPSILON,
EPSILON_DECAY,
GAMMA,
RETURNS_DECAY,
WINNERS_GROWTH,
returns_parameter=current_returns_parameter,
winners_parameter=current_winners_parameter,
batch_size=BATCH_SIZE)
test_score = trainer.test_game()
if test_score < 0:
print('Competitor wins, score was ' + str(test_score))
competitor.save_network(DIRECTORY + '/competitor_save')
champion.load_network(DIRECTORY + '/competitor_save')
else:
print('Champion continues, score was ' + str(test_score))
new_competitor_version = random.randint(max(0, version - 5), version)
print('New competitor version: ' + str(new_competitor_version))
competitor.load_network(DIRECTORY + '/version_' +
str(new_competitor_version))
current_epsilon = trainer.epsilon
print('epsilon is ' + str(current_epsilon))
current_returns_parameter = trainer.returns_parameter
current_winners_parameter = trainer.winners_parameter
trainer = Trainer(Game(GAME_LENGTH, GAME_STEP_TIME),
memory_bank,
champion,
competitor,
current_epsilon,
MIN_EPSILON,
EPSILON_DECAY,
GAMMA,
RETURNS_DECAY,
WINNERS_GROWTH,
returns_parameter=current_returns_parameter,
winners_parameter=current_winners_parameter,
batch_size=BATCH_SIZE)
from lib.trainer import Trainer
# データセットのロード
(x_train, t_train), (x_test, t_test) = load_mnist(normalize=True,
flatten=False,
one_hot_label=True)
# ネットワークを定義
network = MyNeuralNet(cost_function=cross_entropy_error)
# レイヤを追加
conv1 = Conv((30, 1, 5, 5))
pool1 = Pooling(2, 2, stride=2)
affine1 = Affine(30 * 12 * 12, 100, "he")
affine2 = Affine(100, 10, "he")
network.add_layer(conv1)
network.add_layer(Relu())
network.add_layer(pool1)
network.add_layer(affine1)
network.add_layer(Relu())
network.add_layer(affine2)
network.add_layer(Softmax())
# 学習
trainer = Trainer(network, x_train, t_train, x_test, t_test)
trainer.train(lr=0.1, epoch_num=20, batch_size=100, for_cnn=True)
trainer.savefig("simple_cnn", "../data/simple_cnn.png")
from lib.trainer import Trainer
from lib.options import parse_options
# Set logger display format
log.basicConfig(format='[%(asctime)s] [INFO] %(message)s',
datefmt='%d/%m %H:%M:%S',
level=log.INFO)
if __name__ == "__main__":
"""Main program."""
args, args_str = parse_options()
log.info(f'Parameters: \n{args_str}')
if args.analytic:
for filename in glob.glob(os.path.join(args.glsl_path,
'*/*.glsl'))[::-1]:
name = os.path.basename(filename).split('.')[0]
if args.exclude and name in args.exclude:
continue
if args.include and name not in args.include:
continue
log.info(f'Training {name}')
model = Trainer(args, args_str, name)
model.train()
else:
log.info(f'Training on {args.dataset_path}')
model = Trainer(args, args_str, 'mesh')
model.train()
def train_one(model_folder):
new_model_folder_name = model_folder.replace('_created', '_training')
os_rename(model_folder, new_model_folder_name)
frequencies = glob(os_path_join(new_model_folder_name, 'k_*'))
for frequency in frequencies:
# Load model
print('train.py: training {}'.format(frequency))
model_params_path = os_path_join(frequency, model_params_fname)
# create model
model, model_params = get_which_model_from_params_fname(
model_params_path, return_params=True)
if 'cuda' in model_params:
using_cuda = model_params['cuda'] and torch_cuda_is_available()
else:
using_cuda = torch_cuda_is_available()
if using_cuda is True:
model.cuda()
# save initial weights
if 'save_initial' in model_params and model_params[
'save_initial'] and model_params['save_dir']:
suffix = '_initial'
path = add_suffix_to_path(model_params_fname['save_dir'], suffix) # pylint: disable=E1126
ensure_dir(path)
torch_save(model.state_dict(),
os_path_join(path, MODEL_DATA_FNAME))
save_model_params(os_path_join(path, model_params_fname),
model_params)
# loss
if 'cost_function' in model_params:
loss = model_params['cost_function']
elif 'loss_function' in model_params:
loss = model_params['loss_function']
else:
raise ValueError(
'model_params missing key cost_function or loss_function')
if loss not in ['MSE', 'L1', 'SmoothL1']:
raise TypeError('Error must be MSE, L1, or SmoothL1. You gave ' +
str(loss))
if loss == 'MSE':
from torch.nn import MSELoss
loss = MSELoss()
elif loss == 'L1':
from torch.nn import L1Loss
loss = L1Loss()
elif loss == 'SmoothL1':
from torch.nn import SmoothL1Loss
loss = SmoothL1Loss()
# optimizer
if model_params['optimizer'] == 'Adam':
from torch.optim import Adam
optimizer = Adam(model.parameters(),
lr=model_params['learning_rate'],
weight_decay=model_params['weight_decay'])
elif model_params['optimizer'] == 'SGD':
from torch.optim import SGD
optimizer = SGD(model.parameters(),
lr=model_params['learning_rate'],
momentum=model_params['momentum'],
weight_decay=model_params['weight_decay'])
else:
raise ValueError(
'model_params[\'optimizer\'] must be either Adam or SGD. Got '
+ model_params['optimizer'])
logger = Logger()
# Load training, validation, and test data
# Load primary training data
dat_train = ApertureDataset(
model_params['data_train'],
NUM_SAMPLES_TRAIN,
k=model_params['k'],
target_is_data=model_params['data_is_target'])
loader_train = DataLoader(dat_train,
batch_size=model_params['batch_size'],
shuffle=True,
num_workers=DATALOADER_NUM_WORKERS,
pin_memory=using_cuda)
# Load secondary training data - used to evaluate training loss after every epoch
dat_train2 = ApertureDataset(
model_params['data_train'],
NUM_SAMPLES_TRAIN_EVAL,
k=model_params['k'],
target_is_data=model_params['data_is_target'])
loader_train_eval = DataLoader(dat_train2,
batch_size=model_params['batch_size'],
shuffle=False,
num_workers=DATALOADER_NUM_WORKERS,
pin_memory=using_cuda)
# Load validation data - used to evaluate validation loss after every epoch
dat_val = ApertureDataset(
model_params['data_val'],
NUM_SAMPLES_VALID,
k=model_params['k'],
target_is_data=model_params['data_is_target'])
loader_val = DataLoader(dat_val,
batch_size=model_params['batch_size'],
shuffle=False,
num_workers=DATALOADER_NUM_WORKERS,
pin_memory=using_cuda)
trainer = Trainer(
model=model,
loss=loss,
optimizer=optimizer,
patience=model_params['patience'],
loader_train=loader_train,
loader_train_eval=loader_train_eval,
loader_val=loader_val,
cuda=using_cuda,
logger=logger,
data_noise_gaussian=model_params['data_noise_gaussian'],
save_dir=frequency)
# run training
trainer.train()
os_rename(new_model_folder_name,
new_model_folder_name.replace('_training', '_trained'))
embeddings_weights[
train_dataset.vocabulary[word]] = np.random.uniform(
-0.5, 0.5)
with open(params.processed_embeddings_path, 'w') as outfile:
json.dump(embeddings_weights.tolist(), outfile)
print("Embedding weights saved!")
print("Out of %d total words, %d are found in the embedding" %
(len(train_dataset.vocabulary), words_missing))
else:
with open(params.processed_embeddings_path) as json_file:
data = json.load(json_file)
embeddings_weights = torch.Tensor(data)
print("Embedding weights loaded!")
train_loader = DataLoader(train_dataset, batch_size=256)
valid_loader = DataLoader(valid_dataset, batch_size=256)
test_loader = DataLoader(test_dataset, batch_size=256)
nermodel = NERModel(len(train_dataset.vocabulary),
len(train_dataset.label_vocabulary),
params).to(torch.device(params.device))
if (params.embeddings_path != None):
nermodel.word_embedder.weight.data.copy_(torch.Tensor(embeddings_weights))
trainer = Trainer(model=nermodel,
loss_function=nn.CrossEntropyLoss(
ignore_index=train_dataset.label_vocabulary["<pad>"]),
optimizer=optim.Adam(params=nermodel.parameters()))
trainer.train(train_loader, valid_loader, epochs=1000)
logger = Logger(opt)
model = TensorMask(backbone=opt.backbone , num_cls=opt.num_class ,
base_window= opt.base_window ,
freezeBN=opt.frezeBN,freezeLayers=opt.frezeLayer,
align_corners= opt.align_corners)
optimizer = optimer.SGD([{'params':filter(lambda x:len(x.size()) == 4 ,model.parameters()),'weight_decay':0.0001 },
{'params': filter(lambda x:len(x.size()) <4,model.parameters())}],
lr=opt.lr,warm_up=1000,momentum=0.9,nesterov=True)
start_epoch = 0
if opt.weights != '' :
model, optimizer, start_epoch = load_model(
model, opt.weights, optimizer, opt.resume, opt.lr, opt.lr_step)
trainer = Trainer(opt,model,optimizer)
trainer.set_device(opt.gpus,opt.device)
print('Setting up data...')
val_loader = torch.utils.data.DataLoader(
COCO(cfg=opt, split='val',augment=False),
batch_size=8,
shuffle=False,
num_workers=8,
pin_memory=True
)
train_loader = torch.utils.data.DataLoader(
COCO(cfg=opt, split='train',augment=True),
batch_size=opt.batch_size,
shuffle=True,
num_workers=8,
benchmark_set = ThreeDMatchDownsampled(info_benchmark,config, data_augmentation=False)
config.train_loader, neighborhood_limits = get_dataloader(dataset=train_set,
batch_size=config.batch_size,
shuffle=True,
num_workers=config.num_workers,
)
config.val_loader, _ = get_dataloader(dataset=val_set,
batch_size=config.batch_size,
shuffle=False,
num_workers=1,
neighborhood_limits=neighborhood_limits
)
config.test_loader, _ = get_dataloader(dataset=benchmark_set,
batch_size=config.batch_size,
shuffle=False,
num_workers=1,
neighborhood_limits=neighborhood_limits)
# create evaluation metrics
config.desc_loss = MetricLoss(config)
# start to train our model
trainer = Trainer(config)
if(config.mode=='train'):
trainer.train()
elif(config.mode =='val'):
trainer.eval()
else:
trainer.test()
def main():
args = parse_args()
update_config(args.cfg_file)
if args.gpus:
config.GPUS = args.gpus
else:
config.CUDA = False
if args.workers:
config.WORKERS = args.workers
print('Using config:')
pprint.pprint(config)
if args.manualSeed is None:
args.manualSeed = random.randint(1, 10000)
random.seed(args.manualSeed)
np.random.seed(args.manualSeed)
torch.manual_seed(args.manualSeed)
if config.CUDA:
torch.cuda.manual_seed_all(args.manualSeed)
torch.backends.cudnn.benchmark = config.CUDNN.BENCHMARK
torch.backends.cudnn.deterministic = config.CUDNN.DETERMINISTIC
torch.backends.cudnn.enabled = config.CUDNN.ENABLED
if config.CUDA:
os.environ["CUDA_VISIBLE_DEVICES"] = config.GPUS
device = torch.device('cuda' if config.CUDA else 'cpu')
# Redirect print to both console and log file
sys.stdout = Logger(osp.join(config.OUTPUT_DIR, 'log.txt'))
# Create data loaders
dataset = DataSet(config.DATASET.ROOT, config.DATASET.DATASET)
normalizer = T.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_transformer = T.Compose([
T.RandomSizedRectCrop(*config.MODEL.IMAGE_SIZE),
T.RandomHorizontalFlip(),
T.RandomRotation(10),
T.ColorJitter(0.2, 0.2, 0.2),
T.ToTensor(),
normalizer,
T.RandomErasing(EPSILON=config.DATASET.RE),
])
test_transformer = T.Compose([
T.Resize(config.MODEL.IMAGE_SIZE, interpolation=3),
T.ToTensor(),
normalizer,
])
train_loader = DataLoader(UnsupervisedCamStylePreprocessor(
dataset.train,
root=osp.join(dataset.images_dir, dataset.train_path),
camstyle_root=osp.join(dataset.images_dir,
dataset.train_camstyle_path),
num_cam=dataset.num_cam,
use_gan=True,
transform=train_transformer),
batch_size=config.TRAIN.BATCH_SIZE,
num_workers=config.WORKERS,
shuffle=config.TRAIN.SHUFFLE,
pin_memory=True,
drop_last=False)
query_loader = DataLoader(Preprocessor(dataset.query,
root=osp.join(
dataset.images_dir,
dataset.query_path),
transform=test_transformer),
batch_size=config.TEST.BATCH_SIZE,
num_workers=config.WORKERS,
shuffle=False,
pin_memory=True)
gallery_loader = DataLoader(Preprocessor(dataset.gallery,
root=osp.join(
dataset.images_dir,
dataset.gallery_path),
transform=test_transformer),
batch_size=config.TEST.BATCH_SIZE,
num_workers=config.WORKERS,
shuffle=False,
pin_memory=True)
# Create model
model = models.create(config.MODEL.NAME,
pretrained=config.MODEL.PRETRAINED,
num_classes=dataset.num_train_ids)
# Memory Network
num_tgt = len(dataset.train)
memory = models.create('memory', config.MODEL.FEATURES, num_tgt)
# Load from checkpoint
if config.TRAIN.RESUME:
checkpoint = load_checkpoint(config.TRAIN.CHECKPOINT)
model.load_state_dict(checkpoint['state_dict'], strict=False)
memory.load_state_dict(checkpoint['state_dict_memory'], strict=False)
print("=> Start epoch {} ".format(checkpoint['epoch']))
# Set model
model = nn.DataParallel(model).to(device)
memory = memory.to(device)
# Optimizer
base_param_ids = set(map(id, model.module.base.parameters()))
base_params_need_for_grad = filter(lambda p: p.requires_grad,
model.module.base.parameters())
new_params = [p for p in model.parameters() if id(p) not in base_param_ids]
param_groups = [{
'params': base_params_need_for_grad,
'lr_mult': 0.1
}, {
'params': new_params,
'lr_mult': 1.0
}]
optimizer = get_optimizer(config, param_groups)
# Trainer
trainer = Trainer(config, model, memory)
def adjust_lr(epoch):
step_size = config.TRAIN.LR_STEP
lr = config.TRAIN.LR * (config.TRAIN.LR_FACTOR**(epoch // step_size))
for g in optimizer.param_groups:
g['lr'] = lr * g.get('lr_mult', 1)
best_r1 = 0.0
# Start training
for epoch in range(config.TRAIN.BEGIN_EPOCH, config.TRAIN.END_EPOCH):
# lr_scheduler.step()
adjust_lr(epoch)
trainer.train(epoch, train_loader, optimizer)
print('Test with latest model:')
evaluator = Evaluator(model)
r1 = evaluator.evaluate(query_loader, gallery_loader, dataset.query,
dataset.gallery, config.TEST.OUTPUT_FEATURES)
if r1 > best_r1:
best_r1 = r1
save_checkpoint(
{
'state_dict': model.module.state_dict(),
'state_dict_memory': memory.state_dict(),
'epoch': epoch + 1,
},
fpath=osp.join(config.OUTPUT_DIR, 'checkpoint.pth.tar'))
print('\n * Finished epoch {:3d} \n'.format(epoch))
# Final test
print('Test with best model:')
evaluator = Evaluator(model)
checkpoint = load_checkpoint(
osp.join(config.OUTPUT_DIR, 'checkpoint.pth.tar'))
print('best model at epoch: {}'.format(checkpoint['epoch']))
model.module.load_state_dict(checkpoint['state_dict'])
evaluator.evaluate(query_loader, gallery_loader, dataset.query,
dataset.gallery, config.TEST.OUTPUT_FEATURES)