def run(config, folds_dir, balanced):
model = get_model(config[MODEL_NAME], config[MODEL_PARAMS]).cuda()
criterion = get_loss(config[LOSS_NAME], config[LOSS_PARAMS])
optimizer = get_optimizer(config[OPTIM_NAME],
model.parameters(),
optimizer_params=config[OPTIM_PARAMS])
last_epoch = -1
scheduler = get_scheduler(config[SCHEDULER_NAME], optimizer, last_epoch,
config[SCHEDULER_PARAMS])
datasets = {
stage: CustomDataset(folds_dir, stage, config[FOLD_ID],
config[DATA_PREFIX], config[INPUT_SIZE])
for stage in ['train', 'test']
}
print('Loading sampler')
if balanced:
train_sampler = BalancedBatchSampler(datasets['train'])
else:
train_sampler = None
print('Sampler loaded')
dataloaders = {
stage: get_dataloader(datasets[stage], config[BATCH_SIZE],
train_sampler)
for stage in ['train', 'test']
}
writer = SummaryWriter(config[TRAIN_DIR])
train(config, model, dataloaders, criterion, optimizer, scheduler, writer,
last_epoch + 1)
def _init_train_components_ensemble(self, reinitialise=False):
self.val_metrics = {
'loss': metrics.Loss(BCEAndJaccardLoss(eval_ensemble=True, gpu_node=self.config.gpu_node)),
'segment_metrics': SegmentationMetrics(num_classes=self.data_loaders.num_classes,
threshold=self.config.binarize_threshold,
eval_ensemble=True)
}
self.model_cfg.network_params.input_channels = self.data_loaders.input_channels
self.model_cfg.network_params.num_classes = self.data_loaders.num_classes
self.model_cfg.network_params.image_size = self.data_loaders.image_size
self.criterion = self._init_criterion(not reinitialise)
self.ens_models = list()
self.ens_optimizers = list()
self.ens_lr_schedulers = list()
optimizer_cls = get_optimizer(self.optim_cfg)
init_param_names = retrieve_class_init_parameters(optimizer_cls)
optimizer_params = {k: v for k, v in self.optim_cfg.items() if k in init_param_names}
for _ in range(self.len_models):
self.ens_models.append(get_model(self.model_cfg).to(device=self.device))
self.ens_optimizers.append(optimizer_cls(self.ens_models[-1].parameters(), **optimizer_params))
lr_scheduler = self._init_lr_scheduler(self.ens_optimizers[-1])
self.ens_lr_schedulers.append(lr_scheduler)
if not reinitialise:
self.main_logger.info(f'Using ensemble of {self.len_models} {self.ens_models[0]}')
self.main_logger.info(f'Using optimizers {self.ens_optimizers[0].__class__.__name__}')
self.trainer, self.evaluator = self._init_engines()
self._init_handlers()
def get_test_max_landmark_of_one_model(config, gi, best_model_idx, key_group):
test_img_list = gen_test_csv()
print('test_img_list ', len(test_img_list))
# result_set_whole = {}
# for img_id in test_img_list:
#初始化添加
# result_set_whole[img_id] = {}
test_data_set = get_test_loader(config, test_img_list,
get_transform(config, 'val'))
model = get_model(config, gi)
if torch.cuda.is_available():
model = model.cuda()
optimizer = get_optimizer(config, model.parameters())
checkpoint = utils.checkpoint.get_model_saved(config, gi, best_model_idx)
best_epoch, step = utils.checkpoint.load_checkpoint(
model, optimizer, checkpoint)
result_set = test_one_model(test_data_set, model, key_group)
#
result_list_whole = []
for img_ps in result_set.keys():
ps = result_set[img_ps]
max_p_key = max(ps, key=ps.get)
# result_set_whole[img_ps][max_p_key] = ps[max_p_key]
result_list_whole.append((img_ps, max_p_key, ps[max_p_key]))
test_pd = pd.DataFrame.from_records(
result_list_whole, columns=['img_id', 'landmark_id', 'pers'])
output_filename = os.path.join('./results/test/',
'test_img_land_' + str(gi) + '.csv')
test_pd.to_csv(output_filename, index=False)
return
def build_model(self):
"""Creates the model object, which is used to calculate topics, similar words, similar sentences, topic occurrences, and topic similarities
Returns:
model: Keras model object
"""
optimizer = get_optimizer(self.args)
self.logger.info('Building model')
self.logger.info(' Number of training examples: %d',
len(self.train_x))
self.logger.info(' Length of vocab: %d', len(self.vocab))
def max_margin_loss(y_true, y_pred):
return K.mean(y_pred)
model = create_model(self.args, self.overall_maxlen, self.vocab)
# Freeze the word embedding layer
model.get_layer('word_emb').trainable = False
# Check option to fix clusters instead of training them
if self.args.fix_clusters == "yes":
model.get_layer('topic_emb').trainable = False
model.compile(optimizer=optimizer,
loss=max_margin_loss,
metrics=[max_margin_loss])
return model
def __init__(self, args, logger):
self.args = args
self.logger = logger
self.writer = SummaryWriter(args.log_dir)
cudnn.enabled = True
# set up model
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.model = get_aux_net(args.network.arch)(aux_classes=args.aux_classes + 1, classes=args.n_classes)
self.model = self.model.to(self.device)
wandb.watch(self.model)
if args.mode == 'train':
# set up optimizer, lr scheduler and loss functions
optimizer = get_optimizer(self.args.training.optimizer)
optimizer_params = {k: v for k, v in self.args.training.optimizer.items() if k != "name"}
self.optimizer = optimizer(self.model.parameters(), **optimizer_params)
self.scheduler = get_scheduler(self.optimizer, self.args.training.lr_scheduler)
self.class_loss_func = nn.CrossEntropyLoss()
self.start_iter = 0
# resume
if args.training.resume:
self.load(args.model_dir + '/' + args.training.resume)
cudnn.benchmark = True
elif args.mode == 'val':
self.load(os.path.join(args.model_dir, args.validation.model))
else:
self.load(os.path.join(args.model_dir, args.testing.model))
def run(config):
train_dir = config.train.dir
model = get_model(config, model_type).to(device)
print('The nubmer of parameters : %d' % count_parameters(model))
criterion = get_loss(config)
optimizer = get_optimizer(config, model)
checkpoint = utils.checkpoint.get_initial_checkpoint(config,
model_type=model_type)
if checkpoint is not None:
last_epoch, step = utils.checkpoint.load_checkpoint(
model, optimizer, checkpoint, model_type=model_type)
else:
last_epoch, step = -1, -1
print('from checkpoint: {} last epoch:{}'.format(checkpoint, last_epoch))
scheduler = get_scheduler(config, optimizer, last_epoch)
print(config.data)
dataloaders = {
'train': get_train_dataloader(config),
'val': get_valid_dataloader(config),
'test': get_test_dataloader(config)
}
writer = SummaryWriter(config.train[model_type + '_dir'])
visualizer = get_visualizer(config)
train(config, model, dataloaders, criterion, optimizer, scheduler, writer,
visualizer, last_epoch + 1)
def run(config):
teacher_model = get_model(config, 'teacher').to(device)
criterion = get_loss(config)
# for teacher
trainable_params = filter(lambda p: p.requires_grad,
teacher_model.parameters())
optimizer_t = get_optimizer(config, teacher_model.parameters())
checkpoint_t = utils.checkpoint.get_initial_checkpoint(
config, model_type='teacher')
if checkpoint_t is not None:
last_epoch_t, step_t = utils.checkpoint.load_checkpoint(
teacher_model, optimizer_t, checkpoint_t, model_type='teacher')
else:
last_epoch_t, step_t = -1, -1
print('teacher model from checkpoint: {} last epoch:{}'.format(
checkpoint_t, last_epoch_t))
scheduler_t = get_scheduler(config, optimizer_t, last_epoch_t)
print(config.data)
dataloaders = {
'train': get_train_dataloader(config),
'val': get_valid_dataloader(config),
'test': get_test_dataloader(config)
}
writer = SummaryWriter(config.train['teacher' + '_dir'])
visualizer = get_visualizer(config)
train(config, teacher_model, dataloaders, criterion, optimizer_t,
scheduler_t, writer, visualizer, last_epoch_t + 1)
def run(config):
train_dir = config.train.dir
model = get_model(config).cuda()
criterion = get_loss(config)
optimizer = get_optimizer(config, model.parameters())
checkpoint = utils.checkpoint.get_initial_checkpoint(config)
if checkpoint is not None:
last_epoch, step = utils.checkpoint.load_checkpoint(
model, optimizer, checkpoint)
else:
last_epoch, step = -1, -1
print('from checkpoint: {} last epoch:{}'.format(checkpoint, last_epoch))
scheduler = get_scheduler(config, optimizer, last_epoch)
# dataloaders = {split:get_dataloader(config, split, get_transform(config, split))
# for split in ['train', 'val']}
print(config.data)
dataloaders = {
'train': get_train_dataloader(config, get_transform(config)),
'val': get_valid_dataloaders(config)[0]
}
writer = SummaryWriter(train_dir)
train(config, model, dataloaders, criterion, optimizer, scheduler, writer,
last_epoch + 1)
def run(config):
model = get_model(config[MODEL_NAME], config[MODEL_PARAMS]).cuda()
criterion = get_loss(config[LOSS_NAME], config[LOSS_PARAMS])
optimizer = get_optimizer(config[OPTIM_NAME],
model.parameters(),
optimizer_params=config[OPTIM_PARAMS])
last_epoch = -1
scheduler = get_scheduler(config[SCHEDULER_NAME], optimizer, last_epoch,
config[SCHEDULER_PARAMS])
datasets = {
stage: CustomDataset(DATA_DIR, stage, config[FOLD_ID],
config[DATA_PREFIX], config[INPUT_SIZE])
for stage in ['train', 'test']
}
dataloaders = {
stage: get_dataloader(datasets[stage], config[BATCH_SIZE])
for stage in ['train', 'test']
}
writer = SummaryWriter(config[TRAIN_DIR])
clip_grad_value_(model.parameters(), 2.0)
train(config, model, dataloaders, criterion, optimizer, scheduler, writer,
last_epoch + 1)
def initialization(self):
WORK_PATH = self.config.WORK_PATH
os.chdir(WORK_PATH)
os.environ["CUDA_VISIBLE_DEVICES"] = self.config.CUDA_VISIBLE_DEVICES
print("GPU is :", os.environ["CUDA_VISIBLE_DEVICES"])
self.model = get_model(self.config)
self.optimizer = get_optimizer(self.config, self.model.parameters())
checkpoint = get_checkpoint(self.config)
if torch.cuda.device_count() > 1:
self.model = torch.nn.DataParallel(self.model)
self.model = self.model.cuda()
self.last_epoch, self.step = load_checkpoint(self.model,
self.optimizer,
self.center_model,
self.optimizer_center,
checkpoint)
print("from checkpoint {} last epoch: {}".format(
checkpoint, self.last_epoch))
self.collate_fn = get_collate_fn(self.config,
self.config.data.frame_num,
self.sample_type) #
def run(config):
train_dir = config.train.dir
task = get_task(config)
optimizer = get_optimizer(config, task.get_model().parameters())
checkpoint = utils.checkpoint.get_initial_checkpoint(config)
if checkpoint is not None:
last_epoch, step = utils.checkpoint.load_checkpoint(
task.get_model(), optimizer, checkpoint)
else:
last_epoch, step = -1, -1
print('from checkpoint: {} last epoch:{}'.format(checkpoint, last_epoch))
scheduler = get_scheduler(config, optimizer, last_epoch)
preprocess_opt = task.get_preprocess_opt()
dataloaders = {
split: get_dataloader(config, split,
get_transform(config, split, **preprocess_opt))
for split in ['train', 'dev']
}
writer = SummaryWriter(config.train.dir)
train(config, task, dataloaders, optimizer, scheduler, writer,
last_epoch + 1)
def find_lr(optimizer='Adam', verbose=0):
fname = 'plots/mnist_lr_finder_for_{}.png'.format(optimizer)
model = MLP()
criterion = keras.losses.SparseCategoricalCrossentropy()
if optimizer == 'Adam':
optimizer = keras.optimizers.Adam(learning_rate=LEARNING_RATE)
else:
optimizer = get_optimizer(optimizer, learning_rate=LEARNING_RATE)
mnist = keras.datasets.mnist
(x_train, y_train), (x_valid, y_valid) = mnist.load_data()
x_train = x_train.reshape(60000, 784).astype('float32') / 255.0
train_dataset = tf.data.Dataset.from_tensor_slices(
(x_train, y_train)).batch(BATCH_SIZE)
@tf.function
def train_step(x_batch, y_batch):
with tf.GradientTape() as tape:
out = model(x_batch, training=True)
loss = criterion(y_batch, out)
grad = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(grad, model.trainable_variables))
return loss
lr_finder = LRFinder(start_lr=1e-7, max_lr=1e-1)
for idx, (x_batch, y_batch) in enumerate(train_dataset):
loss = train_step(x_batch, y_batch)
new_lr = lr_finder.step(loss.numpy())
optimizer.lr.assign(new_lr)
if lr_finder.done:
break
lr_finder.plot_lr(fname)
if verbose:
print(lr_finder.history)
def run(config_file):
config = load_config(config_file)
os.makedirs(config.work_dir, exist_ok=True)
save_config(config, config.work_dir + '/config.yml')
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
all_transforms = {}
all_transforms['train'] = get_transforms(config.transforms.train)
all_transforms['valid'] = get_transforms(config.transforms.test)
dataloaders = {
phase: make_loader(
data_folder=config.data.train_dir,
df_path=config.data.train_df_path,
phase=phase,
batch_size=config.train.batch_size,
num_workers=config.num_workers,
idx_fold=config.data.params.idx_fold,
transforms=all_transforms[phase],
num_classes=config.data.num_classes,
pseudo_label_path=config.train.pseudo_label_path,
task='cls'
)
for phase in ['train', 'valid']
}
# create model
model = CustomNet(config.model.encoder, config.data.num_classes)
# train setting
criterion = get_loss(config)
params = [
{'params': model.base_params(), 'lr': config.optimizer.params.encoder_lr},
{'params': model.fresh_params(), 'lr': config.optimizer.params.decoder_lr}
]
optimizer = get_optimizer(params, config)
scheduler = get_scheduler(optimizer, config)
# model runner
runner = SupervisedRunner(model=model)
callbacks = [MultiClassAccuracyCallback(threshold=0.5), F1ScoreCallback()]
if os.path.exists(config.work_dir + '/checkpoints/best.pth'):
callbacks.append(CheckpointCallback(resume=config.work_dir + '/checkpoints/best_full.pth'))
# model training
runner.train(
model=model,
criterion=criterion,
optimizer=optimizer,
scheduler=scheduler,
loaders=dataloaders,
logdir=config.work_dir,
num_epochs=config.train.num_epochs,
callbacks=callbacks,
verbose=True,
fp16=True,
)
def test_wrong_optimizer_name(self, model_name):
"""Calling `get_optimizer` with a non-supported optimizer_name will
raise a ValueError
"""
model = get_model(model_name)
with pytest.raises(ValueError):
optimizer = get_optimizer("blah", model.parameters(), lr=0.001)
def run(config):
train_dir = config.train.dir
student_model = get_model(config, model_type).to(device)
criterion = get_loss(config)
trainable_params = filter(lambda p: p.requires_grad,
student_model.parameters())
optimizer = get_optimizer(config, trainable_params)
checkpoint = utils.checkpoint.get_initial_checkpoint(config,
model_type=model_type)
if checkpoint is not None:
last_epoch, step = utils.checkpoint.load_checkpoint(
student_model, optimizer, checkpoint, model_type=model_type)
else:
last_epoch, step = -1, -1
print('student model from checkpoint: {} last epoch:{}'.format(
checkpoint, last_epoch))
scheduler = get_scheduler(config, optimizer, last_epoch)
print(config.data)
dataloaders = {
'train': get_train_dataloader(config),
'val': get_valid_dataloader(config),
'test': get_test_dataloader(config)
}
writer = SummaryWriter(config.train['student' + '_dir'])
visualizer = get_visualizer(config)
train(config, student_model, dataloaders, criterion, optimizer, scheduler,
writer, visualizer, last_epoch + 1)
def main():
# read configuration file
with open(sys.argv[1]) as configs:
config_file = json.load(configs)
# Load all the paths
PATH_TO_IMAGES = config_file["path_to_images"]
TRAIN_DATASET_CSV = config_file["path_to_train_csv"]
TEST_DATASET_CSV = config_file["path_to_test_csv"]
PATH_TO_WEIGHTS = config_file["path_to_weights"]
PATH_TO_RESULTS = config_file["path_to_results"]
WEIGHTS_FILE = PATH_TO_IMAGES + "weights.pt"
# Creates the results folder
# This folder will contain the train and test results, config file and weights of the model
results_directory = PATH_TO_RESULTS + get_time() + "/"
os.mkdir(results_directory)
copy2(sys.argv[1], results_directory)
# Transform of the images
transform = transforms.Compose([
transforms.Resize((config_file["image_size"],
config_file["image_size"])), # Image size
transforms.ToTensor()
])
# Datasets
train_dataset = BoneDataset(TRAIN_DATASET_CSV, PATH_TO_IMAGES, transform,
config_file["region"])
test_dataset = BoneDataset(TEST_DATASET_CSV, PATH_TO_IMAGES, transform,
config_file["region"])
# Train loader
train_loader = torch.utils.data.DataLoader(
dataset=train_dataset,
batch_size=config_file["train_batch_size"],
shuffle=True)
# Test loader
test_loader = torch.utils.data.DataLoader(
dataset=test_dataset,
batch_size=config_file["test_batch_size"],
shuffle=True)
# device, model, optimizer, criterion , MAE and results
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = get_model(config_file["model"], PATH_TO_WEIGHTS).to(device)
optimizer = get_optimizer(model, config_file["optimizer"],
config_file["optimizer_hyperparameters"])
criterion = get_criterion(config_file["criterion"])
results = Results(results_directory)
for epoch in range(1, config_file["epochs"] + 1):
train(model, device, train_loader, optimizer, criterion, epoch,
config_file["log_interval"], config_file["decay_lr_interval"],
results)
test(model, device, test_loader, criterion, results)
torch.save(model.state_dict(), WEIGHTS_FILE)
results.write_results()
def _get_optimizer(self, model_params):
optimizer = get_optimizer(
model_params,
self.config['runner']['total_steps'],
self.config['optimizer']
)
self._load_weight(optimizer, 'Optimizer')
return optimizer
def __init__(self, config):
"""Initialize Trainer
Args:
config (dict): Configuration dictionary
"""
super(Trainer, self).__init__()
# Define multi-task setting
dataset = config['dataset']
dataset_name = dataset['dataset_name']
self.tasks_weighting = dataset['tasks_weighting']
self.tasks = [k for k, v in self.tasks_weighting.items()]
# Setup network
model_config = config['model']
self.model = get_module(model_config, dataset_name, self.tasks)
print('Model constructed for {}'.format(' '.join(self.tasks)))
if 'grouping' in model_config:
print('groups = {}'.format(model_config['grouping']['groups']))
print('grouping method = {}'.format(model_config['grouping']['method']))
self.model = update_module(config, self.model, self.tasks)
# Setup for a task-conditional setting
model_params = config['model']['parameters']
if 'common_mt_params' in model_params:
self.task_conditional = not model_params['common_mt_params']
else:
self.task_conditional = False
# Setup optimizers
optimizer_config = config['optimizer']
optimizer_cls = get_optimizer(optimizer_config['algorithm'])
model_params = get_params(self.model, optimizer_config['parameters']['lr'], len(self.tasks),
self.task_conditional, self.tasks)
self.optimizer = optimizer_cls(model_params, **optimizer_config['parameters'])
# Setup schedulers
scheduler_config = config['scheduler']
scheduler_cls = get_scheduler(scheduler_config['lr_policy'])
self.scheduler = scheduler_cls(self.optimizer, **scheduler_config['parameters'])
# Setup loss function
losses_config = config['loss']
self.criterions = get_loss_functions(self.tasks, losses_config)
# Initialise performance meters
self.best_val_loss = 1e9
self.train_loss = {}
self.val_loss = {}
for task in self.tasks:
self.train_loss[task] = get_running_meter()
self.val_loss[task] = get_running_meter()
# Initialize img logging for visualization
self.img_logging = get_img_logging(dataset_name, self.tasks)
self.pred_decoder = get_pred_decoder(dataset_name, self.tasks)
def start_experiment(self, request):
params = request.get('params')
optimizer_title = request.get('optimizer', 'OnePlusOne')
params_space = self.parse_params_space(params)
self.logger.info(f'Params space defined: {str(params_space)}')
optimizer = get_optimizer(optimizer_title)(parametrization=params_space)
experiment_id = str(uuid.uuid4())
self.experiments[experiment_id] = {
'optimizer_title': optimizer_title,
'optimizer': optimizer
}
return experiment_id
def _get_optimizer(self, model_params):
optimizer = get_optimizer(model_params,
self.config['runner']['total_steps'],
self.config['optimizer'])
init_optimizer = self.init_ckpt.get('Optimizer')
if init_optimizer:
print(
'[Runner] - Loading optimizer weights from the previous experiment'
)
optimizer.load_state_dict(init_optimizer)
return optimizer
def test_optimizer_type(self, optimizer_name, model_name):
"""Test that the specified optimizer is the one actually being loaded.
"""
model = get_model(model_name)
optimizer = get_optimizer(optimizer_name, model.parameters(), lr=0.001)
if optimizer_name.lower() == "sgd":
assert isinstance(optimizer, torch.optim.SGD)
elif optimizer_name.lower() == "adam":
assert isinstance(optimizer, torch.optim.Adam)
else:
raise ValueError("My test is flaky. Got unsupported optimizer "
"{} from conftest.".format(optimizer_name))
def search_once(config, policy):
model = get_model(config).cuda()
criterion = get_loss(config)
optimizer = get_optimizer(config, model.parameters())
scheduler = get_scheduler(config, optimizer, -1)
transforms = {'train': get_transform(config, 'train', params={'policies': policy}),
'val': get_transform(config, 'val')}
dataloaders = {split:get_dataloader(config, split, transforms[split])
for split in ['train', 'val']}
score_dict = train(config, model, dataloaders, criterion, optimizer, scheduler, None, 0)
return score_dict['f1_mavg']
def set_model_and_optimizer(self, model, cfg):
self._model = model.to(self.cfg.local_rank)
if self.cfg.n_gpu > 1:
self._model = torch.nn.parallel.DistributedDataParallel(
self._model, device_ids=[self.cfg.local_rank],
output_device=self.cfg.local_rank,
find_unused_parameters=True)
# Build Optimizer
self._optimizer = get_optimizer(optimizer_type="adam",
learning_rate=cfg.learning_rate,
warmup_proportion=cfg.warmup_proportion,
max_grad_norm=cfg.max_grad_norm,
named_parameters=list(self.model_module.named_parameters()),
gradient_accumulation_steps=cfg.gradient_accumulation_steps,
num_steps_per_epoch=len(self._train_loader),
epoch_num=cfg.epoch_num)
def _init_optimizer(self, init=True):
optimizer_cls = get_optimizer(self.optim_cfg)
init_param_names = retrieve_class_init_parameters(optimizer_cls)
optimizer_params = {k: v for k, v in self.optim_cfg.items() if k in init_param_names}
optimizer = optimizer_cls(self.model.parameters(), **optimizer_params)
if init:
self.main_logger.info(f'Using optimizer {optimizer.__class__.__name__}')
if self.resume_cfg.resume_from is not None and self.resume_cfg.saved_optimizer is not None and init:
optimizer_path = get_resume_optimizer_path(self.resume_cfg.resume_from, self.resume_cfg.saved_optimizer)
if init:
self.main_logger.info(f'Loading optimizer from {optimizer_path}')
optimizer.load_state_dict(torch.load(optimizer_path))
return optimizer
def run(config):
teacher_model = get_model(config, 'teacher').to(device)
student_model = get_model(config, 'student').to(device)
print('The nubmer of parameters : %d'%count_parameters(student_model))
criterion = get_loss(config)
# for teacher
optimizer_t = None
checkpoint_t = utils.checkpoint.get_initial_checkpoint(config,
model_type='teacher')
if checkpoint_t is not None:
last_epoch_t, step_t = utils.checkpoint.load_checkpoint(teacher_model,
optimizer_t, checkpoint_t, model_type='teacher')
else:
last_epoch_t, step_t = -1, -1
print('teacher model from checkpoint: {} last epoch:{}'.format(
checkpoint_t, last_epoch_t))
# for student
optimizer_s = get_optimizer(config, student_model)
checkpoint_s = utils.checkpoint.get_initial_checkpoint(config,
model_type='student')
if checkpoint_s is not None:
last_epoch_s, step_s = utils.checkpoint.load_checkpoint(student_model,
optimizer_s, checkpoint_s, model_type='student')
else:
last_epoch_s, step_s = -1, -1
print('student model from checkpoint: {} last epoch:{}'.format(
checkpoint_s, last_epoch_s))
scheduler_s = get_scheduler(config, optimizer_s, last_epoch_s)
print(config.data)
dataloaders = {'train':get_train_dataloader(config, get_transform(config)),
'val':get_valid_dataloader(config)}
#'test':get_test_dataloader(config)}
writer = SummaryWriter(config.train['student' + '_dir'])
visualizer = get_visualizer(config)
result = train(config, student_model, teacher_model, dataloaders,
criterion, optimizer_s, scheduler_s, writer,
visualizer, last_epoch_s+1)
print('best psnr : %.3f, best epoch: %d'%(result['best_psnr'], result['best_epoch']))
def run_experiment(queries: List[Dict[str,
Any]], best_index: int, n_trials: int,
dfs: Dict[str, pd.DataFrame], dataset_params: Dict[str,
Any],
optimizer_params: Dict[str, Any], output_file: RichPath):
times: List[float] = []
total_time: float = 0.0
best_count: float = 0.0
avg_times: List[float] = []
best_frac: List[float] = []
selected_index: List[int] = []
optimizer = get_optimizer(optimizer_params['name'], queries,
**optimizer_params['params'])
for t in range(1, n_trials + 1):
start = time.time()
query_index, q = optimizer.get_query(time=t)
df = merge_frames(q['from'], q['on'], options=[['home', 'away']])
elapsed = time.time() - start
optimizer.update(query_index, reward=-elapsed)
df.copy()
# Avoid time measurements on the first iteration due to caching
selected_index.append(query_index)
if t > 1:
times.append(elapsed)
total_time += elapsed
avg_times.append(total_time / (t - 1))
best_count += float(query_index == best_index)
best_frac.append(best_count / (t - 1))
if t % 100 == 0:
print(f'Completed {t} trials')
# Collect and write the metrics
metrics = dict(times=times,
avg_times=avg_times,
best_frac=best_frac,
selected_index=selected_index)
output_file.save_as_compressed_file(metrics)
def run(args, log):
df = pd.read_csv(args.df_path)
df_train = df[df['Fold']!=args.fold]
df_valid = df[df['Fold']==args.fold]
dfs = {}
dfs['train'] = df_train
dfs['val'] = df_valid
model = get_model(args).cuda()
if args.mode != 'segmentation':
for param in model.model.encoder.parameters():
param.requires_grad = True
for param in model.model.decoder.parameters():
param.requires_grad = True
for params in model.model.classification_head.parameters():
params.requires_grad = False
elif args.mode == 'classification':
for param in model.model.encoder.parameters():
param.requires_grad = False
for param in model.model.decoder.parameters():
param.requires_grad = False
for param in model.classification_head.parameters():
param.requires_grad = True
criterion = get_loss(args)
optimizer = get_optimizer(args, model)
if args.initial_ckpt is not None:
last_epoch, step = checkpoint.load_checkpoint(args, model, checkpoint=args.initial_ckpt)
log.write(f'Resume training from {args.initial_ckpt} @ {last_epoch}\n')
else:
last_epoch, step = -1, -1
dataloaders = {mode:get_dataloader(args.data_dir, dfs[mode], mode, args.pretrain, args.batch_size) for mode in ['train', 'val']}
seed_everything(seed=123)
clr = CLR(optimizer, len(dataloaders['train']))
train(args, model, dataloaders['train'], criterion, optimizer, clr)
def run(config):
train_group_csv_dir = './data/group_csv/'
writer = SummaryWriter(config.train.dir)
train_filenames = list(glob.glob(os.path.join(train_group_csv_dir, 'data_train_group_*')))[1:]
for ti, train_file in tqdm.tqdm(enumerate(train_filenames)):
gi_tr = train_file.replace('data_train_group_', '')
gi_tr = gi_tr.split('/')[-1]
gi_tr = gi_tr.replace('.csv', '')
group_idx = int(gi_tr)
utils.prepare_train_directories(config, group_idx)
model = get_model(config, group_idx)
if torch.cuda.is_available():
model = model.cuda()
criterion = get_loss(config)
optimizer = get_optimizer(config, model.parameters())
checkpoint = utils.checkpoint.get_initial_checkpoint(config, group_idx)
if checkpoint is not None:
last_epoch, step = utils.checkpoint.load_checkpoint(model, optimizer, checkpoint)
else:
last_epoch, step = -1, -1
if last_epoch > config.train.num_epochs:
print('group -- ', str(group_idx), '-- index-', ti, ' ----已xl，跳过')
continue
print('from checkpoint: {} last epoch:{}'.format(checkpoint, last_epoch))
print('group -- ', str(group_idx), '-- index-', ti)
scheduler = get_scheduler(config, optimizer, last_epoch)
dataloaders = {split:get_dataloader(config, group_idx, split, get_transform(config, split))
for split in ['train', 'val']}
train(config,group_idx, model, dataloaders, criterion, optimizer, scheduler,
writer, last_epoch+1)
def __init__(self,
layers=[],
n_batches=10,
loss='categorical_crossentropy',
metric='accuracy_score',
optimizer='adam',
optimizer_params={},
save_weights=True,
shuffle=True,
random_seed=None):
self.layers = layers
self.n_batches = n_batches # mini-batches will be generated in the stratified manner
self.loss = loss
self.metric = metric
self.optimizer = optimizer
self.optimizer_params = optimizer_params
self.save_weights = save_weights
self.template = 'tmp/nn_{0}_{1:.4f}.json'
self.shuffle = shuffle
self.random_seed = random_seed
self.best_layers_ = None
self.best_epoch_ = None
self.best_val_score_ = 0.
self.n_layers_ = len(self.layers)
self._loss = get_metric(self.loss)
if self.loss == 'categorical_crossentropy':
self._loss_grad = lambda actual, predicted: -(actual - predicted)
self._metric = get_metric(self.metric)
self._optimizer = get_optimizer(self.optimizer,
**self.optimizer_params)
self._tts = TrainTestSplitter(shuffle=self.shuffle,
random_seed=self.random_seed)
self._initialized = False
self._training = False
super(NNClassifier, self).__init__(
_y_required=True) # TODO: split into multiple NNs later
data_len = data.shape[0]
dim = data.shape[1]
while True:
indices = np.random.choice(data_len, batch_size * neg_size)
samples = data[indices].reshape(batch_size, neg_size, dim)
yield samples
###############################################################################################################################
## Optimizaer algorithm
#
from optimizers import get_optimizer
optimizer = get_optimizer(args)
###############################################################################################################################
## Building model
from model import create_model
import keras.backend as K
logger.info(' Building model')
model = create_model(args, overall_maxlen, vocab)
# freeze the word embedding layer
model.get_layer('word_emb').trainable = False
model.compile(optimizer=optimizer,
loss=U.max_margin_loss,
metrics=[U.max_margin_loss])