|
- from __future__ import absolute_import
- from __future__ import division
- from __future__ import print_function
-
- import _init_paths
-
- import os
- # os.environ['CUDA_VISIBLE_DEVICES'] = '0'
-
- import torch
- import torch.utils.data
- from fhd_opts import opts
- # from opts import opts
- from models.model import create_model, load_model, save_model
- from models.data_parallel import DataParallel
- from logger import Logger
- from datasets.dataset_factory import get_dataset
- from trains.train_factory import train_factory
-
- from torch.utils.tensorboard import SummaryWriter
-
-
- import AISyncore
- from collections import OrderedDict
-
-
- def main(opt):
- torch.manual_seed(opt.seed)
- torch.backends.cudnn.benchmark = not opt.not_cuda_benchmark and not opt.test
- Dataset = get_dataset(opt.dataset, opt.task)
- opt = opts().update_dataset_info_and_set_heads(opt, Dataset)
- print(opt)
-
- logger = Logger(opt)
-
- os.environ['CUDA_VISIBLE_DEVICES'] = opt.gpus_str
- opt.device = torch.device('cuda' if opt.gpus[0] >= 0 else 'cpu')
- print(f"--------opt.device:{opt.device}------")
-
- print('Creating model...')
- model = create_model(opt.arch, opt.heads, opt.head_conv)
- optimizer = torch.optim.Adam(model.parameters(), opt.lr)
- start_epoch = 0
- if opt.load_model != '':
- model, optimizer, start_epoch = load_model(
- model, opt.load_model, optimizer, opt.resume, opt.lr, opt.lr_step)
-
- Trainer = train_factory[opt.task]
- trainer = Trainer(opt, model, optimizer)
-
- print(f"--------opt.gpus:{opt.gpus}------")
- print(f"--------opt.chunk_sizes:{opt.chunk_sizes}------")
- print(f"--------opt.device:{opt.device}------")
- print(f"--------os.environ['CUDA_VISIBLE_DEVICES']:{os.environ['CUDA_VISIBLE_DEVICES']}------")
- print(f"---- torch.cuda.is_available():{ torch.cuda.is_available() }------")
-
- # test_tensor=torch.rand(3, 3)
- # print(f"--------test_tensor:{test_tensor.device}------")
- # test_tensor=test_tensor.to("cuda:7")
- # print(f"----2----test_tensor:{test_tensor.device}------")
- trainer.set_device(opt.gpus, opt.chunk_sizes, opt.device)
-
- print('Setting up data...')
- val_loader = torch.utils.data.DataLoader(
- Dataset(opt, 'val'),
- batch_size=16,
- shuffle=False,
- num_workers=1,
- pin_memory=True
- )
-
- if opt.test:
- _, preds = trainer.val(0, val_loader)
- val_loader.dataset.run_eval(preds, opt.save_dir)
- return
-
- train_loader = torch.utils.data.DataLoader(
- Dataset(opt, 'train'),
- batch_size=opt.batch_size,
- shuffle=True,
- num_workers=opt.num_workers,
- pin_memory=True,
- drop_last=True
- )
-
- print('Starting training...')
- #定义Client类
- class ellipseNetClient(AISyncore.client.NumPyClient):
- def __init__(self):
- super().__init__()
- self.trainer=trainer
- self.logger=logger
- self.opt=opt
- self.optimizer=optimizer
- self.model=model
- self.epoch=0
- self.best = 1e10
- self.train_loader=train_loader
- self.val_loader=val_loader
- def get_parameters(self):
- return [val.cpu().numpy() for _, val in self.model.state_dict().items()]
-
- def set_parameters(self, parameters):
- params_dict = zip(self.model.state_dict().keys(), parameters)
- state_dict = OrderedDict({k: torch.tensor(v) for k, v in params_dict})
- self.model.load_state_dict(state_dict, strict=True)
-
- def fit(self, parameters, config):
- self.set_parameters(parameters)
- #一轮训练
- mark = self.epoch if self.opt.save_all else 'last'
- log_dict_train, _ = self.trainer.train(self.epoch, self.train_loader)
- self.logger.write('epoch: {} |'.format(self.epoch))
- for k, v in log_dict_train.items():
- self.logger.scalar_summary('train_{}'.format(k), v, self.epoch)
- self.logger.write('{} {:8f} | '.format(k, v))
- save_model(os.path.join(self.opt.save_dir, 'model_last.pth'),
- self.epoch, self.model, self.optimizer)
-
- self.logger.write('\n')
- if self.epoch in self.opt.lr_step:
- save_model(os.path.join(self.opt.save_dir, 'model_{}.pth'.format(self.epoch)),
- self.epoch, self.model, self.optimizer)
- lr = self.opt.lr * (0.1 ** (self.opt.lr_step.index(self.epoch) + 1))
- print('Drop LR to', lr)
- for param_group in self.optimizer.param_groups:
- param_group['lr'] = lr
- self.epoch+=1
- return self.get_parameters(), len(self.train_loader), {}
-
- def evaluate(self, parameters, config):
- self.set_parameters(parameters)
- #model valuation
- mark = self.epoch if self.opt.save_all else 'last'
- if self.opt.val_intervals > 0 and self.epoch % self.opt.val_intervals == 0:
- save_model(os.path.join(self.opt.save_dir, 'model_{}.pth'.format(mark)),
- self.epoch, self.model, self.optimizer)
- with torch.no_grad():
- log_dict_val, preds = self.trainer.val(self.epoch, self.val_loader)
- for k, v in log_dict_val.items():
- self.logger.scalar_summary('val_{}'.format(k), v, self.epoch)
- self.logger.write('{} {:8f} | '.format(k, v))
- if log_dict_val[self.opt.metric] < self.best:
- self.best = log_dict_val[self.opt.metric]
- save_model(os.path.join(self.opt.save_dir, 'model_best.pth'),
- self.epoch, self.model)
- self.logger.write('\n')
- return float(log_dict_val["loss"]), len(self.val_loader), {"accuracy": float(log_dict_val["loss"])}
-
- # Start client
- server_address = "0.0.0.0"
- AISyncore.client.run_numpy_client(server_address + ":8080", client=ellipseNetClient())
-
- logger.close()
-
- if __name__ == '__main__':
- opt = opts().parse()
- main(opt)
|