import os import math import argparse import random import logging import shutil from tqdm import tqdm import torch from data.data_sampler import DistIterSampler from trainer.eval.evaluator import create_evaluator from utils import util, options as option from data import create_dataloader, create_dataset, get_dataset_debugger from trainer.ExtensibleTrainer import ExtensibleTrainer from time import time from datetime import datetime from utils.util import opt_get, map_cuda_to_correct_device def init_dist(backend, **kwargs): # These packages have globals that screw with Windows, so only import them if needed. import torch.distributed as dist rank = int(os.environ['LOCAL_RANK']) assert rank < torch.cuda.device_count() torch.cuda.set_device(rank) dist.init_process_group(backend=backend, **kwargs) class Trainer: def init(self, opt_path, opt, launcher): self._profile = False self.val_compute_psnr = opt_get(opt, ['eval', 'compute_psnr'], False) self.val_compute_fea = opt_get(opt, ['eval', 'compute_fea'], False) self.current_step = 0 self.total_training_data_encountered = 0 #### loading resume state if exists if opt['path'].get('resume_state', None): # distributed resuming: all load into default GPU resume_state = torch.load(opt['path']['resume_state'], map_location=map_cuda_to_correct_device) else: resume_state = None #### mkdir and loggers if self.rank <= 0: # normal training (self.rank -1) OR distributed training (self.rank 0) if resume_state is None: util.mkdir_and_rename( opt['path']['experiments_root']) # rename experiment folder if exists util.mkdirs( (path for key, path in opt['path'].items() if not key == 'experiments_root' and path is not None and 'pretrain_model' not in key and 'resume' not in key)) shutil.copy(opt_path, os.path.join(opt['path']['experiments_root'], f'{datetime.now().strftime("%d%m%Y_%H%M%S")}_{os.path.basename(opt_path)}')) # config loggers. Before it, the log will not work util.setup_logger('base', opt['path']['log'], 'train_' + opt['name'], level=logging.INFO, screen=True, tofile=True) self.logger = logging.getLogger('base') self.logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: self.tb_logger_path = os.path.join(opt['path']['experiments_root'], 'tb_logger') from torch.utils.tensorboard import SummaryWriter self.tb_logger = SummaryWriter(log_dir=self.tb_logger_path) else: util.setup_logger('base', opt['path']['log'], 'train', level=logging.INFO, screen=True) self.logger = logging.getLogger('base') if resume_state is not None: option.check_resume(opt, resume_state['iter']) # check resume options # convert to NoneDict, which returns None for missing keys opt = option.dict_to_nonedict(opt) self.opt = opt #### wandb init if opt['wandb'] and self.rank <= 0: import wandb os.makedirs(os.path.join(opt['path']['log'], 'wandb'), exist_ok=True) project_name = opt_get(opt, ['wandb_project_name'], opt['name']) run_name = opt_get(opt, ['wandb_run_name'], None) wandb.init(project=project_name, dir=opt['path']['log'], config=opt, name=run_name) #### random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) if self.rank <= 0: self.logger.info('Random seed: {}'.format(seed)) seed += self.rank # Different multiprocessing instances should behave differently. util.set_random_seed(seed) torch.backends.cudnn.benchmark = opt_get(opt, ['cuda_benchmarking_enabled'], True) torch.backends.cuda.matmul.allow_tf32 = True # torch.backends.cudnn.deterministic = True if opt_get(opt, ['anomaly_detection'], False): torch.autograd.set_detect_anomaly(True) # Save the compiled opt dict to the global loaded_options variable. util.loaded_options = opt #### create train and val dataloader dataset_ratio = 1 # enlarge the size of each epoch for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': self.train_set, collate_fn = create_dataset(dataset_opt, return_collate=True) self.dataset_debugger = get_dataset_debugger(dataset_opt) if self.dataset_debugger is not None and resume_state is not None: self.dataset_debugger.load_state(opt_get(resume_state, ['dataset_debugger_state'], {})) train_size = int(math.ceil(len(self.train_set) / dataset_opt['batch_size'])) total_iters = int(opt['train']['niter']) self.total_epochs = int(math.ceil(total_iters / train_size)) if opt['dist']: self.train_sampler = DistIterSampler(self.train_set, self.world_size, self.rank, dataset_ratio) self.total_epochs = int(math.ceil(total_iters / (train_size * dataset_ratio))) shuffle = False else: self.train_sampler = None shuffle = True self.train_loader = create_dataloader(self.train_set, dataset_opt, opt, self.train_sampler, collate_fn=collate_fn, shuffle=shuffle) if self.rank <= 0: self.logger.info('Number of training data elements: {:,d}, iters: {:,d}'.format( len(self.train_set), train_size)) self.logger.info('Total epochs needed: {:d} for iters {:,d}'.format( self.total_epochs, total_iters)) elif phase == 'val': self.val_set, collate_fn = create_dataset(dataset_opt, return_collate=True) self.val_loader = create_dataloader(self.val_set, dataset_opt, opt, None, collate_fn=collate_fn) if self.rank <= 0: self.logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(self.val_set))) else: raise NotImplementedError('Phase [{:s}] is not recognized.'.format(phase)) assert self.train_loader is not None #### create model self.model = ExtensibleTrainer(opt) ### Evaluators self.evaluators = [] if 'eval' in opt.keys() and 'evaluators' in opt['eval'].keys(): # In "pure" mode, we propagate through the normal training steps, but use validation data instead and average # the total loss. A validation dataloader is required. if opt_get(opt, ['eval', 'pure'], False): assert hasattr(self, 'val_loader') for ev_key, ev_opt in opt['eval']['evaluators'].items(): self.evaluators.append(create_evaluator(self.model.networks[ev_opt['for']], ev_opt, self.model.env)) #### resume training if resume_state: self.logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) self.start_epoch = resume_state['epoch'] self.current_step = resume_state['iter'] self.total_training_data_encountered = opt_get(resume_state, ['total_data_processed'], 0) if opt_get(opt, ['path', 'optimizer_reset'], False): print('!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!') print('!! RESETTING OPTIMIZER STATES') print('!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!') else: self.model.resume_training(resume_state, 'amp_opt_level' in opt.keys()) # handle optimizers and schedulers else: self.current_step = -1 if 'start_step' not in opt.keys() else opt['start_step'] self.total_training_data_encountered = 0 if 'training_data_encountered' not in opt.keys() else opt['training_data_encountered'] self.start_epoch = 0 if 'force_start_step' in opt.keys(): self.current_step = opt['force_start_step'] self.total_training_data_encountered = self.current_step * opt['datasets']['train']['batch_size'] opt['current_step'] = self.current_step #### validation if 'val_freq' in opt['train'].keys(): self.val_freq = opt['train']['val_freq'] * opt['datasets']['train']['batch_size'] else: self.val_freq = int(opt['train']['val_freq_megasamples'] * 1000000) self.next_eval_step = self.total_training_data_encountered + self.val_freq del resume_state # For whatever reason, this relieves a memory burden on the first GPU for some training sessions. def do_step(self, train_data): if self._profile: print("Data fetch: %f" % (time() - _t)) _t = time() opt = self.opt batch_size = self.opt['datasets']['train']['batch_size'] # It may seem weird to derive this from opt, rather than train_data. The reason this is done is # because train_data is process-local while the opt variant represents all of the data fed across all GPUs. self.current_step += 1 self.total_training_data_encountered += batch_size will_log = self.current_step % opt['logger']['print_freq'] == 0 #### update learning rate self.model.update_learning_rate(self.current_step, warmup_iter=opt['train']['warmup_iter']) #### training if self._profile: print("Update LR: %f" % (time() - _t)) _t = time() self.model.feed_data(train_data, self.current_step) gradient_norms_dict = self.model.optimize_parameters(self.current_step, return_grad_norms=will_log) iteration_rate = (time() - _t) / batch_size if self._profile: print("Model feed + step: %f" % (time() - _t)) _t = time() #### log if self.dataset_debugger is not None: self.dataset_debugger.update(train_data) if will_log: # Must be run by all instances to gather consensus. current_model_logs = self.model.get_current_log(self.current_step) if will_log and self.rank <= 0: logs = {'step': self.current_step, 'samples': self.total_training_data_encountered, 'megasamples': self.total_training_data_encountered / 1000000, 'iteration_rate': iteration_rate} logs.update(current_model_logs) if self.dataset_debugger is not None: logs.update(self.dataset_debugger.get_debugging_map()) logs.update(gradient_norms_dict) message = '[epoch:{:3d}, iter:{:8,d}, lr:('.format(self.epoch, self.current_step) for v in self.model.get_current_learning_rate(): message += '{:.3e},'.format(v) message += ')] ' for k, v in logs.items(): if 'histogram' in k: self.tb_logger.add_histogram(k, v, self.current_step) elif isinstance(v, dict): self.tb_logger.add_scalars(k, v, self.current_step) else: message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: self.tb_logger.add_scalar(k, v, self.current_step) if opt['wandb'] and self.rank <= 0: import wandb wandb_logs = {} for k, v in logs.items(): if 'histogram' in k: wandb_logs[k] = wandb.Histogram(v) else: wandb_logs[k] = v if opt_get(opt, ['wandb_progress_use_raw_steps'], False): wandb.log(wandb_logs, step=self.current_step) else: wandb.log(wandb_logs, step=self.total_training_data_encountered) self.logger.info(message) #### save models and training states if self.current_step % opt['logger']['save_checkpoint_freq'] == 0: self.model.consolidate_state() if self.rank <= 0: self.logger.info('Saving models and training states.') self.model.save(self.current_step) state = {'epoch': self.epoch, 'iter': self.current_step, 'total_data_processed': self.total_training_data_encountered} if self.dataset_debugger is not None: state['dataset_debugger_state'] = self.dataset_debugger.get_state() self.model.save_training_state(state) if 'alt_path' in opt['path'].keys(): import shutil print("Synchronizing tb_logger to alt_path..") alt_tblogger = os.path.join(opt['path']['alt_path'], "tb_logger") shutil.rmtree(alt_tblogger, ignore_errors=True) shutil.copytree(self.tb_logger_path, alt_tblogger) do_eval = self.total_training_data_encountered > self.next_eval_step if do_eval: self.next_eval_step = self.total_training_data_encountered + self.val_freq if opt_get(opt, ['eval', 'pure'], False) and do_eval: metrics = [] for val_data in tqdm(self.val_loader): self.model.feed_data(val_data, self.current_step, perform_micro_batching=False) metrics.append(self.model.test()) reduced_metrics = {} for metric in metrics: for k, v in metric.as_dict().items(): if isinstance(v, torch.Tensor) and len(v.shape) == 0: if k in reduced_metrics.keys(): reduced_metrics[k].append(v) else: reduced_metrics[k] = [v] if self.rank <= 0: for k, v in reduced_metrics.items(): val = torch.stack(v).mean().item() self.tb_logger.add_scalar(f'val_{k}', val, self.current_step) print(f">>Eval {k}: {val}") if opt['wandb']: import wandb wandb.log({f'eval_{k}': torch.stack(v).mean().item() for k,v in reduced_metrics.items()}) if len(self.evaluators) != 0 and do_eval: eval_dict = {} for eval in self.evaluators: if eval.uses_all_ddp or self.rank <= 0: eval_dict.update(eval.perform_eval()) if self.rank <= 0: print("Evaluator results: ", eval_dict) for ek, ev in eval_dict.items(): self.tb_logger.add_scalar(ek, ev, self.current_step) if opt['wandb']: import wandb wandb.log(eval_dict) # Should not be necessary, but make absolutely sure that there is no grad leakage from validation runs. for net in self.model.networks.values(): net.zero_grad() def do_training(self): self.logger.info('Start training from epoch: {:d}, iter: {:d}'.format(self.start_epoch, self.current_step)) for epoch in range(self.start_epoch, self.total_epochs + 1): self.epoch = epoch if self.opt['dist']: self.train_sampler.set_epoch(epoch) tq_ldr = tqdm(self.train_loader) if self.rank <= 0 else self.train_loader _t = time() for train_data in tq_ldr: self.do_step(train_data) def create_training_generator(self, index): self.logger.info('Start training from epoch: {:d}, iter: {:d}'.format(self.start_epoch, self.current_step)) for epoch in range(self.start_epoch, self.total_epochs + 1): self.epoch = epoch if self.opt['dist']: self.train_sampler.set_epoch(epoch) tq_ldr = tqdm(self.train_loader, position=index) _t = time() for train_data in tq_ldr: yield self.model self.do_step(train_data) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_music_instrument_quantizer.yml') parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher') args = parser.parse_args() opt = option.parse(args.opt, is_train=True) if args.launcher != 'none': # export CUDA_VISIBLE_DEVICES for running in distributed mode. if 'gpu_ids' in opt.keys(): gpu_list = ','.join(str(x) for x in opt['gpu_ids']) os.environ['CUDA_VISIBLE_DEVICES'] = gpu_list print('export CUDA_VISIBLE_DEVICES=' + gpu_list) trainer = Trainer() #### distributed training settings if args.launcher == 'none': # disabled distributed training opt['dist'] = False trainer.rank = -1 if len(opt['gpu_ids']) == 1: torch.cuda.set_device(opt['gpu_ids'][0]) print('Disabled distributed training.') else: opt['dist'] = True init_dist('nccl') trainer.world_size = torch.distributed.get_world_size() trainer.rank = torch.distributed.get_rank() torch.cuda.set_device(torch.distributed.get_rank()) trainer.init(args.opt, opt, args.launcher) trainer.do_training()