def main(args): role = role_maker.PaddleCloudRoleMaker(is_collective=True) fleet.init(role) config = get_config(args.config, overrides=args.override, show=True) gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0)) place = fluid.CUDAPlace(gpu_id) startup_prog = fluid.Program() valid_prog = fluid.Program() valid_dataloader, valid_fetchs = program.build( config, valid_prog, startup_prog, is_train=False) valid_prog = valid_prog.clone(for_test=True) exe = fluid.Executor(place) exe.run(startup_prog) init_model(config, valid_prog, exe) valid_reader = Reader(config, 'valid')() valid_dataloader.set_sample_list_generator(valid_reader, place) compiled_valid_prog = program.compile(config, valid_prog) program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, -1, 'eval', config)
def main(args): config = get_config(args.config, overrides=args.override, show=True) use_gpu = config.get("use_gpu", True) places = fluid.cuda_places() if use_gpu else fluid.cpu_places() startup_prog = fluid.Program() valid_prog = fluid.Program() valid_dataloader, valid_fetchs = program.build(config, valid_prog, startup_prog, is_train=False, is_distributed=False) valid_prog = valid_prog.clone(for_test=True) exe = fluid.Executor(places[0]) exe.run(startup_prog) init_model(config, valid_prog, exe) valid_reader = Reader(config, 'valid')() valid_dataloader.set_sample_list_generator(valid_reader, places) compiled_valid_prog = program.compile(config, valid_prog) program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, -1, 'eval')
def main(args): role = role_maker.PaddleCloudRoleMaker(is_collective=True) fleet.init(role) config = get_config(args.config, overrides=args.override, show=True) # assign the place gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0)) place = fluid.CUDAPlace(gpu_id) # startup_prog is used to do some parameter init work, # and train prog is used to hold the network startup_prog = fluid.Program() train_prog = fluid.Program() train_dataloader, train_fetchs = program.build(config, train_prog, startup_prog, is_train=True) if config.validate: valid_prog = fluid.Program() valid_dataloader, valid_fetchs = program.build(config, valid_prog, startup_prog, is_train=False) # clone to prune some content which is irrelevant in valid_prog valid_prog = valid_prog.clone(for_test=True) # create the "Executor" with the statement of which place exe = fluid.Executor(place=place) # only run startup_prog once to init exe.run(startup_prog) # load model from checkpoint or pretrained model init_model(config, train_prog, exe) train_reader = Reader(config, 'train')() train_dataloader.set_sample_list_generator(train_reader, place) if config.validate: valid_reader = Reader(config, 'valid')() valid_dataloader.set_sample_list_generator(valid_reader, place) compiled_valid_prog = program.compile(config, valid_prog) compiled_train_prog = fleet.main_program for epoch_id in range(config.epochs): # 1. train with train dataset program.run(train_dataloader, exe, compiled_train_prog, train_fetchs, epoch_id, 'train') # 2. validate with validate dataset if config.validate and epoch_id % config.valid_interval == 0: program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, epoch_id, 'valid') # 3. save the persistable model if epoch_id % config.save_interval == 0: model_path = os.path.join(config.model_save_dir, config.ARCHITECTURE["name"]) save_model(train_prog, model_path, epoch_id)
def main(args): role = role_maker.PaddleCloudRoleMaker(is_collective=True) fleet.init(role) config = get_config(args.config, overrides=args.override, show=True) place = env.place() startup_prog = fluid.Program() train_prog = fluid.Program() train_dataloader, train_fetchs = program.build( config, train_prog, startup_prog, is_train=True) if config.validate: valid_prog = fluid.Program() valid_dataloader, valid_fetchs = program.build( config, valid_prog, startup_prog, is_train=False) valid_prog = valid_prog.clone(for_test=True) exe = fluid.Executor(place) exe.run(startup_prog) init_model(config, train_prog, exe) train_reader = Reader(config, 'train')() train_dataloader.set_sample_list_generator(train_reader, place) if config.validate: valid_reader = Reader(config, 'valid')() valid_dataloader.set_sample_list_generator(valid_reader, place) compiled_valid_prog = program.compile(config, valid_prog) compiled_train_prog = fleet.main_program for epoch_id in range(config.epochs): program.run(train_dataloader, exe, compiled_train_prog, train_fetchs, epoch_id, 'train') if config.validate and epoch_id % config.valid_interval == 0: program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, epoch_id, 'valid') if epoch_id % config.save_interval == 0: model_path = os.path.join(config.model_save_dir, config.architecture) save_model(train_prog, model_path, epoch_id)
def main(args): config = get_config(args.config, overrides=args.override, show=True) # assign the place use_gpu = config.get("use_gpu", True) places = fluid.cuda_places() if use_gpu else fluid.cpu_places() # startup_prog is used to do some parameter init work, # and train prog is used to hold the network startup_prog = fluid.Program() train_prog = fluid.Program() best_top1_acc = 0.0 # best top1 acc record if not config.get('use_ema'): train_dataloader, train_fetchs = program.build(config, train_prog, startup_prog, is_train=True, is_distributed=False) else: train_dataloader, train_fetchs, ema = program.build( config, train_prog, startup_prog, is_train=True, is_distributed=False) if config.validate: valid_prog = fluid.Program() valid_dataloader, valid_fetchs = program.build(config, valid_prog, startup_prog, is_train=False, is_distributed=False) # clone to prune some content which is irrelevant in valid_prog valid_prog = valid_prog.clone(for_test=True) # create the "Executor" with the statement of which place exe = fluid.Executor(places[0]) # Parameter initialization exe.run(startup_prog) # load model from 1. checkpoint to resume training, 2. pretrained model to finetune init_model(config, train_prog, exe) train_reader = Reader(config, 'train')() train_dataloader.set_sample_list_generator(train_reader, places) compiled_train_prog = program.compile(config, train_prog, train_fetchs['loss'][0].name) if config.validate: valid_reader = Reader(config, 'valid')() valid_dataloader.set_sample_list_generator(valid_reader, places) compiled_valid_prog = program.compile(config, valid_prog, share_prog=compiled_train_prog) if args.vdl_dir: from visualdl import LogWriter vdl_writer = LogWriter(args.vdl_dir) else: vdl_writer = None for epoch_id in range(config.epochs): # 1. train with train dataset program.run(train_dataloader, exe, compiled_train_prog, train_fetchs, epoch_id, 'train', vdl_writer) # 2. validate with validate dataset if config.validate and epoch_id % config.valid_interval == 0: if config.get('use_ema'): logger.info(logger.coloring("EMA validate start...")) with ema.apply(exe): top1_acc = program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, epoch_id, 'valid') logger.info(logger.coloring("EMA validate over!")) top1_acc = program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, epoch_id, 'valid') if top1_acc > best_top1_acc: best_top1_acc = top1_acc message = "The best top1 acc {:.5f}, in epoch: {:d}".format( best_top1_acc, epoch_id) logger.info("{:s}".format(logger.coloring(message, "RED"))) if epoch_id % config.save_interval == 0: model_path = os.path.join(config.model_save_dir, config.ARCHITECTURE["name"]) save_model(train_prog, model_path, "best_model_in_epoch_" + str(epoch_id)) # 3. save the persistable model if epoch_id % config.save_interval == 0: model_path = os.path.join(config.model_save_dir, config.ARCHITECTURE["name"]) save_model(train_prog, model_path, epoch_id)
def main(args): role = role_maker.PaddleCloudRoleMaker(is_collective=True) fleet.init(role) config = get_config(args.config, overrides=args.override, show=True) # assign the place gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0)) place = fluid.CUDAPlace(gpu_id) # startup_prog is used to do some parameter init work, # and train prog is used to hold the network startup_prog = fluid.Program() train_prog = fluid.Program() best_top1_acc = 0.0 # best top1 acc record if not config.get('use_ema'): train_dataloader, train_fetchs = program.build(config, train_prog, startup_prog, is_train=True) else: train_dataloader, train_fetchs, ema = program.build(config, train_prog, startup_prog, is_train=True) if config.validate: valid_prog = fluid.Program() valid_dataloader, valid_fetchs = program.build(config, valid_prog, startup_prog, is_train=False) # clone to prune some content which is irrelevant in valid_prog valid_prog = valid_prog.clone(for_test=True) # create the "Executor" with the statement of which place exe = fluid.Executor(place) # Parameter initialization exe.run(startup_prog) # load model from 1. checkpoint to resume training, 2. pretrained model to finetune init_model(config, train_prog, exe) train_reader = Reader(config, 'train')() train_dataloader.set_sample_list_generator(train_reader, place) if config.validate: valid_reader = Reader(config, 'valid')() valid_dataloader.set_sample_list_generator(valid_reader, place) compiled_valid_prog = program.compile(config, valid_prog) compiled_train_prog = fleet.main_program vdl_writer = LogWriter(args.vdl_dir) if args.vdl_dir else None for epoch_id in range(config.epochs): # 1. train with train dataset program.run(train_dataloader, exe, compiled_train_prog, train_fetchs, epoch_id, 'train', vdl_writer) if int(os.getenv("PADDLE_TRAINER_ID", 0)) == 0: # 2. validate with validate dataset if config.validate and epoch_id % config.valid_interval == 0: if config.get('use_ema'): logger.info(logger.coloring("EMA validate start...")) with train_fetchs('ema').apply(exe): top1_acc = program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, epoch_id, 'valid') logger.info(logger.coloring("EMA validate over!")) top1_acc = program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, epoch_id, 'valid') if top1_acc > best_top1_acc: best_top1_acc = top1_acc message = "The best top1 acc {:.5f}, in epoch: {:d}".format( best_top1_acc, epoch_id) logger.info("{:s}".format(logger.coloring(message, "RED"))) if epoch_id % config.save_interval == 0: model_path = os.path.join(config.model_save_dir, config.ARCHITECTURE["name"]) save_model(train_prog, model_path, "best_model_in_epoch_" + str(epoch_id)) # 3. save the persistable model if epoch_id % config.save_interval == 0: model_path = os.path.join(config.model_save_dir, config.ARCHITECTURE["name"]) save_model(train_prog, model_path, epoch_id)
def main(args): role = role_maker.PaddleCloudRoleMaker(is_collective=True) fleet.init(role) config = get_config(args.config, overrides=args.override, show=True) use_fp16 = config.get('use_fp16', False) if use_fp16: AMP_RELATED_FLAGS_SETTING = { 'FLAGS_cudnn_exhaustive_search': 1, 'FLAGS_conv_workspace_size_limit': 4000, 'FLAGS_cudnn_batchnorm_spatial_persistent': 1, 'FLAGS_max_inplace_grad_add': 8, } os.environ['FLAGS_cudnn_batchnorm_spatial_persistent'] = '1' paddle.fluid.set_flags(AMP_RELATED_FLAGS_SETTING) # assign the place gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0)) place = fluid.CUDAPlace(gpu_id) # startup_prog is used to do some parameter init work, # and train prog is used to hold the network startup_prog = fluid.Program() train_prog = fluid.Program() best_top1_acc = 0.0 # best top1 acc record if not config.get('use_ema'): train_dataloader, train_fetchs = program.build( config, train_prog, startup_prog, is_train=True) else: train_dataloader, train_fetchs, ema = program.build( config, train_prog, startup_prog, is_train=True) if config.validate: valid_prog = fluid.Program() valid_dataloader, valid_fetchs = program.build( config, valid_prog, startup_prog, is_train=False) # clone to prune some content which is irrelevant in valid_prog valid_prog = valid_prog.clone(for_test=True) # create the "Executor" with the statement of which place exe = fluid.Executor(place) # Parameter initialization exe.run(startup_prog) # load model from 1. checkpoint to resume training, 2. pretrained model to finetune init_model(config, train_prog, exe) if not config.get('use_dali', False): train_reader = Reader(config, 'train')() train_dataloader.set_sample_list_generator(train_reader, place) if config.validate: valid_reader = Reader(config, 'valid')() valid_dataloader.set_sample_list_generator(valid_reader, place) compiled_valid_prog = program.compile(config, valid_prog) else: import dali train_dataloader = dali.train(config) if config.validate and int(os.getenv("PADDLE_TRAINER_ID", 0)): if int(os.getenv("PADDLE_TRAINER_ID", 0)) == 0: valid_dataloader = dali.val(config) compiled_valid_prog = program.compile(config, valid_prog) compiled_train_prog = fleet.main_program vdl_writer = None if args.vdl_dir: if version_info.major == 2: logger.info( "visualdl is just supported for python3, so it is disabled in python2..." ) else: from visualdl import LogWriter vdl_writer = LogWriter(args.vdl_dir) for epoch_id in range(config.epochs): # 1. train with train dataset program.run(train_dataloader, exe, compiled_train_prog, train_fetchs, epoch_id, 'train', config, vdl_writer) if int(os.getenv("PADDLE_TRAINER_ID", 0)) == 0: # 2. validate with validate dataset if config.validate and epoch_id % config.valid_interval == 0: if config.get('use_ema'): logger.info(logger.coloring("EMA validate start...")) with ema.apply(exe): top1_acc = program.run( valid_dataloader, exe, compiled_valid_prog, valid_fetchs, epoch_id, 'valid', config) logger.info(logger.coloring("EMA validate over!")) top1_acc = program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, epoch_id, 'valid', config) if top1_acc > best_top1_acc: best_top1_acc = top1_acc message = "The best top1 acc {:.5f}, in epoch: {:d}".format( best_top1_acc, epoch_id) logger.info("{:s}".format(logger.coloring(message, "RED"))) if epoch_id % config.save_interval == 0: model_path = os.path.join(config.model_save_dir, config.ARCHITECTURE["name"]) save_model(train_prog, model_path, "best_model") # 3. save the persistable model if epoch_id % config.save_interval == 0: model_path = os.path.join(config.model_save_dir, config.ARCHITECTURE["name"]) save_model(train_prog, model_path, epoch_id)