# define lr lr_init = cf.learning_rate if not args_opt.run_distribute else cf.learning_rate * args_opt.device_num lr = get_lr(cf.epoch_size, step_size, lr_init) # define loss loss = CTCLoss(max_sequence_length=cf.captcha_width, max_label_length=max_captcha_digits, batch_size=cf.batch_size) # define net net = StackedRNN(input_size=input_size, batch_size=cf.batch_size, hidden_size=cf.hidden_size) # define opt opt = nn.SGD(params=net.trainable_params(), learning_rate=lr, momentum=cf.momentum) net = WithLossCell(net, loss) net = TrainOneStepCellWithGradClip(net, opt).set_train() # define model model = Model(net) # define callbacks callbacks = [LossMonitor(), TimeMonitor(data_size=step_size)] if cf.save_checkpoint: config_ck = CheckpointConfig( save_checkpoint_steps=cf.save_checkpoint_steps, keep_checkpoint_max=cf.keep_checkpoint_max) ckpt_cb = ModelCheckpoint(prefix="waptctc", directory=cf.save_checkpoint_path, config=config_ck) callbacks.append(ckpt_cb) model.train(cf.epoch_size, dataset, callbacks=callbacks)
def train_eval(config): """ test evaluate """ data_path = config.data_path + config.dataset_type ckpt_path = config.ckpt_path epochs = config.epochs batch_size = config.batch_size if config.dataset_type == "tfrecord": dataset_type = DataType.TFRECORD elif config.dataset_type == "mindrecord": dataset_type = DataType.MINDRECORD else: dataset_type = DataType.H5 ds_train = create_dataset(data_path, train_mode=True, epochs=1, batch_size=batch_size, data_type=dataset_type) print("ds_train.size: {}".format(ds_train.get_dataset_size())) ds_eval = create_dataset(data_path, train_mode=False, epochs=1, batch_size=batch_size, data_type=dataset_type) print("ds_eval.size: {}".format(ds_eval.get_dataset_size())) net_builder = ModelBuilder() train_net, eval_net = net_builder.get_net(config) train_net.set_train() train_model = Model(train_net) train_callback = LossCallBack(config=config) ckptconfig = CheckpointConfig( save_checkpoint_steps=ds_train.get_dataset_size(), keep_checkpoint_max=1) ckpoint_cb = ModelCheckpoint(prefix='widedeep_train', directory=config.ckpt_path, config=ckptconfig) train_model.train(epochs, ds_train, callbacks=[ TimeMonitor(ds_train.get_dataset_size()), train_callback, ckpoint_cb ]) # data download print('Download data from modelarts server to obs.') mox.file.copy_parallel(src_url=config.ckpt_path, dst_url=config.train_url) param_dict = load_checkpoint(find_ckpt(ckpt_path)) load_param_into_net(eval_net, param_dict) auc_metric = AUCMetric() eval_model = Model(train_net, eval_network=eval_net, metrics={"auc": auc_metric}) eval_callback = EvalCallBack(eval_model, ds_eval, auc_metric, config) eval_model.eval(ds_eval, callbacks=eval_callback)
step_size_per_epoch) lr = Tensor(lr) if args.checkpoint != '': param_dict = load_checkpoint(args.pre_trained_model_path) load_param_into_net(model, param_dict) print('Successfully loading the pre-trained model') weights = model.trainable_params() optimizer = Adam(weights, learning_rate=lr, loss_scale=1024.) train_net = TrainOneStepCell(loss_net, optimizer) model = Model(train_net) lr_cb = Monitor(lr) callback_list = [lr_cb] if args.is_distributed: ckpt_path = os.path.join(args.checkpoint_dir, 'ckpt_' + str(get_rank()) + '/') else: ckpt_path = args.checkpoint_dir config_ck = CheckpointConfig(save_checkpoint_steps=step_size_per_epoch, keep_checkpoint_max=10) ckpt_cb = ModelCheckpoint(prefix='wavenet', directory=ckpt_path, config=config_ck) callback_list.append(ckpt_cb) model.train(hparams.nepochs, data_loaders, callbacks=callback_list, dataset_sink_mode=False)
learning_rate=Tensor(lr), momentum=args.momentum, weight_decay=args.weight_decay, loss_scale=args.loss_scale) network = TrainingWrapper(network, opt, sens=args.loss_scale) network.set_train() if args.rank_save_ckpt_flag: # checkpoint save ckpt_max_num = args.max_epoch * args.steps_per_epoch // args.ckpt_interval ckpt_config = CheckpointConfig( save_checkpoint_steps=args.ckpt_interval, keep_checkpoint_max=ckpt_max_num) ckpt_cb = ModelCheckpoint(config=ckpt_config, directory=args.outputs_dir, prefix='{}'.format(args.rank)) cb_params = _InternalCallbackParam() cb_params.train_network = network cb_params.epoch_num = ckpt_max_num cb_params.cur_epoch_num = 1 run_context = RunContext(cb_params) ckpt_cb.begin(run_context) args.logger.info( 'args.steps_per_epoch = {} args.ckpt_interval ={}'.format( args.steps_per_epoch, args.ckpt_interval)) t_end = time.time() for i_all, batch_load in enumerate(data_loader):
max_text_length = config.max_text_length # create dataset dataset = create_dataset(img_dir=os.path.join(dataset_path,'images'), label_dir=os.path.join(dataset_path,'labels'), batch_size=config.batch_size, num_shards=device_num, shard_id=rank, config=config) step_size = dataset.get_dataset_size() lr_init = config.learning_rate lr = nn.dynamic_lr.cosine_decay_lr(0.0, lr_init, config.epoch_size * step_size, step_size, config.epoch_size) loss = CTCLoss(max_sequence_length=config.num_step, max_label_length=max_text_length, batch_size=config.batch_size) net = crnn(config) opt = nn.SGD(params=net.trainable_params(), learning_rate=lr, momentum=config.momentum, nesterov=config.nesterov) net = WithLossCell(net, loss) net = TrainOneStepCellWithGradClip(net, opt).set_train() # define model model = Model(net) # define callbacks callbacks = [LossMonitor(), TimeMonitor(data_size=step_size)] if config.save_checkpoint and rank == 0: config_ck = CheckpointConfig(save_checkpoint_steps=config.save_checkpoint_steps, keep_checkpoint_max=config.keep_checkpoint_max) save_ckpt_path = os.path.join(config.save_checkpoint_path, 'ckpt_' + str(rank) + '/') ckpt_cb = ModelCheckpoint(prefix="crnn", directory=save_ckpt_path, config=config_ck) callbacks.append(ckpt_cb) model.train(config.epoch_size, dataset, callbacks=callbacks, dataset_sink_mode=False)
learning_rate=config.OptimConfig.learning_rate, eps=config.OptimConfig.epsilon, loss_scale=config.OptimConfig.loss_scale) train_net = TrainOneStepCell(loss_net, optimizer) train_net.set_train(True) if args.pre_trained_model_path != '': param_dict = load_checkpoint(args.pre_trained_model_path) load_param_into_net(train_net, param_dict) print('Successfully loading the pre-trained model') model = Model(train_net) lr_cb = Monitor(lr) callback_list = [lr_cb] if args.is_distributed: config.CheckpointConfig.ckpt_file_name_prefix = config.CheckpointConfig.ckpt_file_name_prefix + str( get_rank()) config.CheckpointConfig.ckpt_path = os.path.join( config.CheckpointConfig.ckpt_path, 'ckpt_' + str(get_rank()) + '/') config_ck = CheckpointConfig( save_checkpoint_steps=1, keep_checkpoint_max=config.CheckpointConfig.keep_checkpoint_max) ckpt_cb = ModelCheckpoint( prefix=config.CheckpointConfig.ckpt_file_name_prefix, directory=config.CheckpointConfig.ckpt_path, config=config_ck) callback_list.append(ckpt_cb) model.train(config.TrainingConfig.epochs, ds_train, callbacks=callback_list)
lr_end=config.lr_end, lr_max=config.lr_max, warmup_epochs=config.warmup_epochs, total_epochs=epoch_size, steps_per_epoch=step_size, lr_decay_mode='poly')) opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), lr, config.momentum, config.weight_decay, config.loss_scale) model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale, metrics={'acc'}, amp_level="O2", keep_batchnorm_fp32=False) time_cb = TimeMonitor(data_size=step_size) loss_cb = LossMonitor() cb = [time_cb, loss_cb] if config.save_checkpoint: config_ck = CheckpointConfig( save_checkpoint_steps=config.save_checkpoint_steps, keep_checkpoint_max=config.keep_checkpoint_max) ckpt_cb = ModelCheckpoint(prefix="resnet", directory=config.save_checkpoint_path, config=config_ck) cb += [ckpt_cb] model.train(epoch_size, dataset, callbacks=cb)
loss_scale_manager=None) else: if not args.label_smooth: args.label_smooth_factor = 0.0 loss = CrossEntropy(smooth_factor=args.label_smooth_factor, num_classes=args.num_classes) loss_scale_manager = FixedLossScaleManager(args.loss_scale, drop_overflow_update=False) model = Model(network, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale_manager, amp_level="O2") # checkpoint save progress_cb = ProgressMonitor(args) callbacks = [ progress_cb, ] if args.rank_save_ckpt_flag: ckpt_config = CheckpointConfig( save_checkpoint_steps=args.ckpt_interval * args.steps_per_epoch, keep_checkpoint_max=args.ckpt_save_max) ckpt_cb = ModelCheckpoint(config=ckpt_config, directory=args.outputs_dir, prefix='{}'.format(args.rank)) callbacks.append(ckpt_cb) model.train(args.max_epoch, dataset, callbacks=callbacks)
else: if not args.label_smooth: args.label_smooth_factor = 0.0 loss = CrossEntropy(smooth_factor=args.label_smooth_factor, num_classes=args.num_classes) loss_scale_manager = FixedLossScaleManager(args.loss_scale, drop_overflow_update=False) model = Model(network, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale_manager, amp_level="O2") # define callbacks time_cb = TimeMonitor(data_size=batch_num) loss_cb = LossMonitor(per_print_times=batch_num) callbacks = [time_cb, loss_cb] if args.rank_save_ckpt_flag: ckpt_config = CheckpointConfig( save_checkpoint_steps=args.ckpt_interval * args.steps_per_epoch, keep_checkpoint_max=args.ckpt_save_max) save_ckpt_path = os.path.join(args.outputs_dir, 'ckpt_' + str(args.rank) + '/') ckpt_cb = ModelCheckpoint(config=ckpt_config, directory=save_ckpt_path, prefix='{}'.format(args.rank)) callbacks.append(ckpt_cb) model.train(args.max_epoch, dataset, callbacks=callbacks)
def _build_training_pipeline(config: TransformerConfig, pre_training_dataset=None, fine_tune_dataset=None, test_dataset=None, platform="Ascend"): """ Build training pipeline. Args: config (TransformerConfig): Config of mass model. pre_training_dataset (Dataset): Pre-training dataset. fine_tune_dataset (Dataset): Fine-tune dataset. test_dataset (Dataset): Test dataset. """ net_with_loss = TransformerNetworkWithLoss(config, is_training=True) net_with_loss.init_parameters_data() if config.existed_ckpt: if config.existed_ckpt.endswith(".npz"): weights = np.load(config.existed_ckpt) else: weights = load_checkpoint(config.existed_ckpt) for param in net_with_loss.trainable_params(): weights_name = param.name if weights_name not in weights: raise ValueError( f"Param {weights_name} is not found in ckpt file.") if isinstance(weights[weights_name], Parameter): param.default_input = weights[weights_name].default_input elif isinstance(weights[weights_name], Tensor): param.default_input = Tensor(weights[weights_name].asnumpy(), config.dtype) elif isinstance(weights[weights_name], np.ndarray): param.default_input = Tensor(weights[weights_name], config.dtype) else: param.default_input = weights[weights_name] else: for param in net_with_loss.trainable_params(): name = param.name value = param.default_input if isinstance(value, Tensor): if name.endswith(".gamma"): param.default_input = one_weight(value.asnumpy().shape) elif name.endswith(".beta") or name.endswith(".bias"): param.default_input = zero_weight(value.asnumpy().shape) else: param.default_input = weight_variable( value.asnumpy().shape) dataset = pre_training_dataset if pre_training_dataset is not None \ else fine_tune_dataset if dataset is None: raise ValueError( "pre-training dataset or fine-tuning dataset must be provided one." ) update_steps = dataset.get_repeat_count() * dataset.get_dataset_size() if config.lr_scheduler == "isr": lr = Tensor(square_root_schedule( lr=config.lr, update_num=update_steps, decay_start_step=config.decay_start_step, warmup_steps=config.warmup_steps, min_lr=config.min_lr), dtype=mstype.float32) elif config.lr_scheduler == "poly": lr = Tensor(polynomial_decay_scheduler( lr=config.lr, min_lr=config.min_lr, decay_steps=config.decay_steps, total_update_num=update_steps, warmup_steps=config.warmup_steps, power=config.poly_lr_scheduler_power), dtype=mstype.float32) else: lr = config.lr if config.optimizer.lower() == "adam": optimizer = Adam(net_with_loss.trainable_params(), lr, beta1=0.9, beta2=0.98) elif config.optimizer.lower() == "lamb": lr = BertLearningRate(decay_steps=12000, learning_rate=config.lr, end_learning_rate=config.min_lr, power=10.0, warmup_steps=config.warmup_steps) decay_params = list( filter( lambda x: 'layernorm' not in x.name.lower() and 'bias' not in x .name.lower(), net_with_loss.trainable_params())) other_params = list( filter( lambda x: 'layernorm' in x.name.lower() or 'bias' in x.name. lower(), net_with_loss.trainable_params())) group_params = [{ 'params': decay_params, 'weight_decay': 0.01 }, { 'params': other_params }] optimizer = Lamb(group_params, lr, eps=1e-6) elif config.optimizer.lower() == "momentum": optimizer = Momentum(net_with_loss.trainable_params(), lr, momentum=0.9) else: raise ValueError(f"optimizer only support `adam` and `momentum` now.") # loss scale. if platform == "Ascend": scale_manager = DynamicLossScaleManager( init_loss_scale=config.init_loss_scale, scale_factor=config.loss_scale_factor, scale_window=config.scale_window) else: scale_manager = FixedLossScaleManager(loss_scale=1.0, drop_overflow_update=True) net_with_grads = TransformerTrainOneStepWithLossScaleCell( network=net_with_loss, optimizer=optimizer, scale_update_cell=scale_manager.get_update_cell()) net_with_grads.set_train(True) model = Model(net_with_grads) loss_monitor = LossCallBack(config) ckpt_config = CheckpointConfig( save_checkpoint_steps=config.save_ckpt_steps, keep_checkpoint_max=config.keep_ckpt_max) rank_size = os.getenv('RANK_SIZE') callbacks = [loss_monitor] if rank_size is not None and int( rank_size) > 1 and MultiAscend.get_rank() % 8 == 0: ckpt_callback = ModelCheckpoint( prefix=config.ckpt_prefix, directory=os.path.join(config.ckpt_path, 'ckpt_{}'.format(os.getenv('DEVICE_ID'))), config=ckpt_config) callbacks.append(ckpt_callback) if rank_size is None or int(rank_size) == 1: ckpt_callback = ModelCheckpoint( prefix=config.ckpt_prefix, directory=os.path.join(config.ckpt_path, 'ckpt_{}'.format(os.getenv('DEVICE_ID'))), config=ckpt_config) callbacks.append(ckpt_callback) print(f" | ALL SET, PREPARE TO TRAIN.") _train(model=model, config=config, pre_training_dataset=pre_training_dataset, fine_tune_dataset=fine_tune_dataset, test_dataset=test_dataset, callbacks=callbacks)
get_lr(global_step=0, lr_init=0, lr_end=0, lr_max=config_gpu.lr, warmup_epochs=config_gpu.warmup_epochs, total_epochs=epoch_size, steps_per_epoch=step_size)) opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), lr, config_gpu.momentum, config_gpu.weight_decay, config_gpu.loss_scale) # define model model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale) cb = [Monitor(lr_init=lr.asnumpy())] ckpt_save_dir = config_gpu.save_checkpoint_path + "ckpt_" + str( get_rank()) + "/" if config_gpu.save_checkpoint: config_ck = CheckpointConfig( save_checkpoint_steps=config_gpu.save_checkpoint_epochs * step_size, keep_checkpoint_max=config_gpu.keep_checkpoint_max) ckpt_cb = ModelCheckpoint(prefix="mobilenetV3", directory=ckpt_save_dir, config=config_ck) cb += [ckpt_cb] # begine train model.train(epoch_size, dataset, callbacks=cb)
lr = Tensor(lr) # define optimization opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), lr, config.momentum, config.weight_decay, config.loss_scale) # define model model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale, metrics={'acc'}) print("============== Starting Training ==============") time_callback = TimeMonitor(data_size=step_size) loss_callback = LossMonitor() callbacks = [time_callback, loss_callback] if rank_id == 0: if config.save_checkpoint: config_ckpt = CheckpointConfig( save_checkpoint_steps=config.save_checkpoint_epochs * step_size, keep_checkpoint_max=config.keep_checkpoint_max) ckpt_callback = ModelCheckpoint( prefix="ResNet50", directory=config.save_checkpoint_path, config=config_ckpt) callbacks += [ckpt_callback] model.train(epoch_size, dataset, callbacks=callbacks) print("============== End Training ==============")
net_with_loss = WithLossCell(net, loss) if args_opt.run_distribute: net = TrainOneStepCell(net_with_loss, net, opt, sens=config.loss_scale, reduce_flag=True, mean=True, degree=device_num) else: net = TrainOneStepCell(net_with_loss, net, opt, sens=config.loss_scale) time_cb = TimeMonitor(data_size=dataset_size) loss_cb = LossCallBack() cb = [time_cb, loss_cb] if config.save_checkpoint: ckptconfig = CheckpointConfig( save_checkpoint_steps=config.save_checkpoint_epochs * dataset_size, keep_checkpoint_max=config.keep_checkpoint_max) ckpoint_cb = ModelCheckpoint(prefix='mask_rcnn', directory=config.save_checkpoint_path, config=ckptconfig) cb += [ckpoint_cb] model = Model(net) model.train(config.epoch_size, dataset, callbacks=cb)
def train(cfg): context.set_context(mode=context.GRAPH_MODE, device_target='GPU', save_graphs=False) if cfg['ngpu'] > 1: init("nccl") context.set_auto_parallel_context(device_num=get_group_size(), parallel_mode=ParallelMode.DATA_PARALLEL, gradients_mean=True) cfg['ckpt_path'] = cfg['ckpt_path'] + "ckpt_" + str(get_rank()) + "/" else: raise ValueError('cfg_num_gpu <= 1') batch_size = cfg['batch_size'] max_epoch = cfg['epoch'] momentum = cfg['momentum'] weight_decay = cfg['weight_decay'] initial_lr = cfg['initial_lr'] gamma = cfg['gamma'] training_dataset = cfg['training_dataset'] num_classes = 2 negative_ratio = 7 stepvalues = (cfg['decay1'], cfg['decay2']) ds_train = create_dataset(training_dataset, cfg, batch_size, multiprocessing=True, num_worker=cfg['num_workers']) print('dataset size is : \n', ds_train.get_dataset_size()) steps_per_epoch = math.ceil(ds_train.get_dataset_size()) multibox_loss = MultiBoxLoss(num_classes, cfg['num_anchor'], negative_ratio, cfg['batch_size']) backbone = resnet50(1001) backbone.set_train(True) if cfg['pretrain'] and cfg['resume_net'] is None: pretrained_res50 = cfg['pretrain_path'] param_dict_res50 = load_checkpoint(pretrained_res50) load_param_into_net(backbone, param_dict_res50) print('Load resnet50 from [{}] done.'.format(pretrained_res50)) net = RetinaFace(phase='train', backbone=backbone) net.set_train(True) if cfg['resume_net'] is not None: pretrain_model_path = cfg['resume_net'] param_dict_retinaface = load_checkpoint(pretrain_model_path) load_param_into_net(net, param_dict_retinaface) print('Resume Model from [{}] Done.'.format(cfg['resume_net'])) net = RetinaFaceWithLossCell(net, multibox_loss, cfg) lr = adjust_learning_rate(initial_lr, gamma, stepvalues, steps_per_epoch, max_epoch, warmup_epoch=cfg['warmup_epoch']) if cfg['optim'] == 'momentum': opt = mindspore.nn.Momentum(net.trainable_params(), lr, momentum) elif cfg['optim'] == 'sgd': opt = mindspore.nn.SGD(params=net.trainable_params(), learning_rate=lr, momentum=momentum, weight_decay=weight_decay, loss_scale=1) else: raise ValueError('optim is not define.') net = TrainingWrapper(net, opt) model = Model(net) config_ck = CheckpointConfig(save_checkpoint_steps=cfg['save_checkpoint_steps'], keep_checkpoint_max=cfg['keep_checkpoint_max']) ckpoint_cb = ModelCheckpoint(prefix="RetinaFace", directory=cfg['ckpt_path'], config=config_ck) time_cb = TimeMonitor(data_size=ds_train.get_dataset_size()) callback_list = [LossMonitor(), time_cb, ckpoint_cb] print("============== Starting Training ==============") model.train(max_epoch, ds_train, callbacks=callback_list, dataset_sink_mode=True)
def test_checkpoint_save_ckpt_seconds(): """Test checkpoint save ckpt seconds.""" train_config = CheckpointConfig(save_checkpoint_steps=16, save_checkpoint_seconds=100, keep_checkpoint_max=0, keep_checkpoint_per_n_minutes=1) ckpt_cb = ModelCheckpoint(config=train_config) cb_params = _InternalCallbackParam() net = Net() loss = nn.SoftmaxCrossEntropyWithLogits() optim = Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9) network_ = WithLossCell(net, loss) _train_network = TrainOneStepCell(network_, optim) cb_params.train_network = _train_network cb_params.epoch_num = 10 cb_params.cur_epoch_num = 4 cb_params.cur_step_num = 128 cb_params.batch_num = 32 run_context = RunContext(cb_params) ckpt_cb.begin(run_context) ckpt_cb.step_end(run_context) ckpt_cb2 = ModelCheckpoint(config=train_config) cb_params.cur_epoch_num = 1 cb_params.cur_step_num = 16 ckpt_cb2.begin(run_context) ckpt_cb2.step_end(run_context)
help='dataset_sink_mode is False or True') args = parser.parse_args() context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target) network = AlexNet(cfg.num_classes) loss = nn.SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=True, reduction="mean") lr = Tensor( get_lr(0, cfg.learning_rate, cfg.epoch_size, cfg.save_checkpoint_steps)) opt = nn.Momentum(network.trainable_params(), lr, cfg.momentum) model = Model(network, loss, opt, metrics={"Accuracy": Accuracy()}) # test print("============== Starting Training ==============") ds_train = create_dataset(args.data_path, cfg.batch_size, cfg.epoch_size) time_cb = TimeMonitor(data_size=ds_train.get_dataset_size()) config_ck = CheckpointConfig( save_checkpoint_steps=cfg.save_checkpoint_steps, keep_checkpoint_max=cfg.keep_checkpoint_max) ckpoint_cb = ModelCheckpoint(prefix="checkpoint_alexnet", directory=args.ckpt_path, config=config_ck) model.train(cfg.epoch_size, ds_train, callbacks=[time_cb, ckpoint_cb, LossMonitor()], dataset_sink_mode=args.dataset_sink_mode)
def run_pretrain(): """pre-train bert_clue""" parser = argparse_init() args_opt = parser.parse_args() context.set_context(mode=context.GRAPH_MODE, device_target=args_opt.device_target, device_id=args_opt.device_id) context.set_context(reserve_class_name_in_scope=False) is_auto_enable_graph_kernel = _auto_enable_graph_kernel( args_opt.device_target, args_opt.enable_graph_kernel) _set_graph_kernel_context(args_opt.device_target, args_opt.enable_graph_kernel, is_auto_enable_graph_kernel) ckpt_save_dir = args_opt.save_checkpoint_path if args_opt.distribute == "true": if args_opt.device_target == 'Ascend': D.init() device_num = args_opt.device_num rank = args_opt.device_id % device_num else: D.init() device_num = D.get_group_size() rank = D.get_rank() ckpt_save_dir = args_opt.save_checkpoint_path + 'ckpt_' + str( get_rank()) + '/' context.reset_auto_parallel_context() context.set_auto_parallel_context( parallel_mode=ParallelMode.DATA_PARALLEL, gradients_mean=True, device_num=device_num) _set_bert_all_reduce_split() else: rank = 0 device_num = 1 _check_compute_type(args_opt, is_auto_enable_graph_kernel) if args_opt.accumulation_steps > 1: logger.info("accumulation steps: {}".format( args_opt.accumulation_steps)) logger.info("global batch size: {}".format( cfg.batch_size * args_opt.accumulation_steps)) if args_opt.enable_data_sink == "true": args_opt.data_sink_steps *= args_opt.accumulation_steps logger.info("data sink steps: {}".format(args_opt.data_sink_steps)) if args_opt.enable_save_ckpt == "true": args_opt.save_checkpoint_steps *= args_opt.accumulation_steps logger.info("save checkpoint steps: {}".format( args_opt.save_checkpoint_steps)) ds = create_bert_dataset(device_num, rank, args_opt.do_shuffle, args_opt.data_dir, args_opt.schema_dir) net_with_loss = BertNetworkWithLoss(bert_net_cfg, True) new_repeat_count = args_opt.epoch_size * ds.get_dataset_size( ) // args_opt.data_sink_steps if args_opt.train_steps > 0: train_steps = args_opt.train_steps * args_opt.accumulation_steps new_repeat_count = min(new_repeat_count, train_steps // args_opt.data_sink_steps) else: args_opt.train_steps = args_opt.epoch_size * ds.get_dataset_size( ) // args_opt.accumulation_steps logger.info("train steps: {}".format(args_opt.train_steps)) optimizer = _get_optimizer(args_opt, net_with_loss) callback = [ TimeMonitor(args_opt.data_sink_steps), LossCallBack(ds.get_dataset_size()) ] if args_opt.enable_save_ckpt == "true" and args_opt.device_id % min( 8, device_num) == 0: config_ck = CheckpointConfig( save_checkpoint_steps=args_opt.save_checkpoint_steps, keep_checkpoint_max=args_opt.save_checkpoint_num) ckpoint_cb = ModelCheckpoint( prefix='checkpoint_bert', directory=None if ckpt_save_dir == "" else ckpt_save_dir, config=config_ck) callback.append(ckpoint_cb) if args_opt.load_checkpoint_path: param_dict = load_checkpoint(args_opt.load_checkpoint_path) load_param_into_net(net_with_loss, param_dict) if args_opt.enable_lossscale == "true": update_cell = DynamicLossScaleUpdateCell( loss_scale_value=cfg.loss_scale_value, scale_factor=cfg.scale_factor, scale_window=cfg.scale_window) accumulation_steps = args_opt.accumulation_steps enable_global_norm = cfg.enable_global_norm if accumulation_steps <= 1: if cfg.optimizer == 'AdamWeightDecay' and args_opt.device_target == 'GPU': net_with_grads = BertTrainOneStepWithLossScaleCellForAdam( net_with_loss, optimizer=optimizer, scale_update_cell=update_cell) else: net_with_grads = BertTrainOneStepWithLossScaleCell( net_with_loss, optimizer=optimizer, scale_update_cell=update_cell) else: allreduce_post = args_opt.distribute == "false" or args_opt.allreduce_post_accumulation == "true" net_with_accumulation = ( BertTrainAccumulationAllReducePostWithLossScaleCell if allreduce_post else BertTrainAccumulationAllReduceEachWithLossScaleCell) net_with_grads = net_with_accumulation( net_with_loss, optimizer=optimizer, scale_update_cell=update_cell, accumulation_steps=accumulation_steps, enable_global_norm=enable_global_norm) else: net_with_grads = BertTrainOneStepCell(net_with_loss, optimizer=optimizer) model = Model(net_with_grads) model = ConvertModelUtils().convert_to_thor_model( model, network=net_with_grads, optimizer=optimizer, frequency=cfg.Thor.frequency) model.train(new_repeat_count, ds, callbacks=callback, dataset_sink_mode=(args_opt.enable_data_sink == "true"), sink_size=args_opt.data_sink_steps)
def train(): args = parse_args() # init multicards training if args.is_distributed: init() args.rank = get_rank() args.group_size = get_group_size() parallel_mode = ParallelMode.DATA_PARALLEL context.set_auto_parallel_context(parallel_mode=parallel_mode, gradients_mean=True, device_num=args.group_size) # dataset dataset = data_generator.SegDataset(image_mean=args.image_mean, image_std=args.image_std, data_file=args.data_file, batch_size=args.batch_size, crop_size=args.crop_size, max_scale=args.max_scale, min_scale=args.min_scale, ignore_label=args.ignore_label, num_classes=args.num_classes, num_readers=2, num_parallel_calls=4, shard_id=args.rank, shard_num=args.group_size) dataset = dataset.get_dataset(repeat=1) # network if args.model == 'deeplab_v3_s16': network = net_factory.nets_map[args.model]('train', args.num_classes, 16, args.freeze_bn) elif args.model == 'deeplab_v3_s8': network = net_factory.nets_map[args.model]('train', args.num_classes, 8, args.freeze_bn) else: raise NotImplementedError('model [{:s}] not recognized'.format( args.model)) # loss loss_ = loss.SoftmaxCrossEntropyLoss(args.num_classes, args.ignore_label) loss_.add_flags_recursive(fp32=True) train_net = BuildTrainNetwork(network, loss_) # load pretrained model if args.ckpt_pre_trained: param_dict = load_checkpoint(args.ckpt_pre_trained) load_param_into_net(train_net, param_dict) # optimizer iters_per_epoch = dataset.get_dataset_size() total_train_steps = iters_per_epoch * args.train_epochs if args.lr_type == 'cos': lr_iter = learning_rates.cosine_lr(args.base_lr, total_train_steps, total_train_steps) elif args.lr_type == 'poly': lr_iter = learning_rates.poly_lr(args.base_lr, total_train_steps, total_train_steps, end_lr=0.0, power=0.9) elif args.lr_type == 'exp': lr_iter = learning_rates.exponential_lr(args.base_lr, args.lr_decay_step, args.lr_decay_rate, total_train_steps, staircase=True) else: raise ValueError('unknown learning rate type') opt = nn.Momentum(params=train_net.trainable_params(), learning_rate=lr_iter, momentum=0.9, weight_decay=0.0001, loss_scale=args.loss_scale) # loss scale manager_loss_scale = FixedLossScaleManager(args.loss_scale, drop_overflow_update=False) model = Model(train_net, optimizer=opt, amp_level="O3", loss_scale_manager=manager_loss_scale) # callback for saving ckpts time_cb = TimeMonitor(data_size=iters_per_epoch) loss_cb = LossMonitor() cbs = [time_cb, loss_cb] if args.rank == 0: config_ck = CheckpointConfig( save_checkpoint_steps=args.save_steps, keep_checkpoint_max=args.keep_checkpoint_max) ckpoint_cb = ModelCheckpoint(prefix=args.model, directory=args.train_dir, config=config_ck) cbs.append(ckpoint_cb) model.train(args.train_epochs, dataset, callbacks=cbs)
## fp32 training opt = Momentum( filter(lambda x: x.requires_grad, net.get_parameters()), lr, config.momentum, config.weight_decay) model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc'}) # define callbacks time_cb = TimeMonitor(data_size=step_size) loss_cb = LossMonitor() cb = [time_cb, loss_cb] if config.save_checkpoint: config_ck = CheckpointConfig( save_checkpoint_steps=config.save_checkpoint_epochs * step_size, keep_checkpoint_max=config.keep_checkpoint_max) ckpt_cb = ModelCheckpoint(prefix="resnet", directory=ckpt_save_dir, config=config_ck) cb += [ckpt_cb] if args_opt.elastic: from elastic_schedule import schedule from src.kungfu_mindspore_callbacks import KungFuElasticCallback kungfu_elastic_callback = KungFuElasticCallback(schedule) cb += [kungfu_elastic_callback] dataset_sink_mode = False print('enabled elastic') # from src.debug import DebugStopHook # cb += [DebugStopHook()] cb += [LogStepCallback()]
def main(): parser = argparse.ArgumentParser(description="YOLOv3 train") parser.add_argument("--only_create_dataset", type=bool, default=False, help="If set it true, only create " "Mindrecord, default is false.") parser.add_argument("--distribute", type=bool, default=False, help="Run distribute, default is false.") parser.add_argument("--device_id", type=int, default=0, help="Device id, default is 0.") parser.add_argument("--device_num", type=int, default=1, help="Use device nums, default is 1.") parser.add_argument("--lr", type=float, default=0.001, help="Learning rate, default is 0.001.") parser.add_argument("--mode", type=str, default="sink", help="Run sink mode or not, default is sink") parser.add_argument("--epoch_size", type=int, default=10, help="Epoch size, default is 10") parser.add_argument("--batch_size", type=int, default=32, help="Batch size, default is 32.") parser.add_argument("--pre_trained", type=str, default=None, help="Pretrained checkpoint file path") parser.add_argument("--pre_trained_epoch_size", type=int, default=0, help="Pretrained epoch size") parser.add_argument("--save_checkpoint_epochs", type=int, default=5, help="Save checkpoint epochs, default is 5.") parser.add_argument("--loss_scale", type=int, default=1024, help="Loss scale, default is 1024.") parser.add_argument( "--mindrecord_dir", type=str, default="./Mindrecord_train", help= "Mindrecord directory. If the mindrecord_dir is empty, it wil generate mindrecord file by" "image_dir and anno_path. Note if mindrecord_dir isn't empty, it will use mindrecord_dir " "rather than image_dir and anno_path. Default is ./Mindrecord_train") parser.add_argument("--image_dir", type=str, default="", help="Dataset directory, " "the absolute image path is joined by the image_dir " "and the relative path in anno_path") parser.add_argument("--anno_path", type=str, default="", help="Annotation path.") args_opt = parser.parse_args() context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", device_id=args_opt.device_id) if args_opt.distribute: device_num = args_opt.device_num context.reset_auto_parallel_context() context.set_auto_parallel_context( parallel_mode=ParallelMode.DATA_PARALLEL, mirror_mean=True, device_num=device_num) init() rank = args_opt.device_id % device_num else: rank = 0 device_num = 1 print("Start create dataset!") # It will generate mindrecord file in args_opt.mindrecord_dir, # and the file name is yolo.mindrecord0, 1, ... file_num. if not os.path.isdir(args_opt.mindrecord_dir): os.makedirs(args_opt.mindrecord_dir) prefix = "yolo.mindrecord" mindrecord_file = os.path.join(args_opt.mindrecord_dir, prefix + "0") if not os.path.exists(mindrecord_file): if os.path.isdir(args_opt.image_dir) and os.path.exists( args_opt.anno_path): print("Create Mindrecord.") data_to_mindrecord_byte_image(args_opt.image_dir, args_opt.anno_path, args_opt.mindrecord_dir, prefix=prefix, file_num=8) print("Create Mindrecord Done, at {}".format( args_opt.mindrecord_dir)) else: print("image_dir or anno_path not exits.") if not args_opt.only_create_dataset: loss_scale = float(args_opt.loss_scale) # When create MindDataset, using the fitst mindrecord file, such as yolo.mindrecord0. dataset = create_yolo_dataset(mindrecord_file, repeat_num=args_opt.epoch_size, batch_size=args_opt.batch_size, device_num=device_num, rank=rank) dataset_size = dataset.get_dataset_size() print("Create dataset done!") net = yolov3_resnet18(ConfigYOLOV3ResNet18()) net = YoloWithLossCell(net, ConfigYOLOV3ResNet18()) init_net_param(net, "XavierUniform") # checkpoint ckpt_config = CheckpointConfig(save_checkpoint_steps=dataset_size * args_opt.save_checkpoint_epochs) ckpoint_cb = ModelCheckpoint(prefix="yolov3", directory=None, config=ckpt_config) if args_opt.pre_trained: if args_opt.pre_trained_epoch_size <= 0: raise KeyError( "pre_trained_epoch_size must be greater than 0.") param_dict = load_checkpoint(args_opt.pre_trained) load_param_into_net(net, param_dict) total_epoch_size = 60 if args_opt.distribute: total_epoch_size = 160 lr = Tensor( get_lr(learning_rate=args_opt.lr, start_step=args_opt.pre_trained_epoch_size * dataset_size, global_step=total_epoch_size * dataset_size, decay_step=1000, decay_rate=0.95, steps=True)) opt = nn.Adam(filter(lambda x: x.requires_grad, net.get_parameters()), lr, loss_scale=loss_scale) net = TrainingWrapper(net, opt, loss_scale) callback = [ TimeMonitor(data_size=dataset_size), LossMonitor(), ckpoint_cb ] model = Model(net) dataset_sink_mode = False if args_opt.mode == "sink": print("In sink mode, one epoch return a loss.") dataset_sink_mode = True print( "Start train YOLOv3, the first epoch will be slower because of the graph compilation." ) model.train(args_opt.epoch_size, dataset, callbacks=callback, dataset_sink_mode=dataset_sink_mode)
def do_train(dataset=None, network=None, load_checkpoint_path="", save_checkpoint_path="", epoch_num=1): """ do train """ if load_checkpoint_path == "": raise ValueError( "Pretrain model missed, finetune task must load pretrain model!") steps_per_epoch = dataset.get_dataset_size() # optimizer if optimizer_cfg.optimizer == 'AdamWeightDecay': lr_schedule = BertLearningRate( learning_rate=optimizer_cfg.AdamWeightDecay.learning_rate, end_learning_rate=optimizer_cfg.AdamWeightDecay.end_learning_rate, warmup_steps=int(steps_per_epoch * epoch_num * 0.1), decay_steps=steps_per_epoch * epoch_num, power=optimizer_cfg.AdamWeightDecay.power) params = net_with_loss.trainable_params() decay_params = list( filter(optimizer_cfg.AdamWeightDecay.decay_filter, params)) other_params = list(filter(lambda x: x not in decay_params, params)) group_params = [{ 'params': decay_params, 'weight_decay': optimizer_cfg.AdamWeightDecay.weight_decay }, { 'params': other_params, 'weight_decay': 0.0 }] optimizer = AdamWeightDecay(group_params, lr_schedule, eps=optimizer_cfg.AdamWeightDecay.eps) elif optimizer_cfg.optimizer == 'Lamb': lr_schedule = BertLearningRate( learning_rate=optimizer_cfg.Lamb.learning_rate, end_learning_rate=optimizer_cfg.Lamb.end_learning_rate, warmup_steps=int(steps_per_epoch * epoch_num * 0.1), decay_steps=steps_per_epoch * epoch_num, power=optimizer_cfg.Lamb.power) optimizer = Lamb(network.trainable_params(), learning_rate=lr_schedule) elif optimizer_cfg.optimizer == 'Momentum': optimizer = Momentum( network.trainable_params(), learning_rate=optimizer_cfg.Momentum.learning_rate, momentum=optimizer_cfg.Momentum.momentum) else: raise Exception( "Optimizer not supported. support: [AdamWeightDecay, Lamb, Momentum]" ) # load checkpoint into network ckpt_config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch, keep_checkpoint_max=1) ckpoint_cb = ModelCheckpoint(prefix="classifier", directory=save_checkpoint_path, config=ckpt_config) param_dict = load_checkpoint(load_checkpoint_path) load_param_into_net(network, param_dict) update_cell = DynamicLossScaleUpdateCell(loss_scale_value=2**32, scale_factor=2, scale_window=1000) netwithgrads = BertFinetuneCell(network, optimizer=optimizer, scale_update_cell=update_cell) model = Model(netwithgrads) callbacks = [ TimeMonitor(dataset.get_dataset_size()), LossCallBack(), ckpoint_cb ] model.train(epoch_num, dataset, callbacks=callbacks)
if args_opt.run_distribute: net = TrainOneStepCell(net_with_loss, net, opt, sens=config.loss_scale, reduce_flag=True, mean=True, degree=device_num) else: net = TrainOneStepCell(net_with_loss, net, opt, sens=config.loss_scale) time_cb = TimeMonitor(data_size=dataset_size) loss_cb = LossCallBack(rank_id=rank) cb = [time_cb, loss_cb] if config.save_checkpoint: ckptconfig = CheckpointConfig( save_checkpoint_steps=config.save_checkpoint_epochs * dataset_size, keep_checkpoint_max=config.keep_checkpoint_max) save_checkpoint_path = os.path.join(config.save_checkpoint_path, "ckpt_" + str(rank) + "/") ckpoint_cb = ModelCheckpoint(prefix='ctpn', directory=save_checkpoint_path, config=ckptconfig) cb += [ckpoint_cb] model = Model(net) model.train(training_cfg.total_epoch, dataset, callbacks=cb, dataset_sink_mode=True)
def do_train(dataset=None, network=None, load_checkpoint_path="", save_checkpoint_path="", epoch_num=1): """ Do train Args: dataset: the train dataset. network: the network with loss load_checkpoint_path: the file path which saved pretrain model checkpoint. save_checkpoint_path: the file path which will save finetune model checkpoint. epoch_num: the number of epoch """ if load_checkpoint_path == "": raise ValueError( "Pretrain model missed, finetune task must load pretrain model!") steps_per_epoch = dataset.get_dataset_size( ) # samples / batch_size doing#### if cfg.optimizer == 'AdamWeightDecay': lr_schedule = GPT2LearningRate( learning_rate=cfg.AdamWeightDecay.learning_rate, end_learning_rate=cfg.AdamWeightDecay.end_learning_rate, warmup_steps=int(steps_per_epoch * epoch_num * 0.1), decay_steps=steps_per_epoch * epoch_num, power=cfg.AdamWeightDecay.power) params = network.trainable_params( ) # return a list of all trainable parmeters of the network # Use parameter groups and set different values decay_params = list(filter(cfg.AdamWeightDecay.decay_filter, params)) # without layernorm and bias other_params = list( filter(lambda x: not cfg.AdamWeightDecay.decay_filter(x), params)) # with layernorm and bias group_params = [{ 'params': decay_params, 'weight_decay': cfg.AdamWeightDecay.weight_decay }, { 'params': other_params, 'weight_decay': 0.0 }] optimizer = AdamWeightDecay(group_params, lr_schedule, eps=cfg.AdamWeightDecay.eps) elif cfg.optimizer == 'Lamb': lr_schedule = GPT2LearningRate( learning_rate=cfg.Lamb.learning_rate, end_learning_rate=cfg.Lamb.end_learning_rate, warmup_steps=int(steps_per_epoch * epoch_num * 0.1), decay_steps=steps_per_epoch * epoch_num, power=cfg.Lamb.power) optimizer = Lamb(network.trainable_params(), lr_schedule) elif cfg.optimizer == 'Momentum': optimizer = Momentum(network.trainable_params(), cfg.Momentum.learning_rate, cfg.Momentum.momentum) else: raise Exception( "Optimizer not supported. support: [AdamWeightDecay, Lamb, Momentum]" ) # load checkpoint into network ckpt_config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch, keep_checkpoint_max=1) ckpoint_cb = ModelCheckpoint( prefix="gpt2_language_model_wiki2", directory=None if save_checkpoint_path == "" else save_checkpoint_path, config=ckpt_config) param_dict = load_checkpoint(load_checkpoint_path) final_param_dict = {} for k, v in param_dict.items(): final_param_dict['gpt2_loss.gpt2.gpt2.' + k] = param_dict[k] # set the weights of final linear weights to weights of gpt2 token embedding final_param_dict['gpt2_loss.gpt2.dense1.weight'] = param_dict[ 'gpt2_embedding_lookup.embedding_table'] load_param_into_net(network, final_param_dict) print("Load new parameter successfully!\n") update_cell = DynamicLossScaleUpdateCell(loss_scale_value=2**32, scale_factor=2, scale_window=1000) netwithgrads = GPT2FinetuneCell(network, optimizer=optimizer, scale_update_cell=update_cell) netwithgrads.set_train(True) loss_cb = LossMonitor() model = Model(netwithgrads) # callbacks = [TimeMonitor(dataset.get_dataset_size()), LossCallBack(dataset.get_dataset_size()), ckpoint_cb] callbacks = [TimeMonitor(dataset.get_dataset_size()), loss_cb, ckpoint_cb] print("============== Starting Training ==============") model.train(epoch_num, dataset, callbacks=callbacks) print("============== Training Success ==============")
def do_train(dataset=None, network=None, load_checkpoint_path="", save_checkpoint_path="", epoch_num=1): """ Do train Args: dataset: the train dataset. network: the network with loss load_checkpoint_path: the file path which saved pretrain model checkpoint. save_checkpoint_path: the file path which will save finetune model checkpoint. epoch_num: the number of epoch """ if load_checkpoint_path == "": raise ValueError( "Pretrain model missed, finetune task must load pretrain model!") steps_per_epoch = dataset.get_dataset_size() # optimizer if cfg.optimizer == 'AdamWeightDecay': lr_schedule = GPT2LearningRate( learning_rate=cfg.AdamWeightDecay.learning_rate, end_learning_rate=cfg.AdamWeightDecay.end_learning_rate, warmup_steps=int(steps_per_epoch * epoch_num * 0.1), decay_steps=steps_per_epoch * epoch_num, power=cfg.AdamWeightDecay.power) params = network.trainable_params() decay_params = list(filter(cfg.AdamWeightDecay.decay_filter, params)) other_params = list( filter(lambda x: not cfg.AdamWeightDecay.decay_filter(x), params)) group_params = [{ 'params': decay_params, 'weight_decay': cfg.AdamWeightDecay.weight_decay }, { 'params': other_params, 'weight_decay': 0.0 }] optimizer = AdamWeightDecay(group_params, lr_schedule, eps=cfg.AdamWeightDecay.eps) elif cfg.optimizer == 'Lamb': lr_schedule = GPT2LearningRate( learning_rate=cfg.Lamb.learning_rate, end_learning_rate=cfg.Lamb.end_learning_rate, warmup_steps=int(steps_per_epoch * epoch_num * 0.1), decay_steps=steps_per_epoch * epoch_num, power=cfg.Lamb.power) optimizer = Lamb(network.trainable_params(), lr_schedule) elif cfg.optimizer == 'Momentum': optimizer = Momentum(network.trainable_params(), cfg.Momentum.learning_rate, cfg.Momentum.momentum) else: raise Exception( "Optimizer not supported. support: [AdamWeightDecay, Lamb, Momentum]" ) # load checkpoint into network ckpt_config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch, keep_checkpoint_max=1) prefix_name = "gpt2_summarization_" + str(cfg.gpt2_network) + "_" + str(cfg.optimizer) + "_" \ + str(epoch_num) + "_bs" + str(gpt2_net_cfg.batch_size) ckpoint_cb = ModelCheckpoint( prefix=prefix_name, directory=None if save_checkpoint_path == "" else save_checkpoint_path, config=ckpt_config) param_dict = load_checkpoint(load_checkpoint_path) final_param_dict = {} for name, _ in param_dict.items(): final_param_dict['gpt2.gpt2.' + name] = param_dict[name] final_param_dict['gpt2.lm_head.weight'] = param_dict[ 'gpt2_embedding_lookup.embedding_table'] load_param_into_net(network, final_param_dict) print("Load pretrained parameter successfully!\n") update_cell = DynamicLossScaleUpdateCell(loss_scale_value=2**32, scale_factor=2, scale_window=1000) netwithgrads = GPT2FinetuneCell(network, optimizer=optimizer, scale_update_cell=update_cell) netwithgrads.set_train(True) loss_cb = LossMonitor(per_print_times=1) model = Model(netwithgrads) callbacks = [TimeMonitor(dataset.get_dataset_size()), loss_cb, ckpoint_cb] print("============== Starting Finetuning ==============") model.train(epoch_num, dataset, callbacks=callbacks, dataset_sink_mode=False) print("============== Finetuning Success ==============")
def run_pretrain(): """pre-train bert_clue""" parser = argparse.ArgumentParser(description='bert pre_training') parser.add_argument("--distribute", type=str, default="false", help="Run distribute, default is false.") parser.add_argument("--epoch_size", type=int, default="1", help="Epoch size, default is 1.") parser.add_argument("--device_id", type=int, default=0, help="Device id, default is 0.") parser.add_argument("--device_num", type=int, default=1, help="Use device nums, default is 1.") parser.add_argument("--enable_save_ckpt", type=str, default="true", help="Enable save checkpoint, default is true.") parser.add_argument("--enable_lossscale", type=str, default="true", help="Use lossscale or not, default is not.") parser.add_argument("--do_shuffle", type=str, default="true", help="Enable shuffle for dataset, default is true.") parser.add_argument("--enable_data_sink", type=str, default="true", help="Enable data sink, default is true.") parser.add_argument("--data_sink_steps", type=int, default="1", help="Sink steps for each epoch, default is 1.") parser.add_argument("--checkpoint_path", type=str, default="", help="Checkpoint file path") parser.add_argument("--save_checkpoint_steps", type=int, default=1000, help="Save checkpoint steps, " "default is 1000.") parser.add_argument("--save_checkpoint_num", type=int, default=1, help="Save checkpoint numbers, default is 1.") parser.add_argument("--data_dir", type=str, default="", help="Data path, it is better to use absolute path") parser.add_argument("--schema_dir", type=str, default="", help="Schema path, it is better to use absolute path") args_opt = parser.parse_args() context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", device_id=args_opt.device_id) context.set_context(reserve_class_name_in_scope=False) if args_opt.distribute == "true": device_num = args_opt.device_num context.reset_auto_parallel_context() context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL, mirror_mean=True, device_num=device_num) from mindspore.parallel._auto_parallel_context import auto_parallel_context if bert_net_cfg.num_hidden_layers == 12: auto_parallel_context().set_all_reduce_fusion_split_indices([28, 55, 82, 109, 136, 163, 190, 205]) elif bert_net_cfg.num_hidden_layers == 24: auto_parallel_context().set_all_reduce_fusion_split_indices([38, 93, 148, 203, 258, 313, 368, 397]) D.init() rank = args_opt.device_id % device_num else: rank = 0 device_num = 1 ds, new_repeat_count = create_bert_dataset(args_opt.epoch_size, device_num, rank, args_opt.do_shuffle, args_opt.enable_data_sink, args_opt.data_sink_steps, args_opt.data_dir, args_opt.schema_dir) netwithloss = BertNetworkWithLoss(bert_net_cfg, True) if cfg.optimizer == 'Lamb': optimizer = Lamb(netwithloss.trainable_params(), decay_steps=ds.get_dataset_size() * ds.get_repeat_count(), start_learning_rate=cfg.Lamb.start_learning_rate, end_learning_rate=cfg.Lamb.end_learning_rate, power=cfg.Lamb.power, warmup_steps=cfg.Lamb.warmup_steps, weight_decay=cfg.Lamb.weight_decay, eps=cfg.Lamb.eps) elif cfg.optimizer == 'Momentum': optimizer = Momentum(netwithloss.trainable_params(), learning_rate=cfg.Momentum.learning_rate, momentum=cfg.Momentum.momentum) elif cfg.optimizer == 'AdamWeightDecayDynamicLR': optimizer = AdamWeightDecayDynamicLR(netwithloss.trainable_params(), decay_steps=ds.get_dataset_size() * ds.get_repeat_count(), learning_rate=cfg.AdamWeightDecayDynamicLR.learning_rate, end_learning_rate=cfg.AdamWeightDecayDynamicLR.end_learning_rate, power=cfg.AdamWeightDecayDynamicLR.power, weight_decay=cfg.AdamWeightDecayDynamicLR.weight_decay, eps=cfg.AdamWeightDecayDynamicLR.eps, warmup_steps=cfg.AdamWeightDecayDynamicLR.warmup_steps) else: raise ValueError("Don't support optimizer {}, only support [Lamb, Momentum, AdamWeightDecayDynamicLR]". format(cfg.optimizer)) callback = [TimeMonitor(ds.get_dataset_size()), LossCallBack()] if args_opt.enable_save_ckpt == "true": config_ck = CheckpointConfig(save_checkpoint_steps=args_opt.save_checkpoint_steps, keep_checkpoint_max=args_opt.save_checkpoint_num) ckpoint_cb = ModelCheckpoint(prefix='checkpoint_bert', config=config_ck) callback.append(ckpoint_cb) if args_opt.checkpoint_path: param_dict = load_checkpoint(args_opt.checkpoint_path) load_param_into_net(netwithloss, param_dict) if args_opt.enable_lossscale == "true": update_cell = DynamicLossScaleUpdateCell(loss_scale_value=cfg.loss_scale_value, scale_factor=cfg.scale_factor, scale_window=cfg.scale_window) netwithgrads = BertTrainOneStepWithLossScaleCell(netwithloss, optimizer=optimizer, scale_update_cell=update_cell) else: netwithgrads = BertTrainOneStepCell(netwithloss, optimizer=optimizer) model = Model(netwithgrads) model.train(new_repeat_count, ds, callbacks=callback, dataset_sink_mode=(args_opt.enable_data_sink == "true"))
num_classes=cfg.num_classes, vec_length=cfg.vec_length) # Continue training if set pre_trained to be True if cfg.pre_trained: param_dict = load_checkpoint(cfg.checkpoint_path) load_param_into_net(net, param_dict) opt = nn.Adam(filter(lambda x: x.requires_grad, net.get_parameters()), learning_rate=learning_rate, weight_decay=cfg.weight_decay) loss = SoftmaxCrossEntropyExpand(sparse=True) model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc': Accuracy()}) config_ck = CheckpointConfig(save_checkpoint_steps=int(cfg.epoch_size * batch_num / 2), keep_checkpoint_max=cfg.keep_checkpoint_max) time_cb = TimeMonitor(data_size=batch_num) ckpt_save_dir = "./ckpt_" + str(rank) + "/" ckpoint_cb = ModelCheckpoint(prefix="train_textcnn", directory=ckpt_save_dir, config=config_ck) loss_cb = LossMonitor() model.train(cfg.epoch_size, dataset, callbacks=[time_cb, ckpoint_cb, loss_cb]) print("train success")
def main(): parser = argparse.ArgumentParser(description="SSD training") parser.add_argument("--only_create_dataset", type=bool, default=False, help="If set it true, only create " "Mindrecord, default is false.") parser.add_argument("--distribute", type=bool, default=False, help="Run distribute, default is false.") parser.add_argument("--device_id", type=int, default=0, help="Device id, default is 0.") parser.add_argument("--device_num", type=int, default=1, help="Use device nums, default is 1.") parser.add_argument("--lr", type=float, default=0.25, help="Learning rate, default is 0.25.") parser.add_argument("--mode", type=str, default="sink", help="Run sink mode or not, default is sink.") parser.add_argument("--dataset", type=str, default="coco", help="Dataset, defalut is coco.") parser.add_argument("--epoch_size", type=int, default=70, help="Epoch size, default is 70.") parser.add_argument("--batch_size", type=int, default=32, help="Batch size, default is 32.") parser.add_argument("--checkpoint_path", type=str, default="", help="Checkpoint file path.") parser.add_argument("--save_checkpoint_epochs", type=int, default=5, help="Save checkpoint epochs, default is 5.") parser.add_argument("--loss_scale", type=int, default=1024, help="Loss scale, default is 1024.") args_opt = parser.parse_args() context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", device_id=args_opt.device_id) context.set_context(enable_task_sink=True, enable_loop_sink=True, enable_mem_reuse=True) if args_opt.distribute: device_num = args_opt.device_num context.reset_auto_parallel_context() context.set_auto_parallel_context( parallel_mode=ParallelMode.DATA_PARALLEL, mirror_mean=True, device_num=device_num) init() rank = args_opt.device_id % device_num else: rank = 0 device_num = 1 print("Start create dataset!") # It will generate mindrecord file in args_opt.mindrecord_dir, # and the file name is ssd.mindrecord0, 1, ... file_num. config = ConfigSSD() prefix = "ssd.mindrecord" mindrecord_dir = config.MINDRECORD_DIR mindrecord_file = os.path.join(mindrecord_dir, prefix + "0") if not os.path.exists(mindrecord_file): if not os.path.isdir(mindrecord_dir): os.makedirs(mindrecord_dir) if args_opt.dataset == "coco": if os.path.isdir(config.COCO_ROOT): print("Create Mindrecord.") data_to_mindrecord_byte_image("coco", True, prefix) print("Create Mindrecord Done, at {}".format(mindrecord_dir)) else: print("COCO_ROOT not exits.") else: if os.path.isdir(config.IMAGE_DIR) and os.path.exists( config.ANNO_PATH): print("Create Mindrecord.") data_to_mindrecord_byte_image("other", True, prefix) print("Create Mindrecord Done, at {}".format(mindrecord_dir)) else: print("IMAGE_DIR or ANNO_PATH not exits.") if not args_opt.only_create_dataset: loss_scale = float(args_opt.loss_scale) # When create MindDataset, using the fitst mindrecord file, such as ssd.mindrecord0. dataset = create_ssd_dataset(mindrecord_file, repeat_num=args_opt.epoch_size, batch_size=args_opt.batch_size, device_num=device_num, rank=rank) dataset_size = dataset.get_dataset_size() print("Create dataset done!") ssd = SSD300(backbone=ssd_mobilenet_v2(), config=config) net = SSDWithLossCell(ssd, config) init_net_param(net) # checkpoint ckpt_config = CheckpointConfig(save_checkpoint_steps=dataset_size * args_opt.save_checkpoint_epochs) ckpoint_cb = ModelCheckpoint(prefix="ssd", directory=None, config=ckpt_config) lr = Tensor( get_lr(global_step=0, lr_init=0, lr_end=0, lr_max=args_opt.lr, warmup_epochs=max(args_opt.epoch_size // 20, 1), total_epochs=args_opt.epoch_size, steps_per_epoch=dataset_size)) opt = nn.Momentum( filter(lambda x: x.requires_grad, net.get_parameters()), lr, 0.9, 0.0001, loss_scale) net = TrainingWrapper(net, opt, loss_scale) if args_opt.checkpoint_path != "": param_dict = load_checkpoint(args_opt.checkpoint_path) load_param_into_net(net, param_dict) callback = [ TimeMonitor(data_size=dataset_size), LossMonitor(), ckpoint_cb ] model = Model(net) dataset_sink_mode = False if args_opt.mode == "sink": print("In sink mode, one epoch return a loss.") dataset_sink_mode = True print( "Start train SSD, the first epoch will be slower because of the graph compilation." ) model.train(args_opt.epoch_size, dataset, callbacks=callback, dataset_sink_mode=dataset_sink_mode)
def run_pretrain(): """pre-train bert_clue""" parser = argparse.ArgumentParser(description='bert pre_training') parser.add_argument( '--device_target', type=str, default='Ascend', choices=['Ascend', 'GPU'], help='device where the code will be implemented. (Default: Ascend)') parser.add_argument("--distribute", type=str, default="false", help="Run distribute, default is false.") parser.add_argument("--epoch_size", type=int, default="1", help="Epoch size, default is 1.") parser.add_argument("--device_id", type=int, default=0, help="Device id, default is 0.") parser.add_argument("--device_num", type=int, default=1, help="Use device nums, default is 1.") parser.add_argument("--enable_save_ckpt", type=str, default="true", help="Enable save checkpoint, default is true.") parser.add_argument("--enable_lossscale", type=str, default="true", help="Use lossscale or not, default is not.") parser.add_argument("--do_shuffle", type=str, default="true", help="Enable shuffle for dataset, default is true.") parser.add_argument("--enable_data_sink", type=str, default="true", help="Enable data sink, default is true.") parser.add_argument("--data_sink_steps", type=int, default="1", help="Sink steps for each epoch, default is 1.") parser.add_argument("--save_checkpoint_path", type=str, default="", help="Save checkpoint path") parser.add_argument("--load_checkpoint_path", type=str, default="", help="Load checkpoint file path") parser.add_argument("--save_checkpoint_steps", type=int, default=1000, help="Save checkpoint steps, " "default is 1000.") parser.add_argument("--train_steps", type=int, default=-1, help="Training Steps, default is -1, " "meaning run all steps according to epoch number.") parser.add_argument("--save_checkpoint_num", type=int, default=1, help="Save checkpoint numbers, default is 1.") parser.add_argument("--data_dir", type=str, default="", help="Data path, it is better to use absolute path") parser.add_argument("--schema_dir", type=str, default="", help="Schema path, it is better to use absolute path") args_opt = parser.parse_args() context.set_context(mode=context.GRAPH_MODE, device_target=args_opt.device_target, device_id=args_opt.device_id) context.set_context(reserve_class_name_in_scope=False) context.set_context(variable_memory_max_size="30GB") ckpt_save_dir = args_opt.save_checkpoint_path if args_opt.distribute == "true": if args_opt.device_target == 'Ascend': D.init('hccl') device_num = args_opt.device_num rank = args_opt.device_id % device_num else: D.init('nccl') device_num = D.get_group_size() rank = D.get_rank() ckpt_save_dir = args_opt.save_checkpoint_path + 'ckpt_' + str( rank) + '/' context.reset_auto_parallel_context() context.set_auto_parallel_context( parallel_mode=ParallelMode.DATA_PARALLEL, mirror_mean=True, device_num=device_num) from mindspore.parallel._auto_parallel_context import auto_parallel_context if bert_net_cfg.num_hidden_layers == 12: if bert_net_cfg.use_relative_positions: auto_parallel_context().set_all_reduce_fusion_split_indices( [29, 58, 87, 116, 145, 174, 203, 217]) else: auto_parallel_context().set_all_reduce_fusion_split_indices( [28, 55, 82, 109, 136, 163, 190, 205]) elif bert_net_cfg.num_hidden_layers == 24: if bert_net_cfg.use_relative_positions: auto_parallel_context().set_all_reduce_fusion_split_indices( [30, 90, 150, 210, 270, 330, 390, 421]) else: auto_parallel_context().set_all_reduce_fusion_split_indices( [38, 93, 148, 203, 258, 313, 368, 397]) else: rank = 0 device_num = 1 if args_opt.device_target == 'GPU' and bert_net_cfg.compute_type != mstype.float32: logger.warning('Gpu only support fp32 temporarily, run with fp32.') bert_net_cfg.compute_type = mstype.float32 ds, new_repeat_count = create_bert_dataset(args_opt.epoch_size, device_num, rank, args_opt.do_shuffle, args_opt.enable_data_sink, args_opt.data_sink_steps, args_opt.data_dir, args_opt.schema_dir) if args_opt.train_steps > 0: new_repeat_count = min( new_repeat_count, args_opt.train_steps // args_opt.data_sink_steps) netwithloss = BertNetworkWithLoss(bert_net_cfg, True) if cfg.optimizer == 'Lamb': optimizer = Lamb(netwithloss.trainable_params(), decay_steps=ds.get_dataset_size() * new_repeat_count, start_learning_rate=cfg.Lamb.start_learning_rate, end_learning_rate=cfg.Lamb.end_learning_rate, power=cfg.Lamb.power, warmup_steps=cfg.Lamb.warmup_steps, weight_decay=cfg.Lamb.weight_decay, eps=cfg.Lamb.eps) elif cfg.optimizer == 'Momentum': optimizer = Momentum(netwithloss.trainable_params(), learning_rate=cfg.Momentum.learning_rate, momentum=cfg.Momentum.momentum) elif cfg.optimizer == 'AdamWeightDecayDynamicLR': optimizer = AdamWeightDecayDynamicLR( netwithloss.trainable_params(), decay_steps=ds.get_dataset_size() * new_repeat_count, learning_rate=cfg.AdamWeightDecayDynamicLR.learning_rate, end_learning_rate=cfg.AdamWeightDecayDynamicLR.end_learning_rate, power=cfg.AdamWeightDecayDynamicLR.power, weight_decay=cfg.AdamWeightDecayDynamicLR.weight_decay, eps=cfg.AdamWeightDecayDynamicLR.eps, warmup_steps=cfg.AdamWeightDecayDynamicLR.warmup_steps) else: raise ValueError( "Don't support optimizer {}, only support [Lamb, Momentum, AdamWeightDecayDynamicLR]" .format(cfg.optimizer)) callback = [TimeMonitor(ds.get_dataset_size()), LossCallBack()] if args_opt.enable_save_ckpt == "true": config_ck = CheckpointConfig( save_checkpoint_steps=args_opt.save_checkpoint_steps, keep_checkpoint_max=args_opt.save_checkpoint_num) ckpoint_cb = ModelCheckpoint(prefix='checkpoint_bert', directory=ckpt_save_dir, config=config_ck) callback.append(ckpoint_cb) if args_opt.load_checkpoint_path: param_dict = load_checkpoint(args_opt.load_checkpoint_path) load_param_into_net(netwithloss, param_dict) if args_opt.enable_lossscale == "true": update_cell = DynamicLossScaleUpdateCell( loss_scale_value=cfg.loss_scale_value, scale_factor=cfg.scale_factor, scale_window=cfg.scale_window) netwithgrads = BertTrainOneStepWithLossScaleCell( netwithloss, optimizer=optimizer, scale_update_cell=update_cell) else: netwithgrads = BertTrainOneStepCell(netwithloss, optimizer=optimizer) model = Model(netwithgrads) model.train(new_repeat_count, ds, callbacks=callback, dataset_sink_mode=(args_opt.enable_data_sink == "true"))
# define funsion network network = LeNet5Fusion(cfg.num_classes) # load aware quantizaiton network checkpoint param_dict = load_checkpoint(args.ckpt_path) load_param_into_net(network, param_dict) # convert funsion netwrok to aware quantizaiton network network = quant.convert_quant_network(network, quant_delay=0, bn_fold=False, freeze_bn=10000) net_loss = nn.SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=True, reduction="mean") net_opt = nn.Momentum(network.trainable_params(), cfg.lr, cfg.momentum) time_cb = TimeMonitor(data_size=ds_train.get_dataset_size()) config_ck = CheckpointConfig(save_checkpoint_steps=cfg.epoch_size * step_size, keep_checkpoint_max=cfg.keep_checkpoint_max) ckpoint_cb = ModelCheckpoint(prefix="checkpoint_lenet", config=config_ck) model = Model(network, net_loss, net_opt, metrics={"Accuracy": Accuracy()}) print("============== Starting Training ==============") model.train(cfg['epoch_size'], ds_train, callbacks=[time_cb, ckpoint_cb, LossMonitor()], dataset_sink_mode=args.dataset_sink_mode) print("============== End Training ==============")
init_loss_scale=args.dynamic_init_loss_scale, scale_factor=2, scale_window=2000) model = Model(train_net, optimizer=opt, metrics=None, loss_scale_manager=scale_manager) save_checkpoint_steps = args.ckpt_steps args.logger.info('save_checkpoint_steps:{}'.format(save_checkpoint_steps)) if args.max_ckpts == -1: keep_checkpoint_max = int(args.steps_per_epoch * args.max_epoch / save_checkpoint_steps) + 5 # for more than 5 else: keep_checkpoint_max = args.max_ckpts args.logger.info('keep_checkpoint_max:{}'.format(keep_checkpoint_max)) ckpt_config = CheckpointConfig(save_checkpoint_steps=save_checkpoint_steps, keep_checkpoint_max=keep_checkpoint_max) max_epoch_train = args.max_epoch args.logger.info('max_epoch_train:{}'.format(max_epoch_train)) ckpt_cb = ModelCheckpoint(config=ckpt_config, directory=args.ckpt_path, prefix='{}'.format(args.local_rank)) args.epoch_cnt = 0 progress_cb = ProgressMonitor(args) new_epoch_train = max_epoch_train * steps_per_epoch // args.log_interval model.train(new_epoch_train, de_dataset, callbacks=[progress_cb, ckpt_cb], sink_size=args.log_interval)