def main(): print("IN NEW MAIN XD\n") parser = argparse.ArgumentParser() ## Required parameters parser.add_argument( "--input_dir", default=None, type=str, required=True, help="The input data dir. Should contain .hdf5 files for the task.") parser.add_argument("--config_file", default=None, type=str, required=True, help="The BERT model config") parser.add_argument( "--bert_model", default="bert-large-uncased", type=str, help="Bert pre-trained model selected in the list: bert-base-uncased, " "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese." ) parser.add_argument( "--output_dir", default=None, type=str, required=True, help="The output directory where the model checkpoints will be written." ) ## Other parameters parser.add_argument( "--max_seq_length", default=512, type=int, help= "The maximum total input sequence length after WordPiece tokenization. \n" "Sequences longer than this will be truncated, and sequences shorter \n" "than this will be padded.") parser.add_argument( "--max_predictions_per_seq", default=80, type=int, help="The maximum total of masked tokens in input sequence") parser.add_argument("--train_batch_size", default=32, type=int, help="Total batch size for training.") parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument("--max_steps", default=1000, type=float, help="Total number of training steps to perform.") parser.add_argument( "--warmup_proportion", default=0.01, type=float, help= "Proportion of training to perform linear learning rate warmup for. " "E.g., 0.1 = 10%% of training.") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument( '--gradient_accumulation_steps', type=int, default=1, help= "Number of updates steps to accumualte before performing a backward/update pass." ) parser.add_argument( '--fp16', default=False, action='store_true', help="Whether to use 16-bit float precision instead of 32-bit") parser.add_argument( '--loss_scale', type=float, default=0.0, help= 'Loss scaling, positive power of 2 values can improve fp16 convergence.' ) parser.add_argument('--log_freq', type=float, default=50.0, help='frequency of logging loss.') parser.add_argument('--checkpoint_activations', default=False, action='store_true', help="Whether to use gradient checkpointing") parser.add_argument("--resume_from_checkpoint", default=False, action='store_true', help="Whether to resume training from checkpoint.") parser.add_argument('--resume_step', type=int, default=-1, help="Step to resume training from.") parser.add_argument( '--num_steps_per_checkpoint', type=int, default=2000, help="Number of update steps until a model checkpoint is saved to disk." ) args = parser.parse_args() random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) assert (torch.cuda.is_available()) if args.local_rank == -1: device = torch.device("cuda") n_gpu = torch.cuda.device_count() else: torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) n_gpu = 1 # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.distributed.init_process_group(backend='nccl', init_method='env://') logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) if args.gradient_accumulation_steps < 1: raise ValueError( "Invalid gradient_accumulation_steps parameter: {}, should be >= 1" .format(args.gradient_accumulation_steps)) if args.train_batch_size % args.gradient_accumulation_steps != 0: raise ValueError( "Invalid gradient_accumulation_steps parameter: {}, batch size {} should be divisible" .format(args.gradient_accumulation_steps, args.train_batch_size)) args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps if not args.resume_from_checkpoint and os.path.exists( args.output_dir) and (os.listdir(args.output_dir) and os.listdir( args.output_dir) != ['logfile.txt']): raise ValueError( "Output directory ({}) already exists and is not empty.".format( args.output_dir)) if not args.resume_from_checkpoint: os.makedirs(args.output_dir, exist_ok=True) # Prepare model config = BertConfig.from_json_file(args.config_file) model = BertForPreTraining(config) if not args.resume_from_checkpoint: global_step = 0 else: if args.resume_step == -1: model_names = [ f for f in os.listdir(args.output_dir) if f.endswith(".pt") ] args.resume_step = max([ int(x.split('.pt')[0].split('_')[1].strip()) for x in model_names ]) global_step = args.resume_step checkpoint = torch.load(os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step)), map_location="cpu") model.load_state_dict(checkpoint['model'], strict=False) print("resume step from ", args.resume_step) model.to(device) # Prepare optimizer param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [{ 'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01 }, { 'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0 }] if args.fp16: optimizer = FusedAdam( optimizer_grouped_parameters, lr=args.learning_rate, #warmup=args.warmup_proportion, #t_total=args.max_steps, bias_correction=False, weight_decay=0.01, max_grad_norm=1.0) if args.loss_scale == 0: # optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True) model, optimizer = amp.initialize(model, optimizer, opt_level="O2", keep_batchnorm_fp32=False, loss_scale="dynamic") else: # optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale) model, optimizer = amp.initialize(model, optimizer, opt_level="O2", keep_batchnorm_fp32=False, loss_scale=args.loss_scale) scheduler = LinearWarmUpScheduler(optimizer, warmup=args.warmup_proportion, total_steps=args.max_steps) else: optimizer = BertAdam(optimizer_grouped_parameters, lr=args.learning_rate, warmup=args.warmup_proportion, t_total=args.max_steps) if args.resume_from_checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) # , strict=False) if args.local_rank != -1: model = DDP(model) elif n_gpu > 1: model = torch.nn.DataParallel(model) files = [ os.path.join(args.input_dir, f) for f in os.listdir(args.input_dir) if os.path.isfile(os.path.join(args.input_dir, f)) ] files.sort() num_files = len(files) logger.info("***** Running training *****") # logger.info(" Num examples = %d", len(train_data)) logger.info(" Batch size = %d", args.train_batch_size) print(" LR = ", args.learning_rate) model.train() print("Training. . .") most_recent_ckpts_paths = [] print("Training. . .") tr_loss = 0.0 # total added training loss average_loss = 0.0 # averaged loss every args.log_freq steps epoch = 0 training_steps = 0 while True: if not args.resume_from_checkpoint: random.shuffle(files) f_start_id = 0 else: f_start_id = checkpoint['files'][0] files = checkpoint['files'][1:] args.resume_from_checkpoint = False for f_id in range(f_start_id, len(files)): data_file = files[f_id] logger.info("file no %s file %s" % (f_id, data_file)) train_data = pretraining_dataset( input_file=data_file, max_pred_length=args.max_predictions_per_seq) if args.local_rank == -1: train_sampler = RandomSampler(train_data) train_dataloader = DataLoader( train_data, sampler=train_sampler, batch_size=args.train_batch_size * n_gpu, num_workers=4, pin_memory=True) else: train_sampler = DistributedSampler(train_data) train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size, num_workers=4, pin_memory=True) for step, batch in enumerate( tqdm(train_dataloader, desc="File Iteration")): training_steps += 1 batch = [t.to(device) for t in batch] input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels = batch #\ loss = model( input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask, masked_lm_labels=masked_lm_labels, next_sentence_label=next_sentence_labels, checkpoint_activations=args.checkpoint_activations) if n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu. if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: # optimizer.backward(loss) with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() tr_loss += loss average_loss += loss.item() if training_steps % args.gradient_accumulation_steps == 0: if args.fp16: scheduler.step() optimizer.step() optimizer.zero_grad() global_step += 1 if training_steps == 1 * args.gradient_accumulation_steps: logger.info( "Step:{} Average Loss = {} Step Loss = {} LR {}". format(global_step, average_loss, loss.item(), optimizer.param_groups[0]['lr'])) if training_steps % (args.log_freq * args.gradient_accumulation_steps) == 0: logger.info( "Step:{} Average Loss = {} Step Loss = {} LR {}". format(global_step, average_loss / args.log_freq, loss.item(), optimizer.param_groups[0]['lr'])) average_loss = 0 if global_step >= args.max_steps or training_steps % ( args.num_steps_per_checkpoint * args.gradient_accumulation_steps) == 0: if (not torch.distributed.is_initialized() or (torch.distributed.is_initialized() and torch.distributed.get_rank() == 0)): # Save a trained model logger.info( "** ** * Saving fine - tuned model ** ** * ") model_to_save = model.module if hasattr( model, 'module') else model # Only save the model it-self output_save_file = os.path.join( args.output_dir, "ckpt_{}.pt".format(global_step)) torch.save( { 'model': model_to_save.state_dict(), 'optimizer': optimizer.state_dict(), 'files': [f_id] + files }, output_save_file) most_recent_ckpts_paths.append(output_save_file) if len(most_recent_ckpts_paths) > 3: ckpt_to_be_removed = most_recent_ckpts_paths.pop(0) os.remove(ckpt_to_be_removed) if global_step >= args.max_steps: tr_loss = tr_loss * args.gradient_accumulation_steps / training_steps if (torch.distributed.is_initialized()): tr_loss /= torch.distributed.get_world_size() torch.distributed.all_reduce(tr_loss) logger.info("Total Steps:{} Final Loss = {}".format( training_steps, tr_loss.item())) return del train_dataloader del train_sampler del train_data #for obj in gc.get_objects(): # if torch.is_tensor(obj) or (hasattr(obj, 'data') and torch.is_tensor(obj.data)): # del obj torch.cuda.empty_cache() epoch += 1
def prepare_model_and_optimizer(args, device): # Prepare model config = BertConfig.from_json_file(args.config_file) # Padding for divisibility by 8 if config.vocab_size % 8 != 0: config.vocab_size += 8 - (config.vocab_size % 8) model = BertForPreTraining(config) checkpoint = None if not args.resume_from_checkpoint: global_step = 0 else: if args.resume_step == -1 and not args.init_checkpoint: model_names = [ f for f in os.listdir(args.output_dir) if f.endswith(".pt") ] args.resume_step = max([ int(x.split('.pt')[0].split('_')[1].strip()) for x in model_names ]) global_step = args.resume_step if not args.init_checkpoint else 0 if not args.init_checkpoint: checkpoint = torch.load(os.path.join( args.output_dir, "ckpt_{}.pt".format(global_step)), map_location="cpu") else: checkpoint = torch.load(args.init_checkpoint, map_location="cpu") model.load_state_dict(checkpoint['model'], strict=False) if args.phase2: global_step -= args.phase1_end_step if is_main_process(): print("resume step from ", args.resume_step) model.to(device) param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'gamma', 'beta', 'LayerNorm'] optimizer_grouped_parameters = [{ 'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01 }, { 'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0 }] optimizer = FusedLAMB(optimizer_grouped_parameters, lr=args.learning_rate) lr_scheduler = PolyWarmUpScheduler(optimizer, warmup=args.warmup_proportion, total_steps=args.max_steps) if args.fp16: if args.loss_scale == 0: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale="dynamic") else: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale=args.loss_scale) amp._amp_state.loss_scalers[0]._loss_scale = 2**20 if args.resume_from_checkpoint: if args.phase2 or args.init_checkpoint: keys = list(checkpoint['optimizer']['state'].keys()) #Override hyperparameters from previous checkpoint for key in keys: checkpoint['optimizer']['state'][key]['step'] = global_step for iter, item in enumerate( checkpoint['optimizer']['param_groups']): checkpoint['optimizer']['param_groups'][iter][ 'step'] = global_step checkpoint['optimizer']['param_groups'][iter][ 't_total'] = args.max_steps checkpoint['optimizer']['param_groups'][iter][ 'warmup'] = args.warmup_proportion checkpoint['optimizer']['param_groups'][iter][ 'lr'] = args.learning_rate optimizer.load_state_dict(checkpoint['optimizer']) # , strict=False) # Restore AMP master parameters if args.fp16: optimizer._lazy_init_maybe_master_weights() optimizer._amp_stash.lazy_init_called = True optimizer.load_state_dict(checkpoint['optimizer']) for param, saved_param in zip(amp.master_params(optimizer), checkpoint['master params']): param.data.copy_(saved_param.data) if args.local_rank != -1: if not args.allreduce_post_accumulation: model = DDP( model, message_size=250000000, gradient_predivide_factor=torch.distributed.get_world_size()) else: flat_dist_call([param.data for param in model.parameters()], torch.distributed.broadcast, (0, )) elif args.n_gpu > 1: model = torch.nn.DataParallel(model) return model, optimizer, lr_scheduler, checkpoint, global_step
def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument( "--input_dir", default=None, type=str, required=True, help="The input data dir. Should contain .hdf5 files for the task.") parser.add_argument("--config_file", default="bert_config.json", type=str, required=False, help="The BERT model config") ckpt_group = parser.add_mutually_exclusive_group(required=True) ckpt_group.add_argument("--ckpt_dir", default=None, type=str, help="The ckpt directory, e.g. /results") ckpt_group.add_argument("--ckpt_path", default=None, type=str, help="Path to the specific checkpoint") group = parser.add_mutually_exclusive_group(required=True) group.add_argument('--eval', dest='do_eval', action='store_true') group.add_argument('--prediction', dest='do_eval', action='store_false') ## Other parameters parser.add_argument( "--bert_model", default="bert-large-uncased", type=str, required=False, help="Bert pre-trained model selected in the list: bert-base-uncased, " "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese." ) parser.add_argument( "--max_seq_length", default=512, type=int, help= "The maximum total input sequence length after WordPiece tokenization. \n" "Sequences longer than this will be truncated, and sequences shorter \n" "than this will be padded.") parser.add_argument( "--max_predictions_per_seq", default=80, type=int, help="The maximum total of masked tokens in input sequence") parser.add_argument("--ckpt_step", default=-1, type=int, required=False, help="The model checkpoint iteration, e.g. 1000") parser.add_argument("--eval_batch_size", default=8, type=int, help="Total batch size for training.") parser.add_argument( "--max_steps", default=-1, type=int, help= "Total number of eval steps to perform, otherwise use full dataset") parser.add_argument("--no_cuda", default=False, action='store_true', help="Whether not to use CUDA when available") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument( '--fp16', default=False, action='store_true', help="Whether to use 16-bit float precision instead of 32-bit") parser.add_argument("--log_path", help="Out file for DLLogger", default="/workspace/dllogger_inference.out", type=str) args = parser.parse_args() if 'LOCAL_RANK' in os.environ: args.local_rank = int(os.environ['LOCAL_RANK']) if args.local_rank == -1 or args.no_cuda: device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") else: torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.distributed.init_process_group(backend='nccl', init_method='env://') if is_main_process(): dllogger.init(backends=[ dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE, filename=args.log_path), dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE, step_format=format_step) ]) else: dllogger.init(backends=[]) n_gpu = torch.cuda.device_count() if n_gpu > 1: assert (args.local_rank != -1 ) # only use torch.distributed for multi-gpu dllogger.log( step= "device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}". format(device, n_gpu, bool(args.local_rank != -1), args.fp16), data={}) random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if n_gpu > 0: torch.cuda.manual_seed_all(args.seed) # Prepare model config = BertConfig.from_json_file(args.config_file) # Padding for divisibility by 8 if config.vocab_size % 8 != 0: config.vocab_size += 8 - (config.vocab_size % 8) model = BertForPreTraining(config) if args.ckpt_dir: if args.ckpt_step == -1: #retrieve latest model model_names = [ f for f in os.listdir(args.ckpt_dir) if f.endswith(".pt") ] args.ckpt_step = max([ int(x.split('.pt')[0].split('_')[1].strip()) for x in model_names ]) dllogger.log(step="load model saved at iteration", data={"number": args.ckpt_step}) model_file = os.path.join(args.ckpt_dir, "ckpt_" + str(args.ckpt_step) + ".pt") else: model_file = args.ckpt_path state_dict = torch.load(model_file, map_location="cpu")["model"] model.load_state_dict(state_dict, strict=False) if args.fp16: model.half( ) # all parameters and buffers are converted to half precision model.to(device) multi_gpu_training = args.local_rank != -1 and torch.distributed.is_initialized( ) if multi_gpu_training: model = DDP(model) files = [ os.path.join(args.input_dir, f) for f in os.listdir(args.input_dir) if os.path.isfile(os.path.join(args.input_dir, f)) and 'test' in f ] files.sort() dllogger.log(step="***** Running Inference *****", data={}) dllogger.log(step=" Inference batch", data={"size": args.eval_batch_size}) model.eval() nb_instances = 0 max_steps = args.max_steps if args.max_steps > 0 else np.inf global_step = 0 total_samples = 0 begin_infer = time.time() with torch.no_grad(): if args.do_eval: final_loss = 0.0 # for data_file in files: dllogger.log(step="Opening ", data={"file": data_file}) dataset = pretraining_dataset( input_file=data_file, max_pred_length=args.max_predictions_per_seq) if not multi_gpu_training: train_sampler = RandomSampler(dataset) datasetloader = DataLoader(dataset, sampler=train_sampler, batch_size=args.eval_batch_size, num_workers=4, pin_memory=True) else: train_sampler = DistributedSampler(dataset) datasetloader = DataLoader(dataset, sampler=train_sampler, batch_size=args.eval_batch_size, num_workers=4, pin_memory=True) for step, batch in enumerate( tqdm(datasetloader, desc="Iteration")): if global_step > max_steps: break batch = [t.to(device) for t in batch] input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels = batch #\ loss = model(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask, masked_lm_labels=masked_lm_labels, next_sentence_label=next_sentence_labels) final_loss += loss.item() global_step += 1 total_samples += len(datasetloader) torch.cuda.empty_cache() if global_step > max_steps: break final_loss /= global_step if multi_gpu_training: final_loss = torch.tensor(final_loss, device=device) dist.all_reduce(final_loss) final_loss /= torch.distributed.get_world_size() if (not multi_gpu_training or (multi_gpu_training and torch.distributed.get_rank() == 0)): dllogger.log(step="Inference Loss", data={"final_loss": final_loss.item()}) else: # inference # if multi_gpu_training: # torch.distributed.barrier() # start_t0 = time.time() for data_file in files: dllogger.log(step="Opening ", data={"file": data_file}) dataset = pretraining_dataset( input_file=data_file, max_pred_length=args.max_predictions_per_seq) if not multi_gpu_training: train_sampler = RandomSampler(dataset) datasetloader = DataLoader(dataset, sampler=train_sampler, batch_size=args.eval_batch_size, num_workers=4, pin_memory=True) else: train_sampler = DistributedSampler(dataset) datasetloader = DataLoader(dataset, sampler=train_sampler, batch_size=args.eval_batch_size, num_workers=4, pin_memory=True) for step, batch in enumerate( tqdm(datasetloader, desc="Iteration")): if global_step > max_steps: break batch = [t.to(device) for t in batch] input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels = batch #\ lm_logits, nsp_logits = model(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask, masked_lm_labels=None, next_sentence_label=None) nb_instances += input_ids.size(0) global_step += 1 total_samples += len(datasetloader) torch.cuda.empty_cache() if global_step > max_steps: break # if multi_gpu_training: # torch.distributed.barrier() if (not multi_gpu_training or (multi_gpu_training and torch.distributed.get_rank() == 0)): dllogger.log(step="Done Inferring on samples", data={}) end_infer = time.time() dllogger.log(step="Inference perf", data={ "inference_sequences_per_second": total_samples * args.eval_batch_size / (end_infer - begin_infer) })
def prepare_model_and_optimizer(args, device): # Prepare model config = BertConfig.from_json_file(args.config_file) # Padding for divisibility by 8 if config.vocab_size % 8 != 0: config.vocab_size += 8 - (config.vocab_size % 8) model = BertForPreTraining(config) checkpoint = None if not args.resume_from_checkpoint: global_step = 0 else: if args.resume_step == -1: model_names = [ f for f in os.listdir(args.output_dir) if f.endswith(".pt") ] args.resume_step = max([ int(x.split(".pt")[0].split("_")[1].strip()) for x in model_names ]) global_step = args.resume_step checkpoint = torch.load(os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step)), map_location="cpu") model.load_state_dict(checkpoint["model"], strict=False) if args.phase2: global_step -= args.phase1_end_step if is_main_process(): print("resume step from ", args.resume_step) model.to(device) param_optimizer = list(model.named_parameters()) no_decay = ["bias", "gamma", "beta", "LayerNorm"] optimizer_grouped_parameters = [] names = [] count = 1 for n, p in param_optimizer: count += 1 if not any(nd in n for nd in no_decay): optimizer_grouped_parameters.append({ "params": [p], "weight_decay": 0.01, "name": n }) names.append({"params": [n], "weight_decay": 0.01}) if any(nd in n for nd in no_decay): optimizer_grouped_parameters.append({ "params": [p], "weight_decay": 0.00, "name": n }) names.append({"params": [n], "weight_decay": 0.00}) optimizer = BertLAMB(optimizer_grouped_parameters, lr=args.learning_rate, warmup=args.warmup_proportion, t_total=args.max_steps) if args.fp16: if args.loss_scale == 0: # optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True) model, optimizer = amp.initialize( model, optimizer, opt_level="O2", loss_scale="dynamic", master_weights=False if args.accumulate_into_fp16 else True, ) else: # optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale) model, optimizer = amp.initialize( model, optimizer, opt_level="O2", loss_scale=args.loss_scale, master_weights=False if args.accumulate_into_fp16 else True, ) amp._amp_state.loss_scalers[0]._loss_scale = 2**20 if args.resume_from_checkpoint: if args.phase2: keys = list(checkpoint["optimizer"]["state"].keys()) # Override hyperparameters from Phase 1 for key in keys: checkpoint["optimizer"]["state"][key]["step"] = global_step for iter, item in enumerate( checkpoint["optimizer"]["param_groups"]): checkpoint["optimizer"]["param_groups"][iter][ "t_total"] = args.max_steps checkpoint["optimizer"]["param_groups"][iter][ "warmup"] = args.warmup_proportion checkpoint["optimizer"]["param_groups"][iter][ "lr"] = args.learning_rate optimizer.load_state_dict(checkpoint["optimizer"]) # , strict=False) # Restore AMP master parameters if args.fp16: optimizer._lazy_init_maybe_master_weights() optimizer._amp_stash.lazy_init_called = True optimizer.load_state_dict(checkpoint["optimizer"]) for param, saved_param in zip(amp.master_params(optimizer), checkpoint["master params"]): param.data.copy_(saved_param.data) if args.local_rank != -1: if not args.allreduce_post_accumulation: model = DDP( model, message_size=250000000, gradient_predivide_factor=torch.distributed.get_world_size()) else: flat_dist_call([param.data for param in model.parameters()], torch.distributed.broadcast, (0, )) elif args.n_gpu > 1: model = torch.nn.DataParallel(model) return model, optimizer, checkpoint, global_step
def main(): print("IN NEW MAIN XD\n") parser = argparse.ArgumentParser() ## Required parameters parser.add_argument( "--input_dir", default=None, type=str, required=True, help="The input data dir. Should contain .hdf5 files for the task.") parser.add_argument("--config_file", default="bert_config.json", type=str, required=False, help="The BERT model config") parser.add_argument("--ckpt_dir", default=None, type=str, required=True, help="The ckpt directory, e.g. /results") group = parser.add_mutually_exclusive_group(required=True) group.add_argument('--eval', dest='do_eval', action='store_true') group.add_argument('--prediction', dest='do_eval', action='store_false') ## Other parameters parser.add_argument( "--bert_model", default="bert-large-uncased", type=str, required=False, help="Bert pre-trained model selected in the list: bert-base-uncased, " "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese." ) parser.add_argument( "--max_seq_length", default=512, type=int, help= "The maximum total input sequence length after WordPiece tokenization. \n" "Sequences longer than this will be truncated, and sequences shorter \n" "than this will be padded.") parser.add_argument( "--max_predictions_per_seq", default=80, type=int, help="The maximum total of masked tokens in input sequence") parser.add_argument("--ckpt_step", default=-1, type=int, required=False, help="The model checkpoint iteration, e.g. 1000") parser.add_argument("--eval_batch_size", default=8, type=int, help="Total batch size for training.") parser.add_argument( "--max_steps", default=-1, type=int, help= "Total number of eval steps to perform, otherwise use full dataset") parser.add_argument("--no_cuda", default=False, action='store_true', help="Whether not to use CUDA when available") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument( '--fp16', default=False, action='store_true', help="Whether to use 16-bit float precision instead of 32-bit") args = parser.parse_args() if args.local_rank == -1 or args.no_cuda: device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") else: torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.distributed.init_process_group(backend='nccl', init_method='env://') n_gpu = torch.cuda.device_count() if n_gpu > 1: assert (args.local_rank != -1 ) # only use torch.distributed for multi-gpu logger.info("device %s n_gpu %d distributed inference %r", device, n_gpu, bool(args.local_rank != -1)) random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if n_gpu > 0: torch.cuda.manual_seed_all(args.seed) # Prepare model config = BertConfig.from_json_file(args.config_file) model = BertForPreTraining(config) if args.ckpt_step == -1: #retrieve latest model model_names = [ f for f in os.listdir(args.ckpt_dir) if f.endswith(".model") ] args.ckpt_step = max([ int(x.split('.model')[0].split('_')[1].strip()) for x in model_names ]) print("load model saved at iteraton", args.ckpt_step) model_file = os.path.join(args.ckpt_dir, "ckpt_" + str(args.ckpt_step) + ".model") state_dict = torch.load(model_file, map_location="cpu") model.load_state_dict(state_dict, strict=False) if args.fp16: model.half( ) # all parameters and buffers are converted to half precision model.to(device) multi_gpu_training = args.local_rank != -1 and torch.distributed.is_initialized( ) if multi_gpu_training: model = DDP(model) files = [ os.path.join(args.input_dir, f) for f in os.listdir(args.input_dir) if os.path.isfile(os.path.join(args.input_dir, f)) ] files.sort() logger.info("***** Running evaluation *****") logger.info(" Batch size = %d", args.eval_batch_size) model.eval() print("Evaluation. . .") nb_instances = 0 max_steps = args.max_steps if args.max_steps > 0 else np.inf global_step = 0 with torch.no_grad(): if args.do_eval: final_loss = 0.0 # for data_file in files: logger.info("file %s" % (data_file)) dataset = pretraining_dataset( input_file=data_file, max_pred_length=args.max_predictions_per_seq) if not multi_gpu_training: train_sampler = RandomSampler(dataset) datasetloader = DataLoader(dataset, sampler=train_sampler, batch_size=args.eval_batch_size, num_workers=4, pin_memory=True) else: train_sampler = DistributedSampler(dataset) datasetloader = DataLoader(dataset, sampler=train_sampler, batch_size=args.eval_batch_size, num_workers=4, pin_memory=True) for step, batch in enumerate( tqdm(datasetloader, desc="Iteration")): if global_step > max_steps: break batch = [t.to(device) for t in batch] input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels = batch #\ loss = model(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask, masked_lm_labels=masked_lm_labels, next_sentence_label=next_sentence_labels) final_loss += loss global_step += 1 torch.cuda.empty_cache() if global_step > max_steps: break final_loss /= global_step if multi_gpu_training: final_loss /= torch.distributed.get_world_size() dist.all_reduce(final_loss) if (not multi_gpu_training or (multi_gpu_training and torch.distributed.get_rank() == 0)): logger.info("Finished: Final Loss = {}".format(final_loss)) else: # inference # if multi_gpu_training: # torch.distributed.barrier() # start_t0 = time.time() for data_file in files: logger.info("file %s" % (data_file)) dataset = pretraining_dataset( input_file=data_file, max_pred_length=args.max_predictions_per_seq) if not multi_gpu_training: train_sampler = RandomSampler(dataset) datasetloader = DataLoader(dataset, sampler=train_sampler, batch_size=args.eval_batch_size, num_workers=4, pin_memory=True) else: train_sampler = DistributedSampler(dataset) datasetloader = DataLoader(dataset, sampler=train_sampler, batch_size=args.eval_batch_size, num_workers=4, pin_memory=True) for step, batch in enumerate( tqdm(datasetloader, desc="Iteration")): if global_step > max_steps: break batch = [t.to(device) for t in batch] input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels = batch #\ lm_logits, nsp_logits = model(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask, masked_lm_labels=None, next_sentence_label=None) nb_instances += input_ids.size(0) global_step += 1 torch.cuda.empty_cache() if global_step > max_steps: break # if multi_gpu_training: # torch.distributed.barrier() if (not multi_gpu_training or (multi_gpu_training and torch.distributed.get_rank() == 0)): logger.info("Finished")
def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument( "--data_dir", default=None, type=str, required=True, help= "The input data dir. Should contain the .tsv files (or other data files) for the task." ) parser.add_argument( "--output_dir", default=None, type=str, required=True, help="The output directory where the model checkpoints will be written." ) ## Other parameters parser.add_argument( "--bert_model", default='bert-base-multilingual-cased', type=str, help="Bert pre-trained model selected in the list: bert-base-uncased, " "bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, " "bert-base-multilingual-cased, bert-base-chinese.") parser.add_argument( "--max_seq_length", default=384, type=int, help= "The maximum total input sequence length after WordPiece tokenization. \n" "Sequences longer than this will be truncated, and sequences shorter \n" "than this will be padded.") parser.add_argument("--do_train", action='store_true', help="Whether to run training.") # parser.add_argument("--do_eval", # action='store_true', # help="Whether to run eval on the dev set.") parser.add_argument("--train_batch_size", default=2, type=int, help="Total batch size for training.") # parser.add_argument("--eval_batch_size", # default=2, # type=int, # help="Total batch size for eval.") parser.add_argument("--learning_rate", default=3e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument( "--warmup_proportion", default=0.1, type=float, help= "Proportion of training to perform linear learning rate warmup for. " "E.g., 0.1 = 10%% of training.") parser.add_argument("--no_cuda", action='store_true', help="Whether not to use CUDA when available") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on GPUs") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument( '--gradient_accumulation_steps', type=int, default=1, help= "Number of updates steps to accumualte before performing a backward/update pass." ) parser.add_argument( '--fp16', action='store_true', help="Whether to use 16-bit float precision instead of 32-bit") parser.add_argument( '--loss_scale', type=float, default=0, help= "Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n" "0 (default value): dynamic loss scaling.\n" "Positive power of 2: static loss scaling value.\n") parser.add_argument('--visdom', action='store_true', help='Use visdom for loss visualization') parser.add_argument('--check_saved_model', action='store_true', help='Use visdom for loss visualization') parser.add_argument('--last_final_epoch', type=int, default=-1, help="저번에 이미 최종 학습을 했고, 이에 이어서 트레이닝을 원할때 사용,\n" "기존에 train_epoch를 3으로 세팅했다면, 2가 아닌 3을 입력하세요.") args = parser.parse_args() print(args) if args.visdom: import visdom viz = visdom.Visdom() # visdom을 통해서 loss를 시각화 os.makedirs(args.output_dir, exist_ok=True) if args.local_rank == -1 or args.no_cuda: device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") n_gpu = torch.cuda.device_count() else: torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) n_gpu = 1 # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.distributed.init_process_group(backend='nccl') logger.info( "device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}". format(device, n_gpu, bool(args.local_rank != -1), args.fp16)) if args.gradient_accumulation_steps < 1: raise ValueError( "Invalid gradient_accumulation_steps parameter: {}, should be >= 1" .format(args.gradient_accumulation_steps)) args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if n_gpu > 0: torch.cuda.manual_seed_all(args.seed) if not args.do_train: raise ValueError( "Training is currently the only implemented execution option. Please set `do_train`." ) if os.path.exists(args.output_dir) and os.listdir(args.output_dir): raise ValueError( "Output directory ({}) already exists and is not empty.".format( args.output_dir)) if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=False) processor = DataProcessor() label_list = processor.get_labels() num_train_optimization_steps = None if args.do_train: print("Loading Train Dataset", args.data_dir) train_examples = processor.get_train_examples(args.data_dir) train_dataset = LazyDataset(train_examples, args.max_seq_length, tokenizer) if args.local_rank == -1: train_sampler = RandomSampler(train_dataset) else: train_sampler = DistributedSampler(train_dataset) num_train_optimization_steps = int( len(train_dataset) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs if args.local_rank != -1: num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size( ) # Prepare model loaded_epoch = -1 saved_model_path = -1 if args.last_final_epoch != -1: last_model = os.path.join(args.output_dir, WEIGHTS_NAME) if os.path.exists(last_model): saved_model_path = last_model loaded_epoch = args.last_final_epoch - 1 elif args.check_saved_model: for epoch in range(int(args.num_train_epochs)): tmp = os.path.join(args.output_dir, (f"weight_on_ep{epoch}_" + WEIGHTS_NAME)) if os.path.exists(tmp): saved_model_path = tmp loaded_epoch = epoch if saved_model_path != -1: logger.info(f"Loading on saved model {saved_model_path}") config_file = os.path.join(args.output_dir, CONFIG_NAME) config = BertConfig(config_file) logger.info("Model config {}".format(config)) model = BertForPreTraining(config) model.load_state_dict(torch.load(saved_model_path)) else: loaded_epoch = -1 model = BertForPreTraining.from_pretrained(args.bert_model) if args.fp16: model.half() model.to(device) if args.local_rank != -1: try: from apex.parallel import DistributedDataParallel as DDP except ImportError: raise ImportError( "Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training." ) model = DDP(model) elif n_gpu > 1: model = torch.nn.DataParallel(model) # Prepare optimizer param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [{ 'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01 }, { 'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0 }] if args.fp16: try: from apex.optimizers import FP16_Optimizer from apex.optimizers import FusedAdam except ImportError: raise ImportError( "Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training." ) optimizer = FusedAdam(optimizer_grouped_parameters, lr=args.learning_rate, bias_correction=False, max_grad_norm=1.0) if args.loss_scale == 0: optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True) else: optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale) else: optimizer = BertAdam(optimizer_grouped_parameters, lr=args.learning_rate, warmup=args.warmup_proportion, t_total=num_train_optimization_steps) if args.visdom: # 일단 visdom 기본 figure를 정의 vis_title = f'Baseline on {len(train_dataset)} dataset' vis_legend = ['LM Loss', 'Click Loss', 'Total Loss'] iter_plot = create_vis_plot(viz, 'Iteration', 'Loss', vis_title, vis_legend) epoch_plot = create_vis_plot(viz, 'Epoch', 'Loss', vis_title, vis_legend) # if args.do_eval: # eval_examples = processor.get_dev_examples(args.data_dir) # # logger.info("***** Running evaluation *****") # logger.info(" Num examples = %d", len(eval_examples)) # logger.info(" Batch size = %d", args.eval_batch_size) # # eval_data = LazyDatasetClassifier(eval_examples, label_list, args.max_seq_length, tokenizer) # # Run prediction for full data # """ # cur_tensors = (torch.tensor(f.input_ids), # torch.tensor(f.input_mask), # torch.tensor(f.segment_ids), # torch.tensor(f.lm_label_ids), # torch.tensor(f.label)) # """ # eval_sampler = SequentialSampler(eval_data) # eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size) # save_eval_loss = [] global_step = 0 if args.do_train: logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_dataset)) logger.info(" Batch size = %d", args.train_batch_size) logger.info(" Num steps = %d", num_train_optimization_steps) train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size) """ cur_tensors = (torch.tensor(f.input_ids), torch.tensor(f.input_mask), torch.tensor(f.segment_ids), torch.tensor(f.lm_label_ids), torch.tensor(f.label)) """ save_loss = [] save_epoch_loss = [] save_step = int(len(train_dataloader) // 5) for epoch in trange((loaded_epoch + 1), int(args.num_train_epochs), desc="Epoch"): # if args.do_eval and loaded_epoch != -1: # model.eval() # eval_loss, eval_accuracy = 0, 0 # nb_eval_steps, nb_eval_examples = 0, 0 # # for batch in tqdm(eval_dataloader, desc="Evaluating"): # batch = tuple(t.to(device) for t in batch) # input_ids, input_mask, segment_ids, label_ids = batch # # with torch.no_grad(): # tmp_eval_loss = model(input_ids, segment_ids, input_mask, None, label_ids) # prediction_scores, logits = model(input_ids, segment_ids, input_mask) # # if n_gpu > 1: # tmp_eval_loss = tmp_eval_loss.mean() # mean() to average on multi-gpu. # # logits = logits.detach().cpu().numpy() # label_ids = label_ids.to('cpu').numpy() # tmp_eval_accuracy = accuracy(logits, label_ids) # # eval_loss += tmp_eval_loss.mean().item() # eval_accuracy += tmp_eval_accuracy # # nb_eval_examples += input_ids.size(0) # nb_eval_steps += 1 # # eval_loss = eval_loss / nb_eval_steps # eval_accuracy = eval_accuracy / nb_eval_examples # result = {'eval_loss': eval_loss, # 'eval_accuracy': eval_accuracy, # 'global_step': global_step} # # save_eval_loss.append(eval_loss) # # output_eval_file = os.path.join(args.output_dir, f"Epoch_{epoch}_eval_results.txt") # with open(output_eval_file, "w") as writer: # logger.info(f"***** Eval results on Epoch {epoch} *****") # for key in sorted(result.keys()): # logger.info(" %s = %s", key, str(result[key])) # writer.write("%s = %s\n" % (key, str(result[key]))) model.train() tr_loss = 0 nb_tr_examples, nb_tr_steps = 0, 0 tr_loss_ml = 0 tr_loss_click = 0 for step, batch in enumerate( tqdm(train_dataloader, desc="Iteration")): batch = tuple(t.to(device) for t in batch) input_ids, input_mask, segment_ids, lm_label_ids, label = batch # if global_step == 0: # print(input_ids.shape, input_mask.shape, segment_ids.shape, lm_label_ids.shape, label.shape) loss, loss_ml, loss_click = model(input_ids, segment_ids, input_mask, lm_label_ids, label) if n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu. loss_ml = loss_ml.mean() loss_click = loss_click.mean() if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps loss_ml = loss_ml / args.gradient_accumulation_steps loss_click = loss_click / args.gradient_accumulation_steps if args.fp16: optimizer.backward(loss) else: loss.backward() tr_loss += loss.item() tr_loss_ml += loss_ml.item() tr_loss_click += loss_click.item() nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: # modify learning rate with special warm up BERT uses # if args.fp16 is False, BertAdam is used that handles this automatically lr_this_step = args.learning_rate * warmup_linear( global_step / num_train_optimization_steps, args.warmup_proportion) for param_group in optimizer.param_groups: param_group['lr'] = lr_this_step optimizer.step() optimizer.zero_grad() global_step += 1 if global_step != 0 and global_step % save_step == 0: # 한 에포치당 5번 저장 logger.info(f'Saving state, iter: {global_step}') model_to_save = model.module if hasattr( model, 'module') else model # Only save the model it-self model_name = f"weight_on_{global_step}_" + WEIGHTS_NAME output_model_file = os.path.join(args.output_dir, model_name) torch.save(model_to_save.state_dict(), output_model_file) output_config_file = os.path.join(args.output_dir, CONFIG_NAME) with open(output_config_file, 'w') as f: f.write(model_to_save.config.to_json_string()) print("Loss at ", global_step, loss_ml.item(), loss_click.item(), loss.item()) save_loss.append( [loss_ml.item(), loss_click.item(), loss.item()]) if args.visdom: update_vis_plot(viz, global_step, loss_ml.item(), loss_click.item(), iter_plot, epoch_plot, 'append') if epoch != (int(args.num_train_epochs) - 1): # 각 에포치가 끝날때 마다 저장 logger.info(f'Saving state, epoch: {epoch}') model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self model_name = f"weight_on_ep{epoch}_" + WEIGHTS_NAME output_model_file = os.path.join(args.output_dir, model_name) torch.save(model_to_save.state_dict(), output_model_file) output_config_file = os.path.join(args.output_dir, CONFIG_NAME) with open(output_config_file, 'w') as f: f.write(model_to_save.config.to_json_string()) print("Loss at epoch", epoch, tr_loss_ml, tr_loss_click, tr_loss) save_epoch_loss.append([tr_loss_ml, tr_loss_click, tr_loss]) if args.visdom: update_vis_plot(viz, epoch, tr_loss_ml, tr_loss_click, epoch_plot, None, 'append', len(train_dataset) // args.train_batch_size) # if args.do_eval and loaded_epoch == -1: # # model.eval() # eval_loss, eval_accuracy = 0, 0 # nb_eval_steps, nb_eval_examples = 0, 0 # # for batch in tqdm(eval_dataloader, desc="Evaluating"): # batch = tuple(t.to(device) for t in batch) # input_ids, input_mask, segment_ids, label_ids = batch # # with torch.no_grad(): # tmp_eval_loss = model(input_ids, segment_ids, input_mask, None, label_ids) # prediction_scores, logits = model(input_ids, segment_ids, input_mask) # # if n_gpu > 1: # tmp_eval_loss = tmp_eval_loss.mean() # mean() to average on multi-gpu. # # logits = logits.detach().cpu().numpy() # label_ids = label_ids.to('cpu').numpy() # tmp_eval_accuracy = accuracy(logits, label_ids) # # eval_loss += tmp_eval_loss.mean().item() # eval_accuracy += tmp_eval_accuracy # # nb_eval_examples += input_ids.size(0) # nb_eval_steps += 1 # # eval_loss = eval_loss / nb_eval_steps # eval_accuracy = eval_accuracy / nb_eval_examples # result = {'eval_loss': eval_loss, # 'eval_accuracy': eval_accuracy, # 'global_step': global_step} # # save_eval_loss.append(eval_loss) # # output_eval_file = os.path.join(args.output_dir, f"Epoch_{epoch}_eval_results.txt") # with open(output_eval_file, "w") as writer: # logger.info(f"***** Eval results on Epoch {epoch} *****") # for key in sorted(result.keys()): # logger.info(" %s = %s", key, str(result[key])) # writer.write("%s = %s\n" % (key, str(result[key]))) # Save a trained model logger.info("** ** * Saving fine - tuned model ** ** * ") # model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self # output_model_file = os.path.join(args.output_dir, "pytorch_model.bin") # if args.do_train: # torch.save(model_to_save.state_dict(), output_model_file) save_loss = np.array(save_loss) save_epoch_loss = np.array(save_epoch_loss) np.save(os.path.join(args.output_dir, "save_loss.npy"), save_loss) np.save(os.path.join(args.output_dir, "save_epoch_loss.npy"), save_epoch_loss) # if args.do_eval: # save_eval_loss = np.array(save_eval_loss) # np.save(os.path.join(args.output_dir, "save_eval_loss.npy"), save_eval_loss) model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME) torch.save(model_to_save.state_dict(), output_model_file) output_config_file = os.path.join(args.output_dir, CONFIG_NAME) with open(output_config_file, 'w') as f: f.write(model_to_save.config.to_json_string())
def prepare_model_and_optimizer(args, device): global_step = 0 args.resume_step = 0 checkpoint = None config = BertConfig.from_json_file(args.bert_config_path) config.fused_mha = args.fused_mha config.fused_gelu_bias = args.fused_gelu_bias config.dense_seq_output = args.dense_seq_output config.unpad = args.unpad config.pad = args.pad config.fuse_qkv = not args.disable_fuse_qkv config.fuse_scale = not args.disable_fuse_scale config.fuse_mask = not args.disable_fuse_mask config.fuse_dropout = args.enable_fuse_dropout config.apex_softmax = not args.disable_apex_softmax config.enable_stream = args.enable_stream if config.fuse_mask == True: config.apex_softmax = True if config.pad == False: config.enable_stream = True if config.unpad == True: config.fused_mha = False # Padding for divisibility by 8 if config.vocab_size % 8 != 0: config.vocab_size += 8 - (config.vocab_size % 8) # Load from Pyt checkpoint - either given as init_checkpoint, or picked up from output_dir if found if args.init_checkpoint is not None or found_resume_checkpoint(args): # Prepare model model = BertForPreTraining(config) if args.init_checkpoint is None: # finding checkpoint in output_dir checkpoint_str = "phase2_ckpt_*.pt" if args.phase2 else "phase1_ckpt_*.pt" model_names = [f for f in glob.glob(os.path.join(args.output_dir, checkpoint_str))] global_step = max([int(x.split('.pt')[0].split('_')[-1].strip()) for x in model_names]) args.resume_step = global_step #used for throughput computation resume_init_checkpoint = os.path.join(args.output_dir, checkpoint_str.replace("*", str(global_step))) print("Setting init checkpoint to %s - which is the latest in %s" %(resume_init_checkpoint, args.output_dir)) checkpoint=torch.load(resume_init_checkpoint, map_location="cpu") else: checkpoint=torch.load(args.init_checkpoint, map_location="cpu")["model"] # Fused MHA requires a remapping of checkpoint parameters if config.fused_mha: checkpoint_remapped = remap_attn_parameters(checkpoint) model.load_state_dict(checkpoint_remapped, strict=False) else: model.load_state_dict(checkpoint, strict=True) else: #Load from TF Checkpoint model = BertForPreTraining.from_pretrained(args.init_tf_checkpoint, from_tf=True, config=config) model.to(device) param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'gamma', 'beta', 'LayerNorm'] optimizer_grouped_parameters = [ {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay_rate}, {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}] mlperf_logger.log_event(key=mlperf_logger.constants.OPT_BASE_LR, value=args.learning_rate, sync=False) optimizer = FusedLAMB(optimizer_grouped_parameters, lr=args.learning_rate, betas=(args.opt_lamb_beta_1, args.opt_lamb_beta_2)) mlperf_logger.log_event(key='opt_epsilon', value=optimizer.defaults['eps'], sync=False) b1, b2 = optimizer.defaults['betas'] mlperf_logger.log_event(key='opt_lamb_beta_1', value=b1, sync=False) mlperf_logger.log_event(key='opt_lamb_beta_2', value=b2, sync=False) mlperf_logger.log_event(key='opt_lamb_weight_decay_rate', value=optimizer.defaults['weight_decay'], sync=False) if args.warmup_steps == 0: warmup_steps = int(args.max_steps * args.warmup_proportion) warmup_start = 0 else: warmup_steps = args.warmup_steps warmup_start = args.start_warmup_step lr_scheduler = LinearWarmupPolyDecayScheduler(optimizer, start_warmup_steps=warmup_start, warmup_steps=warmup_steps, total_steps=args.max_steps, end_learning_rate=0.0, degree=1.0) if args.fp16: if args.loss_scale == 0: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale="dynamic") else: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale=args.loss_scale) amp._amp_state.loss_scalers[0]._loss_scale = float(os.getenv("INIT_LOSS_SCALE", 2**20)) if found_resume_checkpoint(args): optimizer.load_state_dict(checkpoint['optimizer']) #restores m,v states (only if resuming checkpoint, not for init_checkpoint and init_tf_checkpoint for now) # Restore AMP master parameters if args.fp16: optimizer._lazy_init_maybe_master_weights() optimizer._amp_stash.lazy_init_called = True optimizer.load_state_dict(checkpoint['optimizer']) for param, saved_param in zip(amp.master_params(optimizer), checkpoint['master params']): param.data.copy_(saved_param.data) if args.local_rank != -1: if not args.allreduce_post_accumulation: model = DDP(model, message_size=250000000, gradient_predivide_factor=torch.distributed.get_world_size()) else: flat_dist_call([param.data for param in model.parameters()], torch.distributed.broadcast, (0,) ) return model, optimizer, lr_scheduler, checkpoint, global_step
def inference(args): start_t = time.time() bert_module = BertForPreTraining( args.vocab_size, args.seq_length, args.hidden_size, args.num_hidden_layers, args.num_attention_heads, args.intermediate_size, nn.GELU(), args.hidden_dropout_prob, args.attention_probs_dropout_prob, args.max_position_embeddings, args.type_vocab_size, args.vocab_size, ) end_t = time.time() print("Initialize model using time: {:.3f}s".format(end_t - start_t)) start_t = time.time() if args.use_lazy_model: from utils.compare_lazy_outputs import load_params_from_lazy load_params_from_lazy( bert_module.state_dict(), args.model_path, ) else: bert_module.load_state_dict(flow.load(args.model_path)) end_t = time.time() print("Loading parameters using time: {:.3f}s".format(end_t - start_t)) bert_module.eval() bert_module.to(args.device) class BertEvalGraph(nn.Graph): def __init__(self): super().__init__() self.bert = bert_module def build(self, input_ids, input_masks, segment_ids): input_ids = input_ids.to(device=args.device) input_masks = input_masks.to(device=args.device) segment_ids = segment_ids.to(device=args.device) with flow.no_grad(): # 1. forward the next_sentence_prediction and masked_lm model _, seq_relationship_scores = self.bert(input_ids, input_masks, segment_ids) return seq_relationship_scores bert_eval_graph = BertEvalGraph() start_t = time.time() inputs = [np.random.randint(0, 20, size=args.seq_length)] inputs = flow.Tensor(inputs, dtype=flow.int64, device=flow.device(args.device)) mask = flow.cast(inputs > 0, dtype=flow.int64) segment_info = flow.zeros_like(inputs) prediction = bert_eval_graph(inputs, mask, segment_info) print(prediction.numpy()) end_t = time.time() print("Inference using time: {:.3f}".format(end_t - start_t))