|
| 1 | +import time |
| 2 | +import random |
| 3 | +import os |
| 4 | +import csv |
| 5 | + |
| 6 | +import torch |
| 7 | +import numpy as np |
| 8 | +from squad_metric import squad_metric |
| 9 | + |
| 10 | +import bmtrain as bmt |
| 11 | + |
| 12 | +from model_center import get_args |
| 13 | +from model_center.model import T5 |
| 14 | +from model_center.generation.t5 import T5BeamSearch |
| 15 | +from model_center.tokenizer import T5Tokenizer |
| 16 | +from model_center.dataset.t5dataset import DATASET |
| 17 | +from model_center.utils import print_inspect |
| 18 | +from model_center.dataset import DistributedDataLoader |
| 19 | +from torch.utils.data import DataLoader |
| 20 | + |
| 21 | + |
| 22 | +def get_tokenizer(args): |
| 23 | + tokenizer = T5Tokenizer.from_pretrained(args.model_config) |
| 24 | + return tokenizer |
| 25 | + |
| 26 | +def get_model(args): |
| 27 | + model = T5.from_pretrained(args.model_config) |
| 28 | + return model |
| 29 | + |
| 30 | +def get_optimizer(args, model): |
| 31 | + optimizer = bmt.optim.AdamOffloadOptimizer(model.parameters(), weight_decay=args.weight_decay) |
| 32 | + return optimizer |
| 33 | + |
| 34 | +def get_learning_rate_scheduler(args, optimizer): |
| 35 | + if args.lr_decay_iters is None: |
| 36 | + args.lr_decay_iters = args.train_iters * args.epochs |
| 37 | + if args.lr_decay_style == "noam": |
| 38 | + lr_scheduler = bmt.lr_scheduler.Noam(optimizer, |
| 39 | + start_lr = args.lr, |
| 40 | + warmup_iter = args.warmup_iters, |
| 41 | + end_iter = args.lr_decay_iters, |
| 42 | + num_iter = args.start_step) |
| 43 | + elif args.lr_decay_style == "constant": |
| 44 | + lr_scheduler = bmt.lr_scheduler.NoDecay(optimizer, |
| 45 | + start_lr = args.lr, |
| 46 | + warmup_iter = args.warmup_iters, |
| 47 | + end_iter = -1, |
| 48 | + num_iter = args.start_step) |
| 49 | + elif args.lr_decay_style == "linear": |
| 50 | + lr_scheduler = bmt.lr_scheduler.Linear(optimizer, |
| 51 | + start_lr = args.lr, |
| 52 | + warmup_iter = args.warmup_iters, |
| 53 | + end_iter = args.lr_decay_iters, |
| 54 | + num_iter = args.start_step) |
| 55 | + elif args.lr_decay_style == "exponential": |
| 56 | + lr_scheduler = bmt.lr_scheduler.Exponential(optimizer, |
| 57 | + start_lr = args.lr, |
| 58 | + warmup_iter = args.warmup_iters, |
| 59 | + end_iter = args.lr_decay_iters, |
| 60 | + num_iter = args.start_step) |
| 61 | + elif args.lr_decay_style == "cosine": |
| 62 | + lr_scheduler = bmt.lr_scheduler.Cosine(optimizer, |
| 63 | + start_lr = args.lr, |
| 64 | + warmup_iter = args.warmup_iters, |
| 65 | + end_iter = args.lr_decay_iters, |
| 66 | + num_iter = args.start_step) |
| 67 | + else: |
| 68 | + raise ValueError(f"lr_scheduler of type {args.lr_decay_style} is not supported yet.") |
| 69 | + |
| 70 | + return lr_scheduler |
| 71 | + |
| 72 | +def setup_model_and_optimizer(args): |
| 73 | + # get the tokenizer |
| 74 | + tokenizer = get_tokenizer(args) |
| 75 | + # get the model |
| 76 | + model = get_model(args) |
| 77 | + bmt.synchronize() |
| 78 | + # get the optimizer and lr_scheduler |
| 79 | + optimizer = get_optimizer(args, model) |
| 80 | + lr_scheduler = get_learning_rate_scheduler(args, optimizer) |
| 81 | + bmt.synchronize() |
| 82 | + # get the memory usage |
| 83 | + bmt.print_rank("Model mem\n", torch.cuda.memory_summary()) |
| 84 | + bmt.synchronize() |
| 85 | + return tokenizer, model, optimizer, lr_scheduler |
| 86 | + |
| 87 | +def initialize(): |
| 88 | + # get arguments |
| 89 | + args = get_args() |
| 90 | + # init bmt |
| 91 | + bmt.init_distributed(seed = args.seed) |
| 92 | + # init save folder |
| 93 | + if args.save != None: |
| 94 | + os.makedirs(args.save, exist_ok=True) |
| 95 | + return args |
| 96 | + |
| 97 | + |
| 98 | +def prepare_dataset(args, tokenizer, base_path, dataset_name): |
| 99 | + splits = ['train', 'dev', 'test'] |
| 100 | + dataset = {} |
| 101 | + for split in splits: |
| 102 | + dataset[split] = DATASET[dataset_name](base_path, split, tokenizer, args.max_encoder_length, args.max_decoder_length) |
| 103 | + return dataset |
| 104 | + |
| 105 | +def collate_fn(data): |
| 106 | + # data: a list of tuples with (input, target) |
| 107 | + return { |
| 108 | + "inputs" : [d['inputs'] for d in data], |
| 109 | + "targets": [d['targets'] for d in data], |
| 110 | + } |
| 111 | + |
| 112 | +def finetune(args, tokenizer, model, optimizer, lr_scheduler, dataset): |
| 113 | + loss_func = bmt.loss.FusedCrossEntropy(ignore_index=-100) |
| 114 | + |
| 115 | + optim_manager = bmt.optim.OptimManager(loss_scale=args.loss_scale, loss_scale_steps=100) |
| 116 | + optim_manager.add_optimizer(optimizer, lr_scheduler) |
| 117 | + |
| 118 | + # print_inspect(model, '*') |
| 119 | + |
| 120 | + for epoch in range(20): |
| 121 | + dataloader = { |
| 122 | + "train": DistributedDataLoader(dataset['train'], batch_size=args.batch_size, shuffle=True), |
| 123 | + "dev": DataLoader(dataset['dev'], batch_size=args.batch_size, shuffle=False, collate_fn=collate_fn), |
| 124 | + } |
| 125 | + |
| 126 | + model.train() |
| 127 | + for it, data in enumerate(dataloader['train']): |
| 128 | + logits = model( |
| 129 | + input_ids = data['input_ids'], |
| 130 | + attention_mask = data['attention_mask'], |
| 131 | + decoder_input_ids = data['decoder_input_ids'], |
| 132 | + decoder_attention_mask = data['decoder_attention_mask'], |
| 133 | + ).logits |
| 134 | + targets = data["targets"] |
| 135 | + |
| 136 | + loss = loss_func(logits.view(-1, logits.shape[-1]), targets.view(-1)) |
| 137 | + global_loss = bmt.sum_loss(loss).item() |
| 138 | + |
| 139 | + optim_manager.zero_grad() |
| 140 | + |
| 141 | + optim_manager.backward(loss) |
| 142 | + grad_norm = optim_manager.clip_grad_norm(optimizer.param_groups, args.clip_grad, norm_type = 2) |
| 143 | + |
| 144 | + optim_manager.step() |
| 145 | + |
| 146 | + bmt.print_rank( |
| 147 | + "train | epoch {:3d} | Iter: {:6d}/{:6d} | loss: {:.4f} | lr: {:.4e}, scale: {:10.4f} | grad_norm: {:.4f} |".format( |
| 148 | + epoch, |
| 149 | + it, |
| 150 | + len(dataloader["train"]), |
| 151 | + global_loss, |
| 152 | + lr_scheduler.current_lr, |
| 153 | + int(optim_manager.loss_scale), |
| 154 | + grad_norm, |
| 155 | + ) |
| 156 | + ) |
| 157 | + # if it % args.inspect_iters == 0: print_inspect(model, "*") |
| 158 | + # if args.save != None and it % args.save_iters == 0: |
| 159 | + # bmt.save(model, os.path.join(args.save, args.save_name+("-%d.pt" % it))) |
| 160 | + |
| 161 | + model.eval() |
| 162 | + beam_search = T5BeamSearch( |
| 163 | + model=model, |
| 164 | + tokenizer=tokenizer, |
| 165 | + ) |
| 166 | + with torch.no_grad(): |
| 167 | + for split in ['dev']: |
| 168 | + pd = [] |
| 169 | + gt = [] |
| 170 | + for it, data in enumerate(dataloader[split]): |
| 171 | + preds = beam_search.generate(data['inputs'], max_length=args.max_decoder_length) |
| 172 | + targets = data["targets"] |
| 173 | + |
| 174 | + pd.extend(preds) |
| 175 | + gt.extend(targets) |
| 176 | + |
| 177 | + bmt.print_rank( |
| 178 | + "{} | epoch {:3d} | Iter: {:6d}/{:6d} |".format( |
| 179 | + split, |
| 180 | + epoch, |
| 181 | + it, |
| 182 | + len(dataloader[split]), |
| 183 | + ) |
| 184 | + ) |
| 185 | + |
| 186 | + metrics = squad_metric(pd, gt, None) |
| 187 | + bmt.print_rank(f"metrics: {metrics}") |
| 188 | + |
| 189 | + |
| 190 | +def main(): |
| 191 | + args = initialize() |
| 192 | + tokenizer, model, optimizer, lr_scheduler = setup_model_and_optimizer(args) |
| 193 | + dataset = prepare_dataset( |
| 194 | + args, |
| 195 | + tokenizer, |
| 196 | + f"{args.base_path}/down_data/squad/", |
| 197 | + args.dataset_name, |
| 198 | + ) |
| 199 | + finetune(args, tokenizer, model, optimizer, lr_scheduler, dataset) |
| 200 | + |
| 201 | +if __name__ == "__main__": |
| 202 | + main() |
0 commit comments