train_gradient_cache.py

# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
import os
import random
import time
from functools import partial

import numpy as np
import paddle
import paddle.nn.functional as F
from data import convert_example, create_dataloader, read_text_pair
from gradient_cache.model import SemanticIndexCacheNeg

import paddlenlp as ppnlp
from paddlenlp.data import Pad, Tuple
from paddlenlp.datasets import load_dataset
from paddlenlp.transformers import LinearDecayWithWarmup

# yapf: disable
parser = argparse.ArgumentParser()
parser.add_argument("--save_dir", default='./checkpoint', type=str, help="The output directory where the model checkpoints will be written.")
parser.add_argument("--max_seq_length", default=128, type=int, help="The maximum total input sequence length after tokenization. Sequences longer than this will be truncated, sequences shorter will be padded.")
parser.add_argument("--batch_size", default=32, type=int, help="Batch size per GPU/CPU for training.")
parser.add_argument("--output_emb_size", default=None, type=int, help="output_embedding_size.")
parser.add_argument("--learning_rate", default=1e-5, type=float, help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
parser.add_argument("--epochs", default=10, type=int, help="Total number of training epochs to perform.")
parser.add_argument("--warmup_proportion", default=0.0, type=float, help="Linear warmup proportion over the training process.")
parser.add_argument("--init_from_ckpt", type=str, default=None, help="The path of checkpoint to be loaded.")
parser.add_argument("--seed", type=int, default=1000, help="random seed for initialization.")
parser.add_argument('--device', choices=['cpu', 'gpu'], default="gpu", help="Select which device to train model, defaults to gpu.")
parser.add_argument('--save_steps', type=int, default=10000, help="Inteval steps to save checkpoint.")
parser.add_argument("--train_set_file", type=str, required=True, help="The full path of train_set_file.")
parser.add_argument("--margin", default=0.3, type=float, help="Margin between pos_sample and neg_samples.")
parser.add_argument("--scale", default=30, type=int, help="Scale for pair-wise margin_rank_loss")
parser.add_argument("--use_amp", action="store_true", help="Whether to use AMP.")
parser.add_argument("--amp_loss_scale", default=32768, type=float, help="The value of scale_loss for fp16. This is only used for AMP training.")
parser.add_argument("--chunk_numbers", type=int, default=50, help="The number of the chunks for model")

args = parser.parse_args()


# yapf: enable


def set_seed(seed):
    """sets random seed"""
    random.seed(seed)
    np.random.seed(seed)
    paddle.seed(seed)


def do_train():
    paddle.set_device(args.device)
    rank = paddle.distributed.get_rank()
    if paddle.distributed.get_world_size() > 1:
        paddle.distributed.init_parallel_env()

    random.seed(args.seed)
    np.random.seed(args.seed)
    paddle.seed(args.seed)

    train_ds = load_dataset(read_text_pair, data_path=args.train_set_file, lazy=False)

    # If you wanna use bert/roberta pretrained model,
    # pretrained_model = ppnlp.transformers.BertModel.from_pretrained('bert-base-chinese')
    # pretrained_model = ppnlp.transformers.RobertaModel.from_pretrained('roberta-wwm-ext')
    pretrained_model = ppnlp.transformers.ErnieModel.from_pretrained("ernie-1.0")

    # If you wanna use bert/roberta pretrained model,
    # tokenizer = ppnlp.transformers.BertTokenizer.from_pretrained('bert-base-chinese')
    # tokenizer = ppnlp.transformers.RobertaTokenizer.from_pretrained('roberta-wwm-ext')
    tokenizer = ppnlp.transformers.ErnieTokenizer.from_pretrained("ernie-1.0")

    trans_func = partial(convert_example, tokenizer=tokenizer, max_seq_length=args.max_seq_length)

    batchify_fn = lambda samples, fn=Tuple(
        Pad(axis=0, pad_val=tokenizer.pad_token_id, dtype="int64"),  # query_input
        Pad(axis=0, pad_val=tokenizer.pad_token_type_id, dtype="int64"),  # query_# query_segment
        Pad(axis=0, pad_val=tokenizer.pad_token_id, dtype="int64"),  # query_# title_input
        Pad(axis=0, pad_val=tokenizer.pad_token_type_id, dtype="int64"),  # title_segment
    ): [data for data in fn(samples)]

    train_data_loader = create_dataloader(
        train_ds, mode="train", batch_size=args.batch_size, batchify_fn=batchify_fn, trans_fn=trans_func
    )

    model = SemanticIndexCacheNeg(
        pretrained_model, margin=args.margin, scale=args.scale, output_emb_size=args.output_emb_size
    )

    if args.init_from_ckpt and os.path.isfile(args.init_from_ckpt):
        state_dict = paddle.load(args.init_from_ckpt)
        model.set_dict(state_dict)
        print("warmup from:{}".format(args.init_from_ckpt))
    model = paddle.DataParallel(model)
    num_training_steps = len(train_data_loader) * args.epochs
    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps, args.warmup_proportion)

    # Generate parameter names needed to perform weight decay.
    # All bias and LayerNorm parameters are excluded.
    decay_params = [p.name for n, p in model.named_parameters() if not any(nd in n for nd in ["bias", "norm"])]
    optimizer = paddle.optimizer.AdamW(
        learning_rate=lr_scheduler,
        parameters=model.parameters(),
        weight_decay=args.weight_decay,
        apply_decay_param_fun=lambda x: x in decay_params,
    )

    if args.use_amp:
        scaler = paddle.amp.GradScaler(init_loss_scaling=args.amp_loss_scale)

    if args.batch_size % args.chunk_numbers == 0:
        chunk_numbers = args.chunk_numbers

    def split(inputs, chunk_numbers, axis=0):
        if inputs.shape[0] % chunk_numbers == 0:
            return paddle.split(inputs, chunk_numbers, axis=0)
        else:
            return paddle.split(inputs, inputs.shape[0], axis=0)

    global_step = 0
    tic_train = time.time()
    for epoch in range(1, args.epochs + 1):
        for step, batch in enumerate(train_data_loader, start=1):
            chunked_x = [split(t, chunk_numbers, axis=0) for t in batch]
            sub_batchs = [list(s) for s in zip(*chunked_x)]

            all_reps = []
            all_grads = []
            all_labels = []
            all_CUDA_rnd_state = []
            all_query = []
            all_title = []

            for sub_batch in sub_batchs:
                all_reps = []
                all_labels = []
                (
                    sub_query_input_ids,
                    sub_query_token_type_ids,
                    sub_title_input_ids,
                    sub_title_token_type_ids,
                ) = sub_batch
                with paddle.amp.auto_cast(args.use_amp, custom_white_list=["layer_norm", "softmax", "gelu"]):

                    with paddle.no_grad():
                        sub_CUDA_rnd_state = paddle.framework.random.get_cuda_rng_state()
                        all_CUDA_rnd_state.append(sub_CUDA_rnd_state)
                        sub_cosine_sim, sub_label, query_embedding, title_embedding = model(
                            query_input_ids=sub_query_input_ids,
                            title_input_ids=sub_title_input_ids,
                            query_token_type_ids=sub_query_token_type_ids,
                            title_token_type_ids=sub_title_token_type_ids,
                        )
                        all_reps.append(sub_cosine_sim)
                        all_labels.append(sub_label)
                        all_title.append(title_embedding)
                        all_query.append(query_embedding)

                model_reps = paddle.concat(all_reps, axis=0)
                model_title = paddle.concat(all_title)
                model_query = paddle.concat(all_query)

                model_title = model_title.detach()
                model_query = model_query.detach()

                model_query.stop_gtadient = False
                model_title.stop_gradient = False
                model_reps.stop_gradient = False

                model_label = paddle.concat(all_labels, axis=0)
                loss = F.cross_entropy(input=model_reps, label=model_label)
                loss.backward()
                all_grads.append(model_reps.grad)

            for sub_batch, CUDA_state, grad in zip(sub_batchs, all_CUDA_rnd_state, all_grads):

                (
                    sub_query_input_ids,
                    sub_query_token_type_ids,
                    sub_title_input_ids,
                    sub_title_token_type_ids,
                ) = sub_batch
                paddle.framework.random.set_cuda_rng_state(CUDA_state)
                cosine_sim, _ = model(
                    query_input_ids=sub_query_input_ids,
                    title_input_ids=sub_title_input_ids,
                    query_token_type_ids=sub_query_token_type_ids,
                    title_token_type_ids=sub_title_token_type_ids,
                )
                surrogate = paddle.dot(cosine_sim, grad)

                if args.use_amp:
                    scaled = scaler.scale(surrogate)
                    scaled.backward()
                else:
                    surrogate.backward()

            if args.use_amp:
                scaler.minimize(optimizer, scaled)
            else:
                optimizer.step()

            global_step += 1
            if global_step % 10 == 0 and rank == 0:
                print(
                    "global step %d, epoch: %d, batch: %d, loss: %.5f, speed: %.2f step/s"
                    % (global_step, epoch, step, loss, 10 / (time.time() - tic_train))
                )
                tic_train = time.time()

            lr_scheduler.step()
            optimizer.clear_grad()

            if global_step % args.save_steps == 0 and rank == 0:
                save_dir = os.path.join(args.save_dir, "model_%d" % global_step)
                if not os.path.exists(save_dir):
                    os.makedirs(save_dir)
                save_param_path = os.path.join(save_dir, "model_state.pdparams")
                paddle.save(model.state_dict(), save_param_path)
                tokenizer.save_pretrained(save_dir)


if __name__ == "__main__":
    do_train()