train_gradient_cache_DPR.py

# Copyright (c) 2022 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 numpy as np
import paddle
from biencoder_base_model import BiEncoder, BiEncoderNllLoss
from NQdataset import DataUtil, NQdataSetForDPR
from paddle.optimizer.lr import LambdaDecay

from paddlenlp.transformers.bert.modeling import BertModel

parser = argparse.ArgumentParser()

parser.add_argument("--batch_size", required=True, type=int, default=None)
parser.add_argument("--learning_rate", required=True, type=float, default=None)
parser.add_argument("--save_dir", required=True, type=str, default=None)
parser.add_argument("--warmup_steps", required=True, type=int)
parser.add_argument("--epoches", required=True, type=int)
parser.add_argument("--max_grad_norm", required=True, type=int)
parser.add_argument("--train_data_path", required=True, type=str)
parser.add_argument("--chunk_size", required=True, type=int)
args = parser.parse_args()

chunk_nums = args.batch_size // args.chunk_size
data_path = args.train_data_path
batch_size = args.batch_size
learning_rate = args.learning_rate
epoches = args.epoches


def dataLoader_for_DPR(batch_size, source_data: list, epochs):
    index = np.arange(0, len(source_data))
    np.random.shuffle(index)
    batch_data = []
    for i in index:
        try:
            batch_data.append(source_data[i])

            if len(batch_data) == batch_size:
                yield batch_data
                batch_data = []

        except Exception:
            import traceback

            traceback.print_exc()
            continue


def get_model(model_name: str):
    question_model = BertModel.from_pretrained(model_name)
    context_model = BertModel.from_pretrained(model_name)
    model = BiEncoder(question_model, context_model)
    return model


model = get_model("bert-base-uncased")


def get_linear_scheduler(warmup_steps, training_steps):
    def lr_lambda(current_step):
        if current_step < warmup_steps:
            return float(current_step) / float(max(1, warmup_steps))
        return max(0.0, float(training_steps - current_step) / float(max(1, training_steps - warmup_steps)))

    return LambdaDecay(learning_rate=args.learning_rate, lr_lambda=lr_lambda, last_epoch=-1, verbose=False)


training_steps = 58880 * args.epoches / args.batch_size
scheduler = get_linear_scheduler(args.warmup_steps, training_steps)
optimizer = paddle.optimizer.AdamW(learning_rate=scheduler, parameters=model.parameters())


def get_dataset(data_path: str):
    data = NQdataSetForDPR(data_path)
    dataset = data.new_data
    return dataset


util = DataUtil()
LOSS = BiEncoderNllLoss()
batch_data = []
dataset = get_dataset(data_path)


def train():

    for epoch in range(epoches):

        index = np.arange(0, len(dataset))
        np.random.shuffle(index)

        batch_data = []

        for i in index:
            # dataLoader
            batch_data.append(dataset[i])
            if len(batch_data) == batch_size:
                all_questions = []
                all_contexts = []
                all_batch_input = util.create_biencoder_input(batch_data, inserted_title=True)

                all_positions = all_batch_input.is_positive

                all_inputs_questions_id = all_batch_input.questions_ids
                all_inputs_questions_segment = all_batch_input.question_segments

                all_inputs_contexts_id = all_batch_input.context_ids
                all_inputs_contexts_segment = all_batch_input.ctx_segments

                sub_q_ids = paddle.split(all_inputs_questions_id, chunk_nums, axis=0)
                sub_c_ids = paddle.split(all_inputs_contexts_id, chunk_nums, axis=0)
                sub_q_segments = paddle.split(all_inputs_questions_segment, chunk_nums, axis=0)
                sub_c_segments = paddle.split(all_inputs_contexts_segment, chunk_nums, axis=0)

                all_questions = []
                all_contexts = []
                all_CUDA_rnd_state_question = []
                all_CUDA_rnd_state_context = []

                for sub_q_id, sub_q_segment in zip(sub_q_ids, sub_q_segments):
                    with paddle.no_grad():
                        sub_CUDA_rnd_state = paddle.framework.random.get_cuda_rng_state()
                        all_CUDA_rnd_state_question.append(sub_CUDA_rnd_state)
                        sub_question_output = model.get_question_pooled_embedding(sub_q_id, sub_q_segment)
                        all_questions.append(sub_question_output)
                for sub_c_id, sub_c_segment in zip(sub_c_ids, sub_c_segments):
                    with paddle.no_grad():
                        sub_CUDA_rnd_state = paddle.framework.random.get_cuda_rng_state()
                        all_CUDA_rnd_state_context.append(sub_CUDA_rnd_state)
                        sub_context_output = model.get_context_pooled_embedding(sub_c_id, sub_c_segment)
                        all_contexts.append(sub_context_output)

                model_questions = paddle.concat(all_questions, axis=0)
                all_questions = []

                model_questions = model_questions.detach()

                model_questions.stop_gradient = False

                model_contexts = paddle.concat(all_contexts, axis=0)

                model_contexts = model_contexts.detach()

                model_contexts.stop_gradient = False

                all_contexts = []

                model_positions = all_positions

                loss, _ = LOSS.calc(model_questions, model_contexts, model_positions)

                print("loss is:")
                print(loss.item())

                loss.backward()

                grads_for_questions = paddle.split(model_questions.grad, chunk_nums, axis=0)
                grads_for_contexts = paddle.split(model_contexts.grad, chunk_nums, axis=0)

                for sub_q_id, sub_q_segment, CUDA_state, grad_for_each_question in zip(
                    sub_q_ids, sub_q_segments, all_CUDA_rnd_state_question, grads_for_questions
                ):

                    paddle.framework.random.set_cuda_rng_state(CUDA_state)

                    sub_question_output = model.get_question_pooled_embedding(sub_q_id, sub_q_segment)

                    finally_question_res_for_backward = paddle.dot(sub_question_output, grad_for_each_question)
                    finally_question_res_for_backward = finally_question_res_for_backward * (1 / 8.0)

                    finally_question_res_for_backward.backward(retain_graph=True)

                for sub_c_id, sub_c_segment, CUDA_state, grad_for_each_context in zip(
                    sub_c_ids, sub_c_segments, all_CUDA_rnd_state_context, grads_for_contexts
                ):
                    paddle.framework.random.set_cuda_rng_state(CUDA_state)

                    sub_context_output = model.get_context_pooled_embedding(sub_c_id, sub_q_segment)

                    finally_context_res_for_backward = paddle.dot(sub_question_output, grad_for_each_context)
                    finally_context_res_for_backward = finally_context_res_for_backward * (1 / 8.0)

                    finally_context_res_for_backward.backward(retain_graph=True)

                paddle.nn.ClipGradByGlobalNorm(clip_norm=args.max_grad_norm, group_name=model.parameters())
                optimizer.step()
                scheduler.step()
                optimizer.clear_grad()

                batch_data = []

        EPOCH = str(epoch)
        save_path_que = args.save_dir + "/question_model_" + EPOCH
        save_path_con = args.save_dir + "/context_model_" + EPOCH
        model.question_encoder.save_pretrained(save_path_que)
        model.context_encoder.save_pretrained(save_path_con)


if __name__ == "__main__":
    train()