test_bert_prim_cinn.py 3.9 KB
Newer Older
W
WangZhen 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
# Copyright (c) 2023 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 os
import time
import unittest

import numpy as np
from bert import Bert, BertPretrainingCriterion, create_pretraining_dataset

import paddle
23
from paddle import fluid
W
WangZhen 已提交
24
from paddle.dataset.common import DATA_HOME, download
25
from paddle.fluid import core
W
WangZhen 已提交
26 27 28 29 30 31 32 33 34

SEED = 2023
BATCH_SIZE = 2

URL = 'https://paddle-ci.gz.bcebos.com/prim_cinn/bert_training_data.npz'
MODULE_NAME = 'test_bert_prim_cinn'
MD5SUM = '71e730ee8d7aa77a215b7e898aa089af'
SAVE_NAME = 'bert_training_data.npz'

35 36

DY2ST_PRIM_CINN_GT = [
37 38 39 40 41 42 43 44 45 46
    10.976988792419434,
    10.345282554626465,
    10.33430004119873,
    10.275235176086426,
    10.22808837890625,
    10.200264930725098,
    10.161019325256348,
    10.10872745513916,
    10.121539115905762,
    9.990568161010742,
47 48
]

49

W
WangZhen 已提交
50 51 52 53 54 55 56 57 58
if core.is_compiled_with_cuda():
    paddle.set_flags({'FLAGS_cudnn_deterministic': True})


def train(to_static, enable_prim, enable_cinn):
    if core.is_compiled_with_cuda():
        paddle.set_device('gpu')
    else:
        paddle.set_device('cpu')
59
    fluid.core._set_prim_all_enabled(enable_prim)
W
WangZhen 已提交
60 61 62 63 64 65 66 67 68 69 70 71 72

    np.random.seed(SEED)
    paddle.seed(SEED)
    # paddle.framework.random._manual_program_seed(SEED)

    train_data_loader = create_pretraining_dataset(
        os.path.join(DATA_HOME, MODULE_NAME, SAVE_NAME),
        20,
        {},
        batch_size=BATCH_SIZE,
        worker_init=None,
    )

73 74
    # Now only apply dy2st for encoder
    bert = Bert(to_static, enable_cinn)
W
WangZhen 已提交
75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109
    criterion = BertPretrainingCriterion()

    optimizer = fluid.optimizer.Adam(parameter_list=bert.parameters())

    losses = []
    for step, batch in enumerate(train_data_loader):
        start_time = time.time()
        (
            input_ids,
            segment_ids,
            input_mask,
            masked_lm_positions,
            masked_lm_labels,
            next_sentence_labels,
            masked_lm_scale,
        ) = batch

        prediction_scores, seq_relationship_score = bert(
            input_ids=input_ids,
            token_type_ids=segment_ids,
            attention_mask=input_mask,
            masked_positions=masked_lm_positions,
        )

        loss = criterion(
            prediction_scores,
            seq_relationship_score,
            masked_lm_labels,
            next_sentence_labels,
            masked_lm_scale,
        )

        loss.backward()
        optimizer.minimize(loss)
        bert.clear_gradients()
110
        losses.append(loss.numpy().item())
W
WangZhen 已提交
111 112 113 114 115 116 117 118 119 120

        print(
            "step: {}, loss: {}, batch_cost: {:.5}".format(
                step,
                loss.numpy(),
                time.time() - start_time,
            )
        )
        if step >= 9:
            break
121
    print(losses)
W
WangZhen 已提交
122 123 124 125 126 127 128 129
    return losses


class TestBert(unittest.TestCase):
    @classmethod
    def setUpClass(cls):
        download(URL, MODULE_NAME, MD5SUM, SAVE_NAME)

130 131 132
    def tearDown(self):
        paddle.set_flags({'FLAGS_deny_cinn_ops': ''})

W
WangZhen 已提交
133
    @unittest.skipIf(
134 135
        not (paddle.is_compiled_with_cinn() and paddle.is_compiled_with_cuda()),
        "paddle is not compiled with CINN and CUDA",
W
WangZhen 已提交
136 137 138 139 140
    )
    def test_prim_cinn(self):
        dy2st_prim_cinn = train(
            to_static=True, enable_prim=True, enable_cinn=True
        )
141 142 143
        np.testing.assert_allclose(
            dy2st_prim_cinn, DY2ST_PRIM_CINN_GT, rtol=1e-5
        )
W
WangZhen 已提交
144 145 146 147


if __name__ == '__main__':
    unittest.main()