classification.py 5.2 KB
Newer Older
D
dongshuilong 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 23 24
# 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.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import time
import platform
import paddle

from ppcls.utils.misc import AverageMeter
from ppcls.utils import logger


W
weishengyu 已提交
25
def classification_eval(engine, epoch_id=0):
D
dongshuilong 已提交
26 27 28 29 30 31 32
    output_info = dict()
    time_info = {
        "batch_cost": AverageMeter(
            "batch_cost", '.5f', postfix=" s,"),
        "reader_cost": AverageMeter(
            "reader_cost", ".5f", postfix=" s,"),
    }
W
weishengyu 已提交
33
    print_batch_step = engine.config["Global"]["print_batch_step"]
D
dongshuilong 已提交
34 35 36

    metric_key = None
    tic = time.time()
D
dongshuilong 已提交
37 38 39 40
    accum_samples = 0
    total_samples = len(
        engine.eval_dataloader.
        dataset) if not engine.use_dali else engine.eval_dataloader.size
W
weishengyu 已提交
41 42 43
    max_iter = len(engine.eval_dataloader) - 1 if platform.system(
    ) == "Windows" else len(engine.eval_dataloader)
    for iter_id, batch in enumerate(engine.eval_dataloader):
D
dongshuilong 已提交
44 45 46 47 48
        if iter_id >= max_iter:
            break
        if iter_id == 5:
            for key in time_info:
                time_info[key].reset()
W
weishengyu 已提交
49
        if engine.use_dali:
D
dongshuilong 已提交
50 51 52 53 54 55 56
            batch = [
                paddle.to_tensor(batch[0]['data']),
                paddle.to_tensor(batch[0]['label'])
            ]
        time_info["reader_cost"].update(time.time() - tic)
        batch_size = batch[0].shape[0]
        batch[0] = paddle.to_tensor(batch[0]).astype("float32")
C
cuicheng01 已提交
57
        if not engine.config["Global"].get("use_multilabel", False):
C
cuicheng01 已提交
58
            batch[1] = batch[1].reshape([-1, 1]).astype("int64")
D
dongshuilong 已提交
59
        # image input
W
weishengyu 已提交
60
        out = engine.model(batch[0])
D
dongshuilong 已提交
61
        # calc loss
W
weishengyu 已提交
62 63
        if engine.eval_loss_func is not None:
            loss_dict = engine.eval_loss_func(out, batch[1])
D
dongshuilong 已提交
64 65 66 67
            for key in loss_dict:
                if key not in output_info:
                    output_info[key] = AverageMeter(key, '7.5f')
                output_info[key].update(loss_dict[key].numpy()[0], batch_size)
D
dongshuilong 已提交
68 69 70 71 72

        # just for DistributedBatchSampler issue: repeat sampling
        current_samples = batch_size * paddle.distributed.get_world_size()
        accum_samples += current_samples

D
dongshuilong 已提交
73
        # calc metric
W
weishengyu 已提交
74
        if engine.eval_metric_func is not None:
D
dongshuilong 已提交
75
            if paddle.distributed.get_world_size() > 1:
D
dongshuilong 已提交
76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92
                pred_list = []
                label_list = []
                if isinstance(out, dict):
                    out = out["logits"]
                paddle.distributed.all_gather(pred_list, out)
                paddle.distributed.all_gather(label_list, batch[1])
                pred = paddle.concat(pred_list, 0)
                labels = paddle.concat(label_list, 0)
                if accum_samples > total_samples:
                    pred = pred[:total_samples + current_samples -
                                accum_samples]
                    labels = labels[:total_samples + current_samples -
                                    accum_samples]
                    current_samples = total_samples + current_samples - accum_samples
                metric_dict = engine.eval_metric_func(pred, labels)
            else:
                metric_dict = engine.eval_metric_func(out, batch[1])
D
dongshuilong 已提交
93 94 95 96 97 98 99
            for key in metric_dict:
                if metric_key is None:
                    metric_key = key
                if key not in output_info:
                    output_info[key] = AverageMeter(key, '7.5f')

                output_info[key].update(metric_dict[key].numpy()[0],
D
dongshuilong 已提交
100
                                        current_samples)
D
dongshuilong 已提交
101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118

        time_info["batch_cost"].update(time.time() - tic)

        if iter_id % print_batch_step == 0:
            time_msg = "s, ".join([
                "{}: {:.5f}".format(key, time_info[key].avg)
                for key in time_info
            ])

            ips_msg = "ips: {:.5f} images/sec".format(
                batch_size / time_info["batch_cost"].avg)

            metric_msg = ", ".join([
                "{}: {:.5f}".format(key, output_info[key].val)
                for key in output_info
            ])
            logger.info("[Eval][Epoch {}][Iter: {}/{}]{}, {}, {}".format(
                epoch_id, iter_id,
W
weishengyu 已提交
119
                len(engine.eval_dataloader), metric_msg, time_msg, ips_msg))
D
dongshuilong 已提交
120 121

        tic = time.time()
W
weishengyu 已提交
122 123
    if engine.use_dali:
        engine.eval_dataloader.reset()
D
dongshuilong 已提交
124 125 126 127 128 129
    metric_msg = ", ".join([
        "{}: {:.5f}".format(key, output_info[key].avg) for key in output_info
    ])
    logger.info("[Eval][Epoch {}][Avg]{}".format(epoch_id, metric_msg))

    # do not try to save best eval.model
W
weishengyu 已提交
130
    if engine.eval_metric_func is None:
D
dongshuilong 已提交
131 132 133
        return -1
    # return 1st metric in the dict
    return output_info[metric_key].avg