code_test/tools/train.py

# Copyright (c) OpenMMLab. All rights reserved.
import argparse
import copy
import os
import os.path as osp
import time
import warnings
import shutil
import sys
path = os.path.dirname(os.path.dirname(__file__))
print(path)
sys.path.append("/tmp/code/code_test")
os.system("pip install --no-cache-dir onnx==1.11.0 onnxruntime==1.11.1 protobuf==3.20.0")
#os.environ['RANK'] = "0"
#os.environ['WORLD_SIZE'] = "8"
#os.environ['MASTER_ADDR'] = "localhost"
#os.environ['MASTER_PORT'] = "1234"
import mmcv
import torch
#os.environ["CUDA_VISIBLE_DEVICES"] = "0, 1, 2, 3"
from mmcv import Config, DictAction
from mmcv.runner import get_dist_info, init_dist
from mmcv.utils import get_git_hash
from pycocotools.coco import COCO

from mmdet import __version__
from mmdet.apis import init_random_seed, set_random_seed, train_detector
from mmdet.datasets import build_dataset
from mmdet.models import build_detector
from mmdet.utils import collect_env, get_root_logger

# Copyright (c) OpenMMLab. All rights reserved.
from functools import partial
import numpy as np
from sklearn.covariance import LedoitWolf
from mmdet.core.export import build_model_from_cfg, preprocess_example_input
from mmdet.core.export.model_wrappers import ONNXRuntimeDetector
from mmdet.apis import (async_inference_detector, inference_detector,
                        init_detector, show_result_pyplot)
import onnxruntime as ort
import onnx
print(f"onnxruntime device: {ort.get_device()}") # output: GPU
print(f'ort avail providers: {ort.get_available_providers()}') # output: ['CUDAExecutionProvider', 'CPUExecutionProvider']

def parse_normalize_cfg(test_pipeline):
    transforms = None
    for pipeline in test_pipeline:
        if 'transforms' in pipeline:
            transforms = pipeline['transforms']
            break
    assert transforms is not None, 'Failed to find `transforms`'
    norm_config_li = [_ for _ in transforms if _['type'] == 'Normalize']
    assert len(norm_config_li) == 1, '`norm_config` should only have one'
    norm_config = norm_config_li[0]
    return norm_config

def pytorch2onnx(model,
                 input_img,
                 input_shape,
                 normalize_cfg,
                 opset_version=11,
                 show=False,
                 output_file='model.onnx',
                 verify=False,
                 test_img=None,
                 do_simplify=False,
                 dynamic_export=True,
                 skip_postprocess=False):

    input_config = {
        'input_shape': input_shape,
        'input_path': input_img,
        'normalize_cfg': normalize_cfg
    }
    # prepare input
    one_img, one_meta = preprocess_example_input(input_config)
    img_list, img_meta_list = [one_img], [[one_meta]]

    if skip_postprocess:
        warnings.warn('Not all models support export onnx without post '
                      'process, especially two stage detectors!')
        model.forward = model.forward_dummy
        torch.onnx.export(
            model,
            one_img,
            output_file,
            input_names=['input'],
            export_params=True,
            keep_initializers_as_inputs=True,
            do_constant_folding=True,
            verbose=show,
            opset_version=opset_version)

        print(f'Successfully exported ONNX model without '
              f'post process: {output_file}')
        return

    # replace original forward function
    origin_forward = model.forward
    model.forward = partial(
        model.forward,
        img_metas=img_meta_list,
        return_loss=False,
        rescale=False)

    output_names = ['dets', 'labels', 'feature', 'entropy', 'learning_loss']
    if model.with_mask:
        output_names.append('masks')
    input_name = 'input'
    dynamic_axes = None
    if dynamic_export:
        dynamic_axes = {
            input_name: {
                0: 'batch',
                2: 'height',
                3: 'width'
            },
            'dets': {
                0: 'batch',
                1: 'num_dets',
            },
            'labels': {
                0: 'batch',
                1: 'num_dets',
            },
            'feature': {
                0: 'batch',
                1: 'feat_dim',
            },
            'entropy': {
                0: 'batch',
                1: '1',
            },
            'learning_loss': {
                0: 'batch',
                1: '1',
            },
        }
        if model.with_mask:
            dynamic_axes['masks'] = {0: 'batch', 1: 'num_dets'}

    torch.onnx.export(
        model,
        img_list,
        output_file,
        input_names=[input_name],
        output_names=output_names,
        export_params=True,
        keep_initializers_as_inputs=True,
        do_constant_folding=True,
        verbose=show,
        opset_version=opset_version,
        dynamic_axes=dynamic_axes)

    model.forward = origin_forward

    # get the custom op path
    ort_custom_op_path = ''
    try:
        from mmcv.ops import get_onnxruntime_op_path
        ort_custom_op_path = get_onnxruntime_op_path()
    except (ImportError, ModuleNotFoundError):
        warnings.warn('If input model has custom op from mmcv, \
            you may have to build mmcv with ONNXRuntime from source.')

    if do_simplify:
        import onnxsim

        from mmdet import digit_version

        min_required_version = '0.3.0'
        assert digit_version(onnxsim.__version__) >= digit_version(
            min_required_version
        ), f'Requires to install onnx-simplify>={min_required_version}'

        input_dic = {'input': img_list[0].detach().cpu().numpy()}
        model_opt, check_ok = onnxsim.simplify(
            output_file,
            input_data=input_dic,
            custom_lib=ort_custom_op_path,
            dynamic_input_shape=dynamic_export)
        if check_ok:
            onnx.save(model_opt, output_file)
            print(f'Successfully simplified ONNX model: {output_file}')
        else:
            warnings.warn('Failed to simplify ONNX model.')
    print(f'Successfully exported ONNX model: {output_file}')

    if verify:
        # check by onnx
        onnx_model = onnx.load(output_file)
        onnx.checker.check_model(onnx_model)

        # wrap onnx model
        onnx_model = ONNXRuntimeDetector(output_file, model.CLASSES, 0)
        if dynamic_export:
            # scale up to test dynamic shape
            h, w = [int((_ * 1.5) // 32 * 32) for _ in input_shape[2:]]
            h, w = min(1344, h), min(1344, w)
            input_config['input_shape'] = (1, 3, h, w)

        if test_img is None:
            input_config['input_path'] = input_img

        # prepare input once again
        one_img, one_meta = preprocess_example_input(input_config)
        img_list, img_meta_list = [one_img], [[one_meta]]

        # get pytorch output
        with torch.no_grad():
            pytorch_results = model(
                img_list,
                img_metas=img_meta_list,
                return_loss=False,
                rescale=True)[0]

        img_list = [_.cuda().contiguous() for _ in img_list]
        if dynamic_export:
            img_list = img_list + [_.flip(-1).contiguous() for _ in img_list]
            img_meta_list = img_meta_list * 2
        # get onnx output
        onnx_results = onnx_model(
            img_list, img_metas=img_meta_list, return_loss=False)[0]
        # visualize predictions
        score_thr = 0.3
        if show:
            out_file_ort, out_file_pt = None, None
        else:
            out_file_ort, out_file_pt = 'show-ort.png', 'show-pt.png'

        show_img = one_meta['show_img']
        model.show_result(
            show_img,
            pytorch_results,
            score_thr=score_thr,
            show=True,
            win_name='PyTorch',
            out_file=out_file_pt)
        onnx_model.show_result(
            show_img,
            onnx_results,
            score_thr=score_thr,
            show=True,
            win_name='ONNXRuntime',
            out_file=out_file_ort)

        # compare a part of result
        '''print(input_config['input_shape'])
        print(one_img)
        print(len(onnx_results))
        print(len(pytorch_results))
        print(onnx_results)
        print(pytorch_results)'''
        for i in range(len(onnx_results)):
            print(onnx_results[i].shape)
        print("***************")
        for i in range(len(pytorch_results)):
            print(pytorch_results[i].shape)
        if model.with_mask:
            compare_pairs = list(zip(onnx_results, pytorch_results))
        else:
            compare_pairs = [(onnx_results, pytorch_results)]
        err_msg = 'The numerical values are different between Pytorch' + \
                  ' and ONNX, but it does not necessarily mean the' + \
                  ' exported ONNX model is problematic.'
        # check the numerical value
        for onnx_res, pytorch_res in compare_pairs:
            for o_res, p_res in zip(onnx_res, pytorch_res):
                np.testing.assert_allclose(
                    o_res, p_res, rtol=1e-03, atol=1e-05, err_msg=err_msg)
        print('The numerical values are the same between Pytorch and ONNX')
def parse_args():
    parser = argparse.ArgumentParser(description='Train a detector')
    parser.add_argument('--config', default='/tmp/code/code_test/configs/AD_mlops/AD_mlops_test18.py', help='train config file path')
    parser.add_argument('--work-dir',default='/tmp/output', help='the dir to save logs and models')
    parser.add_argument(
        '--resume-from', help='the checkpoint file to resume from')
    parser.add_argument(
        '--no-validate',
        action='store_true',
        help='whether not to evaluate the checkpoint during training')
    parser.add_argument(
        '--shape',
        help='infer image shape')
    group_gpus = parser.add_mutually_exclusive_group()
    group_gpus.add_argument(
        '--gpus',
        type=int,
        help='number of gpus to use '
        '(only applicable to non-distributed training)')
    group_gpus.add_argument(
        '--gpu-ids',
        type=int,
        nargs='+',
        help='ids of gpus to use '
        '(only applicable to non-distributed training)')
    parser.add_argument('--seed', type=int, default=None, help='random seed')
    parser.add_argument(
        '--deterministic',
        action='store_true',
        help='whether to set deterministic options for CUDNN backend.')
    parser.add_argument(
        '--options',
        nargs='+',
        action=DictAction,
        help='override some settings in the used config, the key-value pair '
        'in xxx=yyy format will be merged into config file (deprecate), '
        'change to --cfg-options instead.')
    parser.add_argument(
        '--cfg-options',
        nargs='+',
        action=DictAction,
        help='override some settings in the used config, the key-value pair '
        'in xxx=yyy format will be merged into config file. If the value to '
        'be overwritten is a list, it should be like key="[a,b]" or key=a,b '
        'It also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" '
        'Note that the quotation marks are necessary and that no white space '
        'is allowed.')
    parser.add_argument(
        '--launcher',
        choices=['none', 'pytorch', 'slurm', 'mpi'],
        default='none',
        help='job launcher')
    parser.add_argument('--local_rank', type=int, default=0)
    parser.add_argument(
        '--data-path', default='/tmp/dataset', help='dataset path')
    parser.add_argument(
        '--batchsize',
        type=int,
        default=8,
        help='training batch size')
    parser.add_argument(
        '--epoch',
        type=int,
        default=2,
        help='training epoch')
    parser.add_argument(
        '--warmup_iters',
        type=int,
        default=500,
        help='training warmup_iters')
    parser.add_argument(
        '--lr',
        type=float,
        default=0.001,
        help='learning rate')
    parser.add_argument('--train_image_size', 
        type=list, 
        default=[(100, 100)],
        help='train image size')
    parser.add_argument('--test_image_size', 
        type=list, 
        default=[(100, 100)], 
        help='test image size')
    args = parser.parse_args()
    if 'LOCAL_RANK' not in os.environ:
        os.environ['LOCAL_RANK'] = str(args.local_rank)

    if args.options and args.cfg_options:
        raise ValueError(
            '--options and --cfg-options cannot be both '
            'specified, --options is deprecated in favor of --cfg-options')
    if args.options:
        warnings.warn('--options is deprecated in favor of --cfg-options')
        args.cfg_options = args.options

    return args


def main():
    args = parse_args()

    cfg = Config.fromfile(args.config)
    if args.cfg_options is not None:
        cfg.merge_from_dict(args.cfg_options)
    # import modules from string list.
    if cfg.get('custom_imports', None):
        from mmcv.utils import import_modules_from_strings
        import_modules_from_strings(**cfg['custom_imports'])
    # set cudnn_benchmark
    if cfg.get('cudnn_benchmark', False):
        torch.backends.cudnn.benchmark = True

    if args.batchsize is not None:
        cfg.data.samples_per_gpu = args.batchsize
    if args.epoch is not None:
        cfg.runner.max_epochs = args.epoch
    if args.warmup_iters is not None:
        cfg.lr_config.warmup_iters = args.warmup_iters
    if args.lr is not None:
        cfg.optimizer.lr = args.lr
    '''if args.train_image_size is not None:
        cfg.train_pipeline[2].img_scale = args.train_image_size
        cfg.data.train.dataset.pipeline[2].img_scale = args.train_image_size'''
    if args.test_image_size is not None:
        cfg.test_pipeline[1].img_scale = args.test_image_size
        cfg.data.val.pipeline[1].img_scale = args.test_image_size
        cfg.data.test.pipeline[1].img_scale = args.test_image_size
    #if on platform, change the classnum fit the user define dataset
    if args.data_path is not None:
        coco_config=COCO(os.path.join(args.data_path,"annotations/instances_annotations.json"))
        cfg.data.train.img_prefix = os.path.join(args.data_path,"images")
        cfg.data.train.ann_file =  os.path.join(args.data_path,"annotations/instances_annotations.json")

        cfg.data.val.img_prefix = os.path.join(args.data_path,"images")
        cfg.data.val.ann_file =  os.path.join(args.data_path,"annotations/instances_annotations.json")

        cfg.data.test.img_prefix = os.path.join(args.data_path,"images")
        cfg.data.test.ann_file =  os.path.join(args.data_path,"annotations/instances_annotations.json")
        cfg.classes = ()
        for cat in coco_config.cats.values():
            cfg.classes = cfg.classes + tuple([cat['name']])

        cfg.data.train.classes = cfg.classes
        cfg.data.val.classes = cfg.classes
        cfg.data.test.classes = cfg.classes

        #some model will RepeatDataset to speed up training, make sure all dataset path replace to data_path
        #cfg = Config.fromstring(cfg.dump().replace("ann_file='data/coco/annotations/instances_train2017.json',","ann_file='{}',".format(os.path.join(args.data_path,"annotations/instances_annotations.json"))), ".py")
        #cfg = Config.fromstring(cfg.dump().replace("img_prefix='data/coco/train2017/',","img_prefix='{}',".format(os.path.join(args.data_path,"images"))), ".py")

        # replace the classes num fit userdefine dataset
        #cfg = Config.fromstring(cfg.dump().replace("num_classes=80","num_classes={0}".format(len(coco_config.getCatIds()))), ".py")
        cfg.model.bbox_head.num_classes = len(coco_config.getCatIds())
    print(cfg.dump())
    # work_dir is determined in this priority: CLI > segment in file > filename
    if args.work_dir is not None:
        # update configs according to CLI args if args.work_dir is not None
        cfg.work_dir = args.work_dir
    elif cfg.get('work_dir', None) is None:
        # use config filename as default work_dir if cfg.work_dir is None
        cfg.work_dir = osp.join('./work_dirs',
                                osp.splitext(osp.basename(args.config))[0])
    if args.resume_from is not None:
        cfg.resume_from = args.resume_from
    if args.gpu_ids is not None:
        cfg.gpu_ids = args.gpu_ids
    else:
        cfg.gpu_ids = range(1) if args.gpus is None else range(args.gpus)

    # init distributed env first, since logger depends on the dist info.
    if args.launcher == 'none':
        distributed = False
    else:
        distributed = True
        init_dist(args.launcher, **cfg.dist_params)
        # re-set gpu_ids with distributed training mode
        _, world_size = get_dist_info()
        cfg.gpu_ids = range(world_size)

    # create work_dir
    mmcv.mkdir_or_exist(osp.abspath(cfg.work_dir))
    # dump config
    cfg.dump(osp.join(cfg.work_dir, 'config.py'))
    # init the logger before other steps
    timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
    log_file = osp.join(cfg.work_dir, f'{timestamp}.log')
    logger = get_root_logger(log_file=log_file, log_level=cfg.log_level)

    # init the meta dict to record some important information such as
    # environment info and seed, which will be logged
    meta = dict()
    # log env info
    env_info_dict = collect_env()
    env_info = '\n'.join([(f'{k}: {v}') for k, v in env_info_dict.items()])
    dash_line = '-' * 60 + '\n'
    logger.info('Environment info:\n' + dash_line + env_info + '\n' +
                dash_line)
    meta['env_info'] = env_info
    meta['config'] = cfg.pretty_text
    # log some basic info
    logger.info(f'Distributed training: {distributed}')
    logger.info(f'Config:\n{cfg.pretty_text}')

    # set random seeds
    #seed = init_random_seed(args.seed)
    seed = 965702173
    logger.info(f'Set random seed to {seed}, '
                f'deterministic: {args.deterministic}')
    set_random_seed(seed, deterministic=args.deterministic)
    #set_random_seed(seed, deterministic=True)
    cfg.seed = seed
    meta['seed'] = seed
    meta['exp_name'] = osp.basename(args.config)

    model = build_detector(
        cfg.model,
        train_cfg=cfg.get('train_cfg'),
        test_cfg=cfg.get('test_cfg'))
    model.init_weights()

    datasets = [build_dataset(cfg.data.train)]
    if len(cfg.workflow) == 2:
        val_dataset = copy.deepcopy(cfg.data.val)
        val_dataset.pipeline = cfg.data.train.pipeline
        datasets.append(build_dataset(val_dataset))
    if cfg.checkpoint_config is not None:
        # save mmdet version, config file content and class names in
        # checkpoints as meta data
        cfg.checkpoint_config.meta = dict(
            mmdet_version=__version__ + get_git_hash()[:7],
            CLASSES=datasets[0].CLASSES)
    # add an attribute for visualization convenience
    model.CLASSES = datasets[0].CLASSES
    train_detector(
        model,
        datasets,
        cfg,
        distributed=distributed,
        validate=(not args.no_validate),
        timestamp=timestamp,
        meta=meta)

    if args.shape is None:
        img_scale = cfg.test_pipeline[1]['img_scale'][0]
        print(img_scale)
        input_shape = (1, 3, img_scale[1], img_scale[0])
    elif len(args.shape) == 1:
        input_shape = (1, 3, args.shape[0], args.shape[0])
    elif len(args.shape) == 2:
        input_shape = (1, 3) + tuple(args.shape)
    else:
        raise ValueError('invalid input shape')
    # create onnx dir
    onnx_path = osp.join(args.work_dir)
    model = build_model_from_cfg(osp.join(args.work_dir, 'config.py'), osp.join(args.work_dir, "latest.pth"))
    input_img = osp.join(osp.dirname(__file__), 'demo.jpg')
    normalize_cfg = parse_normalize_cfg(cfg.test_pipeline)
    
    # convert model to onnx file
    pytorch2onnx(
        model,
        input_img,
        input_shape,
        normalize_cfg,
        output_file=osp.join(onnx_path,'model.onnx'),
        test_img=input_img)
    
    #启智平台
    shutil.copytree(osp.abspath(osp.join(osp.dirname(__file__),'../transformer/')), osp.join(args.work_dir, "transformer"))
    #shutil.copy(osp.join(args.train_work_dir, "config.py"), osp.join(args.work_dir, "config.py"))
    class_name_file = open(osp.join(args.work_dir, "class_names.txt"), 'w')
    for name in cfg.classes:
        class_name_file.write(name+'\n')
    shutil.copy(osp.abspath(osp.join(osp.dirname(__file__),'serve_desc.yaml')), osp.join(args.work_dir, "serve_desc.yaml"))
    


if __name__ == '__main__':
    main()