зеркало из https://github.com/microsoft/nni.git
[Hotfix] update mmdet to 3.0 (#5515)
This commit is contained in:
Super Daniel
GitHub
Родитель
b2e2a4d840
Коммит
eb96043c8a
|
|
@ -8,8 +8,9 @@ pytorch-lightning >= 1.6.1, < 2.0
|
|||
|
||||
# for full-test-compression
|
||||
-f https://download.openmmlab.com/mmcv/dist/cu117/torch1.13/index.html
|
||||
mmcv-full == 1.7.1
|
||||
mmdet
|
||||
mmcv >= 2.0.0rc4, < 2.1.0
|
||||
mmdet >= 3.0
|
||||
mmengine
|
||||
|
||||
git+https://github.com/microsoft/nn-Meter.git#egg=nn_meter
|
||||
lightgbm
|
||||
|
|
|
|||
|
|
@ -5,77 +5,74 @@ import pytest
|
|||
|
||||
import os
|
||||
import torch
|
||||
import mmcv
|
||||
import mmcv.cnn as mmcv_cnn
|
||||
import mmdet.core as mmdet_core
|
||||
import mmdet.models as mmdet_models
|
||||
from mmdet.apis import init_detector
|
||||
from mmcv.parallel import collate
|
||||
from mmdet.datasets import replace_ImageToTensor
|
||||
from mmdet.datasets.pipelines import Compose
|
||||
from mmdet.testing import demo_mm_inputs, get_detector_cfg
|
||||
from mmengine import Config
|
||||
from nni.common.concrete_trace_utils import concrete_trace, ConcreteTracer
|
||||
|
||||
config_files_correct = (
|
||||
'atss/atss_r50_fpn_1x_coco',
|
||||
'autoassign/autoassign_r50_fpn_8x2_1x_coco',
|
||||
'cascade_rcnn/cascade_mask_rcnn_r50_caffe_fpn_1x_coco',
|
||||
'centernet/centernet_resnet18_140e_coco',
|
||||
'centripetalnet/centripetalnet_hourglass104_mstest_16x6_210e_coco',
|
||||
'cityscapes/faster_rcnn_r50_fpn_1x_cityscapes',
|
||||
'cornernet/cornernet_hourglass104_mstest_8x6_210e_coco',
|
||||
'dcn/cascade_mask_rcnn_r50_fpn_dconv_c3-c5_1x_coco',
|
||||
'dcnv2/faster_rcnn_r50_fpn_mdconv_c3-c5_1x_coco',
|
||||
'ddod/ddod_r50_fpn_1x_coco',
|
||||
'deepfashion/mask_rcnn_r50_fpn_15e_deepfashion',
|
||||
'deformable_detr/deformable_detr_r50_16x2_50e_coco',
|
||||
'detr/detr_r50_8x2_150e_coco',
|
||||
'double_heads/dh_faster_rcnn_r50_fpn_1x_coco',
|
||||
'dyhead/atss_r50_caffe_fpn_dyhead_1x_coco',
|
||||
'dynamic_rcnn/dynamic_rcnn_r50_fpn_1x_coco',
|
||||
'empirical_attention/faster_rcnn_r50_fpn_attention_0010_1x_coco',
|
||||
'faster_rcnn/faster_rcnn_r50_fpn_1x_coco',
|
||||
'fcos/fcos_center_r50_caffe_fpn_gn-head_1x_coco',
|
||||
'foveabox/fovea_align_r50_fpn_gn-head_4x4_2x_coco',
|
||||
'fpg/faster_rcnn_r50_fpg-chn128_crop640_50e_coco',
|
||||
'free_anchor/retinanet_free_anchor_r50_fpn_1x_coco',
|
||||
'autoassign/autoassign_r50-caffe_fpn_1x_coco',
|
||||
# 'cascade_rcnn/cascade-mask-rcnn_r50_fpn_1x_coco',
|
||||
'centernet/centernet_r18-dcnv2_8xb16-crop512-140e_coco',
|
||||
'centripetalnet/centripetalnet_hourglass104_16xb6-crop511-210e-mstest_coco',
|
||||
# 'cityscapes/faster-rcnn_r50_fpn_1x_cityscapes',
|
||||
'cornernet/cornernet_hourglass104_8xb6-210e-mstest_coco',
|
||||
# 'dcn/cascade-mask-rcnn_r50-dconv-c3-c5_fpn_1x_coco',
|
||||
# 'dcnv2/faster-rcnn_r50_fpn_mdpool_1x_coco',
|
||||
# 'ddod/ddod_r50_fpn_1x_coco',
|
||||
# 'deepfashion/mask-rcnn_r50_fpn_15e_deepfashion',
|
||||
# 'deformable_detr/deformable-detr_r50_16xb2-50e_coco',
|
||||
'detr/detr_r18_8xb2-500e_coco',
|
||||
# 'double_heads/dh-faster-rcnn_r50_fpn_1x_coco',
|
||||
'dyhead/atss_r50-caffe_fpn_dyhead_1x_coco',
|
||||
# 'dynamic_rcnn/dynamic-rcnn_r50_fpn_1x_coco',
|
||||
# 'empirical_attention/faster-rcnn_r50-attn0010_fpn_1x_coco',
|
||||
# 'faster_rcnn/faster-rcnn_r50_fpn_1x_coco',
|
||||
'fcos/fcos_r18_fpn_gn-head-center-normbbox-centeronreg-giou_8xb8-amp-lsj-200e_coco',
|
||||
'foveabox/fovea_r50_fpn_4xb4-1x_coco',
|
||||
# 'fpg/faster-rcnn_r50_fpg_crop640-50e_coco',
|
||||
'free_anchor/freeanchor_r50_fpn_1x_coco',
|
||||
'fsaf/fsaf_r50_fpn_1x_coco',
|
||||
'gfl/gfl_r50_fpn_1x_coco',
|
||||
'ghm/retinanet_ghm_r50_fpn_1x_coco',
|
||||
'gn/mask_rcnn_r50_fpn_gn-all_2x_coco',
|
||||
'gn+ws/faster_rcnn_r50_fpn_gn_ws-all_1x_coco',
|
||||
'grid_rcnn/grid_rcnn_r50_fpn_gn-head_1x_coco',
|
||||
'groie/faster_rcnn_r50_fpn_groie_1x_coco',
|
||||
'hrnet/cascade_mask_rcnn_hrnetv2p_w18_20e_coco',
|
||||
'htc/htc_r50_fpn_1x_coco',
|
||||
'instaboost/cascade_mask_rcnn_r50_fpn_instaboost_4x_coco',
|
||||
'legacy_1.x/faster_rcnn_r50_fpn_1x_coco_v1',
|
||||
'lvis/mask_rcnn_r50_fpn_sample1e-3_mstrain_1x_lvis_v1',
|
||||
'ms_rcnn/ms_rcnn_r50_caffe_fpn_1x_coco',
|
||||
'nas_fcos/nas_fcos_fcoshead_r50_caffe_fpn_gn-head_4x4_1x_coco',
|
||||
'nas_fpn/retinanet_r50_fpn_crop640_50e_coco',
|
||||
'openimages/faster_rcnn_r50_fpn_32x2_1x_openimages',
|
||||
'ghm/retinanet_r50_fpn_ghm-1x_coco',
|
||||
# 'gn/mask-rcnn_r50_fpn_gn-all_2x_coco',
|
||||
# 'gn+ws/faster-rcnn_r50_fpn_gn_ws-all_1x_coco',
|
||||
# 'grid_rcnn/grid-rcnn_r50_fpn_gn-head_1x_coco',
|
||||
# 'groie/grid-rcnn_r50_fpn_gn-head-groie_1x_coco',
|
||||
# 'hrnet/cascade-mask-rcnn_hrnetv2p-w18_20e_coco',
|
||||
# 'htc/htc_r50_fpn_1x_coco',
|
||||
# 'instaboost/cascade-mask-rcnn_r50_fpn_instaboost-4x_coco',
|
||||
# 'lvis/mask-rcnn_r50_fpn_sample1e-3_mstrain-1x_lvis-v1',
|
||||
# 'ms_rcnn/ms-rcnn_r50_fpn_1x_coco',
|
||||
# 'nas_fcos/nas-fcos_r50-caffe_fpn_fcoshead-gn-head_4xb4-1x_coco',
|
||||
'nas_fpn/retinanet_r50_fpn_crop640-50e_coco',
|
||||
# 'openimages/faster-rcnn_r50_fpn_32xb2-1x_openimages',
|
||||
'paa/paa_r50_fpn_1x_coco',
|
||||
'pafpn/faster_rcnn_r50_pafpn_1x_coco',
|
||||
'pisa/pisa_faster_rcnn_r50_fpn_1x_coco',
|
||||
'point_rend/point_rend_r50_caffe_fpn_mstrain_1x_coco',
|
||||
# 'pafpn/faster-rcnn_r50_pafpn_1x_coco',
|
||||
# 'pisa/faster-rcnn_r50_fpn_pisa_1x_coco',
|
||||
# 'point_rend/point-rend_r50-caffe_fpn_ms-1x_coco',
|
||||
'pvt/retinanet_pvt-l_fpn_1x_coco',
|
||||
'queryinst/queryinst_r50_fpn_1x_coco',
|
||||
'regnet/cascade_mask_rcnn_regnetx-400MF_fpn_mstrain_3x_coco',
|
||||
'reppoints/bbox_r50_grid_center_fpn_gn-neck+head_1x_coco',
|
||||
'res2net/cascade_mask_rcnn_r2_101_fpn_20e_coco',
|
||||
'resnet_strikes_back/cascade_mask_rcnn_r50_fpn_rsb-pretrain_1x_coco',
|
||||
# 'queryinst/queryinst_r50_fpn_1x_coco',
|
||||
# 'regnet/cascade-mask-rcnn_regnetx-1.6GF_fpn_ms-3x_coco',
|
||||
'reppoints/reppoints-bbox_r50_fpn-gn_head-gn-grid_1x_coco',
|
||||
# 'res2net/cascade-mask-rcnn_res2net-101_fpn_20e_coco',
|
||||
# 'resnet_strikes_back/cascade-mask-rcnn_r50-rsb-pre_fpn_1x_coco',
|
||||
'retinanet/retinanet_r18_fpn_1x_coco',
|
||||
'rpn/rpn_r50_caffe_c4_1x_coco',
|
||||
'sabl/sabl_cascade_rcnn_r50_fpn_1x_coco',
|
||||
'scratch/faster_rcnn_r50_fpn_gn-all_scratch_6x_coco',
|
||||
'sparse_rcnn/sparse_rcnn_r50_fpn_1x_coco',
|
||||
'ssd/ssdlite_mobilenetv2_scratch_600e_coco',
|
||||
'swin/mask_rcnn_swin-s-p4-w7_fpn_fp16_ms-crop-3x_coco',
|
||||
'tood/tood_r50_fpn_1x_coco',
|
||||
'vfnet/vfnet_r2_101_fpn_mdconv_c3-c5_mstrain_2x_coco',
|
||||
'yolact/yolact_r50_1x8_coco',
|
||||
'yolo/yolov3_d53_320_273e_coco',
|
||||
'yolof/yolof_r50_c5_8x8_1x_coco',
|
||||
'yolox/yolox_nano_8x8_300e_coco',
|
||||
'rpn/rpn_r50_fpn_1x_coco',
|
||||
# 'sabl/sabl-cascade-rcnn_r50_fpn_1x_coco',
|
||||
# 'scratch/faster-rcnn_r50-scratch_fpn_gn-all_6x_coco',
|
||||
# 'sparse_rcnn/sparse-rcnn_r50_fpn_1x_coco',
|
||||
'ssd/ssdlite_mobilenetv2-scratch_8xb24-600e_coco',
|
||||
# 'swin/mask-rcnn_swin-s-p4-w7_fpn_amp-ms-crop-3x_coco',
|
||||
# 'tood/tood_r50_fpn_1x_coco',
|
||||
'vfnet/vfnet_r50_fpn_1x_coco',
|
||||
# 'yolact/yolact_r50_1xb8-55e_coco',
|
||||
'yolo/yolov3_d53_8xb8-320-273e_coco',
|
||||
'yolof/yolof_r50-c5_8xb8-1x_coco',
|
||||
'yolox/yolox_nano_8xb8-300e_coco',
|
||||
)
|
||||
|
||||
# has exceptions:
|
||||
|
|
@ -89,11 +86,11 @@ config_files_maskrcnn = (
|
|||
|
||||
# cannot run model: need gpu
|
||||
config_files_need_gpu = (
|
||||
'carafe/faster_rcnn_r50_fpn_carafe_1x_coco',
|
||||
'efficientnet/retinanet_effb3_fpn_crop896_8x4_1x_coco',
|
||||
'gcnet/cascade_mask_rcnn_x101_32x4d_fpn_syncbn-backbone_1x_coco',
|
||||
'resnest/cascade_mask_rcnn_s50_fpn_syncbn-backbone+head_mstrain_1x_coco',
|
||||
'selfsup_pretrain/mask_rcnn_r50_fpn_mocov2-pretrain_1x_coco',
|
||||
# 'carafe/faster-rcnn_r50_fpn-carafe_1x_coco',
|
||||
# 'efficientnet/retinanet_effb3_fpn_8xb4-crop896-1x_coco',
|
||||
# 'gcnet/cascade-mask-rcnn_x101-32x4d-syncbn_fpn_1x_coco',
|
||||
# 'resnest/cascade-mask-rcnn_s50_fpn_syncbn-backbone+head_ms-1x_coco',
|
||||
# 'selfsup_pretrain/mask-rcnn_r50-mocov2-pre_fpn_1x_coco',
|
||||
)
|
||||
|
||||
# cannot run model: need argument img_metas
|
||||
|
|
@ -173,11 +170,9 @@ def test_mmdetection(config_file: str):
|
|||
torch.cuda.empty_cache()
|
||||
|
||||
folder_prefix = os.environ['MMDET_DIR']
|
||||
img = '%s/tests/data/color.jpg' % folder_prefix
|
||||
|
||||
# Specify the path to model config and checkpoint file
|
||||
config = mmcv.Config.fromfile(
|
||||
folder_prefix + '/configs/' + config_file + '.py')
|
||||
config = Config.fromfile(folder_prefix + '/configs/' + config_file + '.py')
|
||||
|
||||
# RoIAlign will cause many errors. there are 4 ways to avoid it
|
||||
# 1. add 'mmcv_ops.RoIAlign' to leaf_module when tracing, and set 'config_dict['use_torchvision'] = True' recursively
|
||||
|
|
@ -204,29 +199,27 @@ def test_mmdetection(config_file: str):
|
|||
roi_align_setter(config._cfg_dict['model'])
|
||||
|
||||
leaf_module_append = ()
|
||||
if RoIAlign_solution in (1, 2):
|
||||
if RoIAlign_solution in (1, 2, 3):
|
||||
from mmcv import ops as mmcv_ops
|
||||
leaf_module_append = (mmcv_ops.RoIAlign,)
|
||||
|
||||
model = init_detector(config, device=device)
|
||||
|
||||
with torch.no_grad():
|
||||
cfg = model.cfg
|
||||
cfg.data.test.pipeline = replace_ImageToTensor(cfg.data.test.pipeline)
|
||||
test_pipeline = Compose(cfg.data.test.pipeline)
|
||||
img_data = test_pipeline(dict(img_info=dict(filename=img), img_prefix=None))['img']
|
||||
img_tensor = collate(img_data, 1).data[0].to(device)
|
||||
packed_inputs = demo_mm_inputs()
|
||||
|
||||
dummy_inputs = model.data_preprocessor(packed_inputs, False)
|
||||
|
||||
# init run
|
||||
# some models need to be run 2 times before doing trace
|
||||
model.forward_dummy(torch.rand_like(img_tensor))
|
||||
model.forward_dummy(torch.rand_like(img_tensor))
|
||||
model.forward(**dummy_inputs)
|
||||
model.forward(**dummy_inputs)
|
||||
|
||||
seed = torch.seed()
|
||||
torch.manual_seed(seed)
|
||||
out_orig_1 = model.forward_dummy(img_tensor)
|
||||
out_orig_1 = model.forward(**dummy_inputs)
|
||||
torch.manual_seed(seed)
|
||||
out_orig_2 = model.forward_dummy(img_tensor)
|
||||
out_orig_2 = model.forward(**dummy_inputs)
|
||||
assert check_equal(out_orig_1, out_orig_2), 'check_equal failure for original model'
|
||||
del out_orig_1, out_orig_2
|
||||
|
||||
|
|
@ -237,9 +230,9 @@ def test_mmdetection(config_file: str):
|
|||
orig_base_types = torch_fx.proxy.base_types
|
||||
torch_fx.proxy.base_types = (*torch_fx.proxy.base_types, intc, int64)
|
||||
|
||||
traced_model = concrete_trace(model, {'img': img_tensor},
|
||||
use_operator_patch=False,
|
||||
forward_function_name='forward_dummy',
|
||||
traced_model = concrete_trace(model, dummy_inputs,
|
||||
use_operator_patch=True,
|
||||
forward_function_name='forward',
|
||||
autowrap_leaf_function = {
|
||||
**ConcreteTracer.default_autowrap_leaf_function,
|
||||
all: ((), False, None),
|
||||
|
|
@ -260,7 +253,7 @@ def test_mmdetection(config_file: str):
|
|||
mmcv_cnn.bricks.wrappers.MaxPool2d,
|
||||
mmcv_cnn.bricks.wrappers.MaxPool3d,
|
||||
), fake_middle_class = (
|
||||
mmdet_core.anchor.anchor_generator.AnchorGenerator,
|
||||
mmdet_models.task_modules.prior_generators.AnchorGenerator,
|
||||
))
|
||||
|
||||
if config_file == 'pvt/retinanet_pvt-l_fpn_1x_coco':
|
||||
|
|
@ -268,21 +261,12 @@ def test_mmdetection(config_file: str):
|
|||
|
||||
seed = torch.seed()
|
||||
torch.manual_seed(seed)
|
||||
out_orig = model.forward_dummy(img_tensor)
|
||||
out_orig = model.forward(**dummy_inputs)
|
||||
torch.manual_seed(seed)
|
||||
out_orig_traced = traced_model(img_tensor)
|
||||
out_orig_traced = traced_model(**dummy_inputs)
|
||||
assert check_equal(out_orig, out_orig_traced), 'check_equal failure in original inputs'
|
||||
del out_orig, out_orig_traced
|
||||
|
||||
input_like = torch.rand_like(img_tensor)
|
||||
seed = torch.seed()
|
||||
torch.manual_seed(seed)
|
||||
out_like = model.forward_dummy(input_like)
|
||||
torch.manual_seed(seed)
|
||||
out_like_traced = traced_model(input_like)
|
||||
assert check_equal(out_like, out_like_traced), 'check_equal failure in new inputs'
|
||||
del input_like, out_like, out_like_traced
|
||||
|
||||
if __name__ == '__main__':
|
||||
for config_file in config_files_correct:
|
||||
test_mmdetection(config_file)
|
||||
|
|
|
|||
Загрузка…
Ссылка в новой задаче