the show result is error #4
Comments
|
|
|
It seems the model aims to detect the background rather than the object. How's your mAP result? Is it also strange? FYII've tested the model in COCO dataset, the result is good. #1 (comment) TipsMaybe you can load the official pth model to check the results. |
|
Something must happen in your code. I've tested in your demo image, the result is shown below: My vscode config code: {
"env": {"PYTHONPATH" : "${workspaceRoot}"},
"name": "solov2-coco:demo",
"type": "python",
"request": "launch",
"program": "${workspaceRoot}/tools_det/demo.py",
"console": "integratedTerminal",
"justMyCode": false,
"args": ["work_dirs/result.jpg", "work_dirs/tmp/class_1_ins_2/demo_solov2.png", "configs/det/solo/solov2_r50_fpn_coco.py", "/disk1/lihao/model_zoo/solo/SOLOv2_fpn_R50_1x.pth"]
}, |
|
configs\det_base_\datasets\coco-instance.py |
|
Yes, just delete it. I was testing low-data dataset in that moment. |
|
Hi,I meet some problem when train solov2. To make sure it's not the problem of coco data or too few train iterations ,after all it is 12 epochs not 36 epochs. |
|
Again, it would be useful if you can provide your training logs. Because the demo you train on official project(WXlong/SOLO) is also strange, it contains overlap parts. Yet the Matrix NMS should have avoided this situation. |
|
I think the trained model is no problem . The problem is no matrix nms in test phase when run tool-det/demo.py, so all the boxes without nms are drawn in image. why the result your shown above is only one box? can you show your test code ? |
|
A personal question, do you speak Chinese? We can talk in chinese. DEMO codeThe demo.py directly comes from the official mmdetection project. '''
Author: your name
Date: 2021-12-20 16:47:41
LastEditTime: 2022-04-21 15:48:09
LastEditors: Please set LastEditors
Description: 打开koroFileHeader查看配置 进行设置: https://github.com/OBKoro1/koro1FileHeader/wiki/%E9%85%8D%E7%BD%AE
FilePath: /research_workspace/tools_det/demo.py
'''
from argparse import ArgumentParser
import cv2
from mmdet.apis import (async_inference_detector, inference_detector,
init_detector, show_result_pyplot)
def parse_args():
parser = ArgumentParser()
parser.add_argument('img', help='Image file')
parser.add_argument('out', help='Image file')
parser.add_argument('config', help='Config file')
parser.add_argument('checkpoint', help='Checkpoint file')
parser.add_argument(
'--device', default='cuda:0', help='Device used for inference')
parser.add_argument(
'--score-thr', type=float, default=0.3, help='bbox score threshold')
args = parser.parse_args()
return args
def show_result_pyplot(model,
img,
result,
out_file, #加入out_file，运行时把改行注释删除
score_thr=0.5):
if hasattr(model, 'module'):
model = model.module
res_img = model.show_result(
img,
result,
score_thr=score_thr,
show=False
)
cv2.imwrite(out_file, res_img)
def main(args):
# build the model from a config file and a checkpoint file
model = init_detector(args.config, args.checkpoint, device=args.device)
# test a single image
result = inference_detector(model, args.img)
# show the results
show_result_pyplot(model, args.img, result, args.out, score_thr=args.score_thr)
if __name__ == '__main__':
args = parse_args()
main(args)Pretrain Model convert codeThis code can helps you to convert official solov2 pth to our mmdet format. # disclaimer: inspired by MoCo official repo and PyContrast repo
import argparse
import pickle as pkl
import sys
import torch
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Convert Models')
parser.add_argument('input', metavar='I', default="/disk1/lihao/model_zoo/InstanceLoc/insloc_c4_400ep.pkl", help='input model path')
parser.add_argument('output', metavar='O', default="/disk1/lihao/model_zoo/InstanceLoc/insloc_c4_400ep.pth", help='output path')
parser.add_argument('--c4', action='store_true', help='using ema model')
args = parser.parse_args()
print('=========================')
print(f'converting {args.input}')
print('=========================')
torch_weight = torch.load(args.input, encoding='ascii')
# torch_weight = torch.load(args.input)['state_dict']
new_state_dict = {}
for k, v in torch_weight.items():
# C4 backbone treats C5 layers as the shared head
if 'roi_head.shared_head' in k:
old_k = k
k = k.replace('roi_head.shared_head', 'backbone')
print(old_k, '---->', k)
if not args.c4 and 'roi_head.bbox_head.shared_fcs' in k:
old_k = k
k = k.replace(f'roi_head.bbox_head.shared_fcs.0.weight',
f'roi_heads.box_head.fc1.weight')
k = k.replace(f'roi_head.bbox_head.shared_fcs.0.bias',
f'roi_heads.box_head.fc1.bias')
print(old_k, '---->', k)
new_state_dict[k] = v.numpy()
continue
if 'backbone' in k and 'layer' not in k and 'backbone_k' not in k:
old_k = k
if args.c4:
# C4 stem
k = k.replace('backbone',
'backbone.stem').replace('bn1', 'conv1.norm')
else:
# FPN stem
k = k.replace('backbone', 'backbone.bottom_up.stem').replace(
'bn1', 'conv1.norm')
print(old_k, '---->', k)
elif 'backbone' in k and 'backbone_k' not in k and 'layer' in k:
if args.c4:
# C4 Backbone
old_k = k
if 'layer4' not in k:
k = k.replace("layer1", "res2")
k = k.replace("layer2", "res3")
k = k.replace("layer3", "res4")
else:
k = k.replace("backbone", "roi_heads")
k = k.replace("layer4", "res5")
k = k.replace("bn1", "conv1.norm")
k = k.replace("bn2", "conv2.norm")
k = k.replace("bn3", "conv3.norm")
k = k.replace("downsample.0", "shortcut")
k = k.replace("downsample.1", "shortcut.norm")
else:
# FPN backbone
old_k = k
k = k.replace('backbone', 'backbone.bottom_up')
k = k.replace("layer1", "res2")
k = k.replace("layer2", "res3")
k = k.replace("layer3", "res4")
k = k.replace("layer4", "res5")
k = k.replace("bn1", "conv1.norm")
k = k.replace("bn2", "conv2.norm")
k = k.replace("bn3", "conv3.norm")
k = k.replace("downsample.0", "shortcut")
k = k.replace("downsample.1", "shortcut.norm")
print(old_k, '--->', k)
elif 'neck' in k and 'neck_k' not in k:
# FPN neck
old_k = k
# replace lateral conv
k = k.replace('neck.lateral_convs.0.bn',
'backbone.fpn_lateral2.norm')
k = k.replace('neck.lateral_convs.1.bn',
'backbone.fpn_lateral3.norm')
k = k.replace('neck.lateral_convs.2.bn',
'backbone.fpn_lateral4.norm')
k = k.replace('neck.lateral_convs.3.bn',
'backbone.fpn_lateral5.norm')
k = k.replace('neck.lateral_convs.0.conv', 'backbone.fpn_lateral2')
k = k.replace('neck.lateral_convs.1.conv', 'backbone.fpn_lateral3')
k = k.replace('neck.lateral_convs.2.conv', 'backbone.fpn_lateral4')
k = k.replace('neck.lateral_convs.3.conv', 'backbone.fpn_lateral5')
# replace fpn conv
k = k.replace('neck.fpn_convs.0.bn', 'backbone.fpn_output2.norm')
k = k.replace('neck.fpn_convs.1.bn', 'backbone.fpn_output3.norm')
k = k.replace('neck.fpn_convs.2.bn', 'backbone.fpn_output4.norm')
k = k.replace('neck.fpn_convs.3.bn', 'backbone.fpn_output5.norm')
k = k.replace('neck.fpn_convs.0.conv', 'backbone.fpn_output2')
k = k.replace('neck.fpn_convs.1.conv', 'backbone.fpn_output3')
k = k.replace('neck.fpn_convs.2.conv', 'backbone.fpn_output4')
k = k.replace('neck.fpn_convs.3.conv', 'backbone.fpn_output5')
print(old_k, '--->', k)
elif 'roi_head.bbox_head.shared_convs' in k:
# 4conv in det head
old_k = k
conv_idx = int(k.split('.')[3])
k = k.replace(
f'roi_head.bbox_head.shared_convs.{conv_idx}.conv.weight',
f'roi_heads.box_head.conv{conv_idx+1}.weight')
k = k.replace(
f'roi_head.bbox_head.shared_convs.{conv_idx}.bn.weight',
f'roi_heads.box_head.conv{conv_idx+1}.norm.weight')
k = k.replace(
f'roi_head.bbox_head.shared_convs.{conv_idx}.bn.bias',
f'roi_heads.box_head.conv{conv_idx+1}.norm.bias')
k = k.replace(
f'roi_head.bbox_head.shared_convs.{conv_idx}.bn.running_mean',
f'roi_heads.box_head.conv{conv_idx+1}.norm.running_mean')
k = k.replace(
f'roi_head.bbox_head.shared_convs.{conv_idx}.bn.running_var',
f'roi_heads.box_head.conv{conv_idx+1}.norm.running_var')
k = k.replace(
f'roi_head.bbox_head.shared_convs.{conv_idx}.bn.num_batches_tracked',
f'roi_heads.box_head.conv{conv_idx+1}.norm.num_batches_tracked'
)
print(old_k, '--->', k)
else:
continue
new_state_dict[k] = v.numpy()
res = {
"model": new_state_dict,
"__author__": "Ceyuan",
"matching_heuristics": True
}
with open(args.output, "wb") as f:
pkl.dump(res, f) |
|
Hello~，按照你提供的方法在mmdet2.19.0上训练solov2模型，结果一直异常满屏蓝色轮廓。之前以为是训练的模型本身是正常的，只是测试时没有做matrix-nms，因为人像轮廓感觉已经出来了（如上面图所示）。现在我把pth模型转成onnx用c++前传，测试结果仍然是异常，现象和之前一模一样，这应该是模型本身的问题了。 |
|
您好，我在mmdet2.19版本上面加入您的代码，用coco-train数据集上面重新训练模型。作为对照，我先用mmdet上开源的mask-rcnn训练模型和测试模型，确定整体代码测试脚本和训练数据都没有异常。然后我再加入你的det和custom目录到mmdet2.19中训练solov2模型，用相同的脚本在coco-val上测试，结果还是异常。 另外，我在使用您网站上下载的模型进行测试时，会报错unexpected key in source state_dict: mask_feat_head，missing keys in source state_dict: mask_head，对模型转换之后测试不报错，但是测试结果图片仍然有上面异常点1和异常点4的问题 感觉还是有不对劲的地方，无法复现你展示的结果。 |
Hi,Thansk your work.
I am using mmdet-2.23.0 to train solov2 and succuss to get pth model
However,some problem occur when i test model result using demo.py
when i run " python demo.py ./imgs/30208.jpg ./res.jpg ./configs/solo/solo_r101_fpn_3x_coco.py ./SOLO_R101_3x.pth"
the result image show is strange, and other models show same result,why it happen?
The text was updated successfully, but these errors were encountered: