mmdet 安装与使用
!pip install openmim
!mim install mmdet
上述指令即可完成对mmdet的安装，而在kaggle上大概要花20min左右，主要是造轮子的时间太久了。
这边有一个取巧的办法，可以直接copy&edit这个工程可以很快的完成mmdet的安装。
mmdet + tensorboard
在kaggle上我们无法直接访问tensorboard，这里可以用ngrok做转发，代码如下
import datetime
import tensorflow as tf
from tensorflow import summary
%load_ext tensorboard
log_dir="runs/"
# summary_writer = tf.summary.create_file_writer(
# log_dir + "fit/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
!wget https://bin.equinox.io/c/4VmDzA7iaHb/ngrok-stable-linux-amd64.zip
!unzip ngrok-stable-linux-amd64.zip
!./ngrok authtoken xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
import os
import multiprocessing
pool = multiprocessing.Pool(processes = 10)
results_of_processes = [pool.apply_async(os.system, args=(cmd, ), callback = None )
for cmd in [
f"tensorboard --logdir ./runs/ --host 0.0.0.0 --port 6006 &",
"./ngrok http 6006 &"
]]
! curl -s http://localhost:4040/api/tunnels | python3 -c \
"import sys, json; print(json.load(sys.stdin)['tunnels'][0]['public_url'])"
其中authtoken需要注册ngrok可以获得，ngrok的链接:ngrok
config文件与训练
%%writefile configs/myconfig/xxxxxxxxx.py
使用上述指令可以自行创建一个config文件，方便训练与修改。
!python tools/train.py configs/myconfig/xxxxxxxxx.py
验证
from mmcv import Config
cfg = Config.fromfile('configs/myconfig/yolox_l_4x4_80e_coco.py')
#--------------------------------------------------------------------------------#
def xyxy2xywh(bbox):
"""Convert ``xyxy`` style bounding boxes to ``xywh`` style for COCO
evaluation.
Args:
bbox (numpy.ndarray): The bounding boxes, shape (4, ), in
``xyxy`` order.
Returns:
list[float]: The converted bounding boxes, in ``xywh`` order.
"""
_bbox = bbox.tolist()
return [
_bbox[0],
_bbox[1],
_bbox[2] - _bbox[0],
_bbox[3] - _bbox[1],
]
#--------------------------------------------------------------------------------#
from mmdet.apis import inference_detector, init_detector, show_result_pyplot
import numpy as np
from pycocotools.coco import COCO
import json
from tqdm import tqdm
import mmcv
model = init_detector(cfg, '/kaggle/working/runs/epoch_80.pth')
# test_json_path = '/kaggle/working/mmdetection/data_anno/test2017.json'
test_json_path = '/kaggle/working/mmdetection/data_anno/instances_val2017.json'
coco = COCO(test_json_path)
imgIds = coco.getImgIds()
res = []
with tqdm(total=len(imgIds)) as pbar:
for i in imgIds:
img_info = coco.loadImgs(i)
img = mmcv.imread('/kaggle/input/fewshotlogodetection/val/images/' + img_info[0]['file_name'])
result = inference_detector(model, img)
bboxes = np.vstack(result)
labels = [
np.full(bbox.shape[0], j, dtype=np.int32)
for j, bbox in enumerate(result)
]
labels = np.concatenate(labels)
for k in range(len(labels)):
res_tmp = {}
box = xyxy2xywh(bboxes[k,:4])
res_tmp['image_id'] = int(i)
res_tmp['category_id'] = int(labels[k] + 1)
res_tmp['bbox'] = box
res_tmp['score'] = float(bboxes[k,4])
res.append(res_tmp)
pbar.update(1)
当然也可以直接使用现成的。
!python tools/test.py configs/myconfig/yolox_l_4x4_80e_coco.py /kaggle/input/modelpara/yolox_l_0.58.pth --format-only --options "jsonfile_prefix=/kaggle/working/res"
