In [1]:
# Check nvcc version
!nvcc -V
# Check GCC version
!gcc --version
In [2]:
# install dependencies: (use cu101 because colab has CUDA 10.1)
!pip install -U torch==1.5.1+cu101 torchvision==0.6.1+cu101 -f https://download.pytorch.org/whl/torch_stable.html
# install mmcv-full thus we could use CUDA operators
!pip install mmcv-full
# Install mmdetection
!rm -rf mmdetection
!git clone https://github.com/open-mmlab/mmdetection.git
%cd mmdetection
!pip install -e .
# install Pillow 7.0.0 back in order to avoid bug in colab
!pip install Pillow==7.0.0
In [3]:
# Check Pytorch installation
import torch, torchvision
print(torch.__version__, torch.cuda.is_available())
# Check MMDetection installation
import mmdet
print(mmdet.__version__)
# Check mmcv installation
from mmcv.ops import get_compiling_cuda_version, get_compiler_version
print(get_compiling_cuda_version())
print(get_compiler_version())
In [4]:
!mkdir checkpoints
!wget -c https://download.openmmlab.com/mmdetection/v2.0/mask_rcnn/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco_bbox_mAP-0.408__segm_mAP-0.37_20200504_163245-42aa3d00.pth \
-O checkpoints/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco_bbox_mAP-0.408__segm_mAP-0.37_20200504_163245-42aa3d00.pth
In [5]:
from mmdet.apis import inference_detector, init_detector, show_result_pyplot
# Choose to use a config and initialize the detector
config = 'configs/mask_rcnn/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco.py'
# Setup a checkpoint file to load
checkpoint = 'checkpoints/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco_bbox_mAP-0.408__segm_mAP-0.37_20200504_163245-42aa3d00.pth'
# initialize the detector
model = init_detector(config, checkpoint, device='cuda:0')
In [6]:
# Use the detector to do inference
img = 'demo/demo.jpg'
result = inference_detector(model, img)
In [7]:
# Let's plot the result
show_result_pyplot(model, img, result, score_thr=0.3)
There are three ways to support a new dataset in MMDetection:
Usually we recommend to use the first two methods which are usually easier than the third.
In this tutorial, we gives an example that converting the data into the format of existing datasets like COCO, VOC, etc. Other methods and more advanced usages can be found in the doc.
Firstly, let's download a tiny dataset obtained from KITTI. We select the first 75 images and their annotations from the 3D object detection dataset (it is the same dataset as the 2D object detection dataset but has 3D annotations). We convert the original images from PNG to JPEG format with 80% quality to reduce the size of dataset.
In [8]:
# download, decompress the data
!wget https://download.openmmlab.com/mmdetection/data/kitti_tiny.zip
!unzip kitti_tiny.zip > /dev/null
In [9]:
# Check the directory structure of the tiny data
# Install tree first
!apt-get -q install tree
!tree kitti_tiny
In [10]:
# Let's take a look at the dataset image
import mmcv
import matplotlib.pyplot as plt
img = mmcv.imread('kitti_tiny/training/image_2/000073.jpeg')
plt.figure(figsize=(15, 10))
plt.imshow(mmcv.bgr2rgb(img))
plt.show()
After downloading the data, we need to implement a function to convert the kitti annotation format into the middle format. In this tutorial we choose to convert them in load_annotations function in a newly implemented KittiTinyDataset.
Let's take a look at the annotation txt file.
In [11]:
# Check the label of a single image
!cat kitti_tiny/training/label_2/000000.txt
According to the KITTI's documentation, the first column indicates the class of the object, and the 5th to 8th columns indicates the bboxes. We need to read annotations of each image and convert them into middle format MMDetection accept is as below:
[
{
'filename': 'a.jpg',
'width': 1280,
'height': 720,
'ann': {
'bboxes': <np.ndarray> (n, 4),
'labels': <np.ndarray> (n, ),
'bboxes_ignore': <np.ndarray> (k, 4), (optional field)
'labels_ignore': <np.ndarray> (k, 4) (optional field)
}
},
...
]
In [12]:
import copy
import os.path as osp
import mmcv
import numpy as np
from mmdet.datasets.builder import DATASETS
from mmdet.datasets.custom import CustomDataset
@DATASETS.register_module()
class KittiTinyDataset(CustomDataset):
CLASSES = ('Car', 'Pedestrian', 'Cyclist')
def load_annotations(self, ann_file):
cat2label = {k: i for i, k in enumerate(self.CLASSES)}
# load image list from file
image_list = mmcv.list_from_file(self.ann_file)
data_infos = []
# convert annotations to middle format
for image_id in image_list:
filename = f'{self.img_prefix}/{image_id}.jpeg'
image = mmcv.imread(filename)
height, width = image.shape[:2]
data_info = dict(filename=f'{image_id}.jpeg', width=width, height=height)
# load annotations
label_prefix = self.img_prefix.replace('image_2', 'label_2')
lines = mmcv.list_from_file(osp.join(label_prefix, f'{image_id}.txt'))
content = [line.strip().split(' ') for line in lines]
bbox_names = [x[0] for x in content]
bboxes = [[float(info) for info in x[4:8]] for x in content]
gt_bboxes = []
gt_labels = []
gt_bboxes_ignore = []
gt_labels_ignore = []
# filter 'DontCare'
for bbox_name, bbox in zip(bbox_names, bboxes):
if bbox_name in cat2label:
gt_labels.append(cat2label[bbox_name])
gt_bboxes.append(bbox)
else:
gt_labels_ignore.append(-1)
gt_bboxes_ignore.append(bbox)
data_anno = dict(
bboxes=np.array(gt_bboxes, dtype=np.float32).reshape(-1, 4),
labels=np.array(gt_labels, dtype=np.long),
bboxes_ignore=np.array(gt_bboxes_ignore,
dtype=np.float32).reshape(-1, 4),
labels_ignore=np.array(gt_labels_ignore, dtype=np.long))
data_info.update(ann=data_anno)
data_infos.append(data_info)
return data_infos
In [13]:
from mmcv import Config
cfg = Config.fromfile('./configs/faster_rcnn/faster_rcnn_r50_caffe_fpn_mstrain_1x_coco.py')
Given a config that trains a Faster R-CNN on COCO dataset, we need to modify some values to use it for training Faster R-CNN on KITTI dataset.
In [14]:
from mmdet.apis import set_random_seed
# Modify dataset type and path
cfg.dataset_type = 'KittiTinyDataset'
cfg.data_root = 'kitti_tiny/'
cfg.data.test.type = 'KittiTinyDataset'
cfg.data.test.data_root = 'kitti_tiny/'
cfg.data.test.ann_file = 'train.txt'
cfg.data.test.img_prefix = 'training/image_2'
cfg.data.train.type = 'KittiTinyDataset'
cfg.data.train.data_root = 'kitti_tiny/'
cfg.data.train.ann_file = 'train.txt'
cfg.data.train.img_prefix = 'training/image_2'
cfg.data.val.type = 'KittiTinyDataset'
cfg.data.val.data_root = 'kitti_tiny/'
cfg.data.val.ann_file = 'val.txt'
cfg.data.val.img_prefix = 'training/image_2'
# modify num classes of the model in box head
cfg.model.roi_head.bbox_head.num_classes = 3
# We can still use the pre-trained Mask RCNN model though we do not need to
# use the mask branch
cfg.load_from = 'checkpoints/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco_bbox_mAP-0.408__segm_mAP-0.37_20200504_163245-42aa3d00.pth'
# Set up working dir to save files and logs.
cfg.work_dir = './tutorial_exps'
# The original learning rate (LR) is set for 8-GPU training.
# We divide it by 8 since we only use one GPU.
cfg.optimizer.lr = 0.02 / 8
cfg.lr_config.warmup = None
cfg.log_config.interval = 10
# Change the evaluation metric since we use customized dataset.
cfg.evaluation.metric = 'mAP'
# We can set the evaluation interval to reduce the evaluation times
cfg.evaluation.interval = 12
# We can set the checkpoint saving interval to reduce the storage cost
cfg.checkpoint_config.interval = 12
# Set seed thus the results are more reproducible
cfg.seed = 0
set_random_seed(0, deterministic=False)
cfg.gpu_ids = range(1)
# We can initialize the logger for training and have a look
# at the final config used for training
print(f'Config:\n{cfg.pretty_text}')
In [15]:
from mmdet.datasets import build_dataset
from mmdet.models import build_detector
from mmdet.apis import train_detector
# Build dataset
datasets = [build_dataset(cfg.data.train)]
# Build the detector
model = build_detector(
cfg.model, train_cfg=cfg.get('train_cfg'), test_cfg=cfg.get('test_cfg'))
# Add an attribute for visualization convenience
model.CLASSES = datasets[0].CLASSES
# Create work_dir
mmcv.mkdir_or_exist(osp.abspath(cfg.work_dir))
train_detector(model, datasets, cfg, distributed=False, validate=True)
From the log, we can have a basic understanding the training process and know how well the detector is trained.
Firstly, the ResNet-50 backbone pre-trained on ImageNet is loaded, this is a common practice since training from scratch is more cost. The log shows that all the weights of the ResNet-50 backbone are loaded except the conv1.bias, which has been merged into conv.weights.
Second, since the dataset we are using is small, we loaded a Mask R-CNN model and finetune it for detection. Because the detector we actually using is Faster R-CNN, the weights in mask branch, e.g. roi_head.mask_head, are unexpected key in source state_dict and not loaded.
The original Mask R-CNN is trained on COCO dataset which contains 80 classes but KITTI Tiny dataset only have 3 classes. Therefore, the last FC layer of the pre-trained Mask R-CNN for classification has different weight shape and is not used.
Third, after training, the detector is evaluated by the default VOC-style evaluation. The results show that the detector achieves 54.1 mAP on the val dataset, not bad!
In [16]:
img = mmcv.imread('kitti_tiny/training/image_2/000068.jpeg')
model.cfg = cfg
result = inference_detector(model, img)
show_result_pyplot(model, img, result)
In [ ]: