Where is the faster RCNN code for training with own dataset? about chainercv HOT 9 CLOSED

Here is a train code.
You can train with you own dataset by changing the dataset used in the code from VOC to yours.

https://github.com/chainer/chainercv/blob/master/examples/faster_rcnn/train.py

from chainercv.

Thank you for the reply.
For the sake of other people who have similar questions (I have seen many such inquiries in other github codes, but so many with incomplete answers), could you clarify the points?

In train.py, lines 62-64:
train_data = VOCDetectionDataset(split='trainval', year='2007')
test_data = VOCDetectionDataset(split='test', year='2007',
use_difficult=True, return_difficult=True)

Are these the portion that needs to be changed? If yes, should I just specify the path to the train_data and test_data instead?

Thank you for helping!

from chainercv.

Yes, as you said, these are the lines that you need to change.
Your detection dataset should return three objects when indexed: img, bbox and label.
It is important that the shape and convention used in your dataset are consistent with ChainerCV.
Please check this.
https://github.com/chainer/chainercv#data-conventions

Also, the label_names is probably different with your dataset, so you need to change that too.
https://github.com/chainer/chainercv/blob/master/examples/faster_rcnn/train.py#L124
https://github.com/chainer/chainercv/blob/master/examples/faster_rcnn/train.py#L65

from chainercv.

Thank you for the info. Let me try these.

from chainercv.

Hello,
I tried faster RCNN with my own dataset (Ubuntu 14.04, anaconda=4.2.16 that created an env, chainer=2.0.1, chainercv=0.6.0, numpy=1.13.1, cupy=1.0.0.1).

I replaced one class of VOC2012 with my own class. The training went fine until iteration 3000 when there was an error message as shown below. Multiple trials yielded the same error, even after changing the plot_interval (in train.py) from 3000 to 10000. Is this a numpy error?

--------------------------------------error message below
2980 0 613.412 0.001 1.0425 0.34825 0.417773 0.0704124 0.206067
Process Process-12:...........................................] 4.27%
Traceback (most recent call last):............................] 25.91%
File "/home/owner/anaconda3/envs/chainercv/lib/python3.5/multiprocessing/process.py", line 249, in _bootstrap
self.run()rs/sec. Estimated time to finish: 3:51:27.651041.
File "/home/owner/anaconda3/envs/chainercv/lib/python3.5/multiprocessing/process.py", line 93, in run
self._target(*self._args, **self._kwargs)
File "/home/owner/anaconda3/envs/chainercv/lib/python3.5/site-packages/chainer/iterators/multiprocess_iterator.py", line 390, in _worker
data = _pack(dataset[index], mem)
File "/home/owner/anaconda3/envs/chainercv/lib/python3.5/site-packages/chainercv/datasets/transform_dataset.py", line 54, in getitem
in_data = self._dataset[index]
File "/home/owner/anaconda3/envs/chainercv/lib/python3.5/site-packages/chainer/dataset/dataset_mixin.py", line 67, in getitem
return self.get_example(index)
File "/home/owner/anaconda3/envs/chainercv/lib/python3.5/site-packages/chainercv/datasets/voc/voc_detection_dataset.py", line 138, in get_example
bbox = np.stack(bbox).astype(np.float32)
File "/home/owner/anaconda3/envs/chainercv/lib/python3.5/site-packages/numpy/core/shape_base.py", line 350, in stack
raise ValueError('need at least one array to stack')
ValueError: need at least one array to stack

from chainercv.

There may be an image without any ground truth bounding boxes. Please make sure that each training image has at least one annotation.

from chainercv.

Yeah, as @Hakuyume said, you have an image with no annotation.
You can make it work by filtering all images without annotations from the dataset.

from chainercv.

Thank you both for the comment! I will carefully check the files.

from chainercv.

Continued discussion in #391

from chainercv.