AssertionError

[kimo20200]

Hello,

I got an error when I tried to train the model. I have attached a link in the box. Do you have any idea about it?
Thank you.

link: https://drive.google.com/file/d/14cI1BLo7FZQ8IhGBuGoUzsXJtEEtlp65/view?usp=sharing

Best wishes,
Qing

[LLYXC]

Thanks for the interest in OXnet. I can't find the screenshot anywhere. Would you please try again?

[kimo20200]

I have updated it, sorry for that. Please see the first box

[LLYXC]

In the current version of OXnet, we are sampling two types of data in one batch: labeled data and unlabeled data. Hence, you can see the usage of a two-stream batch sampler. The reported error is caused by

OXnet/retinanet/dataloader.py

Line 532 in 2c99bb5

assert len(self.secondary_indices) >= self.secondary_batch_size > 0

which means:

1. the size of the second batch of data (for unlabeled data) should be at least 1;
2. the total number of the unlabeled data should be larger than or equal to the second batch size (defined as follows:)
   

   OXnet/train_oxnet.py

   Line 123 in 2c99bb5

   labeled_batch, unlabeled_batch = 2, 2

[kimo20200]

Thanks for your reply, I have prepared my own train (labeled) dataset with coco-liked format, but where the unlabeled data I should include. I am confused about it.

[LLYXC]

You can create two COCO-style json files for labeled data and unlabeled data and merge these two in one single json file. Say you have N labeled data, and M unlabeled data, then the beginning N data in the merged file should be labeled ones and the followed M data are the unlabeled ones. Set N and N+M to be num_labeled_data and num_data in the following:

OXnet/train_oxnet.py

Lines 121 to 125 in 2c99bb5

	labeled_idxs = list(range(parser.num_labeled_data))
	unlabeled_idxs = list(range(parser.num_labeled_data, parser.num_data))
	labeled_batch, unlabeled_batch = 2, 2
	batch_sampler = TwoStreamBatchSampler(labeled_idxs, unlabeled_idxs, labeled_batch, unlabeled_batch)
	dataloader_train = DataLoader(dataset_train, num_workers=8, collate_fn=tuple_collater, batch_sampler=batch_sampler)

Note you can give empty or negative values to the annotations for the unlabeled data, these would not be included in the training.
Thanks for pointing this out. I will clarify the instruction.

[kimo20200]

Very appreciate your clear demonstration and could you show an example of the json file or structure? I think it is easy for users to understand it.

[LLYXC]

A COCO-style json file would be a dictionary with the structure like follows:
dict_keys(['categories', 'images', 'annotations'])
Say you now have a json file for the labeled data (labeled.json), and a json file the for unlabeled data (unlabeled.json), you can merge these into a single file for training as follows:

import json as js
js_l = js.load('labeled.json')
js_u = js.load('unlabeled.json')
merged = {'categories': js_l['categories'],
          'images' : js_l['images'] + js_u ['images'] ,
          'annotations': js_l['annotations'] + js_u ['annotations']}
with open('train.json', 'w') as f:
    js.dump(merged, f)

[kimo20200]

Thank you a lot, I will try it later

[kimo20200]

Hi, when I was training, the Cls Cons loss and Reg Cons loss were always 0. They were calculated by soft focal loss. Does it means I add a incorrect unlabeled dataset?

[LLYXC]

You may need to pre-train a RetinaNet and use the weights to initialize the teacher model so that the teacher model could generate meaningful pseudo labels to be used in the soft focal loss.
I think the mentioned problem might be that the teacher model does not generate pseudo labels.

[kimo20200]

Cool. Thank you.