request for README of pytorch-ssd-mmd-coral

[chrisliu54]

Hi, I run the examples/ssd/train_visda.py script after figure out the directory(using visda18 dataset), and I came across several errors as follows:

After I run next time(without any change), I got the following error message:

When debugging, I found tensor operation buggy in the following code:

visda-2018-public/detection/pytorch-ssd-mmd-coral/torchcv/transforms/random_flip.py

Lines 24 to 25 in 1ef7600

	xmin = w - boxes[:,2]
	xmax = w - boxes[:,0]

Specifically, when tensor pass throught the line above, the actual dimension is of 1d, not 2d.

I'm still puzzled about the dataset folder configuration. Hope the README file of the code coming out soon : )

[MInner]

As for the "Connection refused" error, install and run wisdom server in parallel, or comment lines that start it. Not sure about the latter, are you sure you downloaded our gt files and provided paths to them?

[chrisliu54]

Yep, I just downloaded the gt files mentioned in this page and modified the corresponding path in the source code. I used the images extracted from train.tar as training data and visda18-detection-train.txt/visda18-detection-test.txt as gt boxes.

The latter problem was actually caused by dimension mismatch. Sometimes a boxes with only 1 dimension was passed to the function. I fixed this problem by modifying the code this way:

xmin = w - boxes[..., 2]
xmax = w - boxes[..., 0]

[chrisliu54]

BTW, according to this page, I found the "Connection refued" error can be solved by setting num_worker=0 in when defining torch.utils.data.DataLoader.

[chrisliu54]

Well, another issue appeared when processing this line:

visda-2018-public/detection/pytorch-ssd-mmd-coral/torchcv/loss/ssd_loss.py

Line 63 in 5c5854c

cls_loss[cls_targets<0] = 0 # set ignored loss to 0

I got the error message below:

[heinzermch]

The whole code is written in a very sloppy way. I recommend to use TorchCV directly, which is what this project is based on anyway. Plus at least the implementation of Self-Ensemble is wrong (not that it would work for this task). I did not check the implementations of the other baselines, as I don't have a good understanding of them.

[MInner]

@heinzermch yep, I suspected that self-assembling code was probably wrong logically, but I would appreciate if you could point me more specifically towards what exactly is wrong with this implementation. As for the quality of code in the repo, it is pretty much torchcv SSD example with data readers for this specific dataset and somewhat different run logic since you have to train supervisely on source, unsupervisely on target, and eval on different splits of both. Pull requests are welcome :)

[krumo]

@chrisliu54 hi, did you solve this error?

[chrisliu54]

After I turned to use official repo of TorchCV, some weird errors has gone since then : ）