Doesn't train actually.

[chichivica]

Many thanks for great opensource implementation of the semantic segmentation in pytorch ever!

I'm trying to proceed through training 'segnet' model on 'pascal' dataset. What I've done: 1) installed pytorch: 0.2.0_4 and python: 2.7.13 2) downloaded VOCtrainval_11-May-2012.tar from http://host.robots.ox.ac.uk/pascal/VOC/voc2012/#Development%20Kit 3) downloaded "Semantic Boundaries Dataset and Benchmark" from http://home.bharathh.info/pubs/codes/SBD/download.html 4) as stated in Readme, extracted and pointed to them in config.json file 5) started training process

• started visdom server

  python -m visdom.server

  started training as:

  python train.py --arch segnet --dataset pascal

  training successfully started and going looks well:

after completion, generated 100 segnet_pascal1%2d.pkl files

6) So, after that I'm trying to test newly trained model on the simple pictures:

python test.py --model_path segnet_pascal_1_99.pkl --dataset pascal --img_path 2007_000033.jpg --out_path result_33.jpg

But result is quite wrong:

for some reasons, output resolution differ and segmentation was not produced correctly.

Could you please give me some advises what I'm doing wrong?

Many thanks, Ivan

[chichivica]

In addition, trained FCN8S,

[meetps]

Refer #20 , Try training fcn-8s without polynomial learning rate. Should work fine.

[chichivica]

@meetshah1995 trained SEGNET with disabled polynomial learning rate (commented out two lines):

iter = len(trainloader)*epoch + i

poly_lr_scheduler(optimizer, args.l_rate, iter)

as well I had to change here like this:

rgb = np.zeros((temp.shape[1], temp.shape[2], 3))

It may be because I'm using python3, or because you previously made a conversion NHWC -> NCWH. I don't know. Have you ever tested Segnet architecture?

So, now I'm waiting util fcn-8s get trained to see results without polynomial learning rate.

[chichivica]

At least fcn-8s got trained. It took approximately 3 day on nvidia 1080.

[XavierCHEN34]

hi， why does your results looks much worse than the examples?

[meetps]

PolyLR doesn't seem to work with FCN. FCN-8s in default settings work fine, with mIoU > 60

[HelloAlone]

hi, @chichivica if you want to train SegNet, you should set batch_size and l_rate, because of the BN layer. For example , you can set batch_size=16 and l_rate=1e-4.

[Galto2000]

I am getting the same (bad) results as @chichivica, with SegNet, even with the settings @HelloAlone suggested. My GTX1080Ti just could not handle a batch size of 16, so I used a batch size of 14 instead.

Also, the noise on my training curves increases significantly over time.

[xdsonglinliu]

@chichivica @Galto2000 Thank you for your analysis. I meet the same problem. Have you finally trained the SegNet network successfully? or Is there any other better SegNet implementation based on pytorch?