Hi, thanks for your implementation.

Have you run the experiments in the original paper?

Are there any improvements by using GAN in semantic segmentation tasks?

GAN often generate weird results which seem like "ground truth", but poorly in the evaluation.

Hi, When I tried ./03-inference.py -w pretrained_model/vgg16.npz -i JPEGImages/2012_001258.jpg I got the error: File "./03-inference.py", line 113, in <module> main() File "./03-inference.py", line 95, in main chainer.serializers.load_npz(args.weight, model) File "/media/vip/Data1/cj_project/chainer_ws/local/lib/python2.7/site-packages/chainer/serializers/npz.py", line 153, in load_npz d.load(obj) File "/media/vip/Data1/cj_project/chainer_ws/local/lib/python2.7/site-packages/chainer/serializer.py", line 83, in load obj.serialize(self) File "/media/vip/Data1/cj_project/chainer_ws/local/lib/python2.7/site-packages/chainer/link.py", line 795, in serialize d[name].serialize(serializer[name]) File "/media/vip/Data1/cj_project/chainer_ws/local/lib/python2.7/site-packages/chainer/link.py", line 551, in serialize data = serializer(name, param.data) File "/media/vip/Data1/cj_project/chainer_ws/local/lib/python2.7/site-packages/chainer/serializers/npz.py", line 123, in __call__ numpy.copyto(value, dataset) ValueError: could not broadcast input array from shape (4096,25088) into shape (4096,512,7,7) Could you tell me what should I did? Thank you so much.

Run the training /home/jcc/Semantic-Segmentation-using-Adversarial-Networks/utils.py:193: RuntimeWarning: invalid value encountered in true_divide acc_cls = np.diag(hist) / hist.sum(axis=1) /home/jcc/Semantic-Segmentation-using-Adversarial-Networks/utils.py:195: RuntimeWarning: invalid value encountered in true_divide iu = np.diag(hist) / (hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist))

and the training just seems to stop here, and nothing indicating the training process is showing here! And if I ignore this by: import numpy as np np.seterr(divide='ignore', invalid='ignore'), the training stop here too. Is this cause by not finding the data? But my dataset is already here:

Hello Oyam,

You said you used FCN-VGG16. However, there is " n_class=1000 " in your models/vgg16.py code. It should use ImageNet with 1000 classes.

If I add "-s fcn32s" in command, I always got the shape error " ValueError: could not broadcast input array from shape (4096,25088) into shape (4096,512,7,7)"

Do I need to reshape the input data?

I only changed some parameters as follow, and change the dataset path. parser.add_argument('--iteration', '-i', type=int, default=5,... parser.add_argument('--evaluate_interval', type=int, default=1,... parser.add_argument('--snapshot_interval', type=int, default=5,... parser.add_argument('--display_interval', type=int, default=1,...

I run the 02-train.py without command line parameter and just use the default setting.

And if I set the iteration more than 5, the loss also start to become NaN after the 4th iteration.

Here is the log file { "dis/loss": 1.4394323825836182, "gen/loss": 549001.1875, "gen/accuracy": 0.8932809831192742, "gen/accuracy_cls": 0.5, "gen/iu": 0.4466404915596371, "gen/fwavacc": 0.797950914802537, "epoch": 0, "iteration": 1, "elapsed_time": 50.618666648864746 }, { "dis/loss": 1135330.0, "gen/loss": 487891.59375, "gen/accuracy": 0.8661764545563562, "gen/accuracy_cls": 0.5, "gen/iu": 0.4330882272781781, "gen/fwavacc": 0.7502616504278193, "epoch": 0, "iteration": 2, "elapsed_time": 96.50169348716736 }, { "dis/loss": 1.1617913440729532e+22, "gen/loss": 6.914105166024373e+21, "gen/accuracy": 0.0, "gen/accuracy_cls": 0.0, "gen/iu": 0.0, "gen/fwavacc": 0.0, "epoch": 0, "iteration": 3, "elapsed_time": 143.30005288124084 }, { "dis/loss": NaN, "gen/loss": NaN, "gen/accuracy": 0.46102918492009626, "gen/accuracy_cls": 0.3333333333333333, "gen/iu": 0.1536763949733654, "gen/fwavacc": 0.21254790934808832, "epoch": 0, "iteration": 4, "elapsed_time": 183.56889081001282 }

I use chainer 1.23.0; numpy 1.15.2 matplotlib 3.0.0 Has anyone run into the same problem? need help ;)

The code isn't compatible with any version of chainer Version>=2.0 Specifically, this is the error that is being thrown ValueError: invoke_before_training has been removed since Chainer v2.0.0. Use initializer= instead.

1.python download_dataset.py emerging questions ， File "download_dataset.py", line 7, in import utils ImportError: No module named utils ，I want know how to solve 2.I want to know How is this project used?

Thanks for sharing the implementation. My question is similar question #2 . Could you update/show the performance of the implementation with/without GAN? I would like to know if my running is correct or not. Thanks