First of all I just want to thank you for the great work. I am having an issue during training, my loss and val_loss is nan, however I am still getting values for accuracy and val_acc. I am training on the PASCAL_VOC 2012 dataset with the segmentation class pngs.

keras 1.2.1 & 2.0.6 tensorflow-gpu 1.2.1 python 3.6.1

Hello,

You have provided the pre-trained model of VOC. I have a small dataset with 2 classes, which I annotated based on VOC and I want to fine-tune it. Would you please guide me through the process?

Hi Nicolov,

Thanks for the great work! I tried to train new dataset by generating my own set of jpg and png masks. However I realized it only works for pre-defined 20 classes. For example I wanted to re-train this network to segment screws from background, I wasn't able to find way to add new classes but to use a existed color 0x181818 which was originally trained for cats. After training it did segmented the screw. However I'm still wondering is there any way to change the number of classes and specify which color value are associated with certain class?

Hi! I have val accuracy = 1, but when i am trying to predict mask on the image from train set it displays me black image. Does anybody know what is the reason of this behaviour?

Hi, @nicolov ,

For the caffe weight conversion, I got the following error:

(tf_1.0) root@milton-All-Series:/data/code/segmentation_keras/conversion# docker run -v $(pwd):/workspace -ti `docker build -q .`
Cannot connect to the Docker daemon at unix:///var/run/docker.sock. Is the docker daemon running?
"docker run" requires at least 1 argument(s).
See 'docker run --help'.

Usage:  docker run [OPTIONS] IMAGE [COMMAND] [ARG...]

Run a command in a new container
(tf_1.0) root@milton-All-Series:/data/code/segmentation_keras/conversion#

It shows that docker daemon is not running. Any other command should I input before it?

Thanks

Hi nicolov,

In the post, you explained how to do the weight conversion. Due to the development environment constraints, it is a little bit hard for me to follow exactly your step.

In keras blog, author also show a way to load VGG16 weight from Keras directly. Do you think this weight can be used for your implementation? Do we have to use the converted caffe model weight for pascal_voc. The data set I will be using is of different domain with the data set published in the paper. Thanks for your advice.

While running train.py, this is the error message: Epoch 1/20 E tensorflow/stream_executor/cuda/cuda_dnn.cc:378] Loaded runtime CuDNN library: 6021 (compatibility version 6000) but source was compiled with 5105 (compatibility version 5100). If using a binary install, upgrade your CuDNN library to match. If building from sources, make sure the library loaded at runtime matches a compatible version specified during compile configuration.

Since I don't have the root account, I can't install CuDNN v5. Do you know how I can fix this? Thanks!

Have you by any chance compared this to the original implementation with regards to the mean IoU? If so, what implementation of IoU did you use and what were your results?

Hi, @nicolov ,

According to this line, vgg_conv.npy is needed as pre-trained vgg model in training. Could you list the download location for corresponding caffemodel and prototxt file? And, is the conversion step the same as here?

Thanks!

Hi nicolov,

In the train.py, you have included the function of load_weights(model, weights_path):. My understanding is that you are trying to load a pre-training vcg model. If I do not want to use this pretrained model because the problem I am working one may belong to a totally different domain, should I just skip calling this load_weights function? Or using a pre-trained model is always preferable, I am kind of confusing about this.

In the notes, you mentioned that The training code is currently limited to the frontend module, and thus only outputs 16x16 segmentation maps. If I would like to leverage this code for my own data set, what are the modifications that I have to make? Do I still have to load the weights?

Thank you very much!

Probably a rookie error but when I am trying to run the conversion step in conversion by running the docker I get the following error:

$sudo docker run -v $(pwd):/workspace -ti `docker build -q .`
time="2017-02-09T09:15:11-08:00" level=fatal msg="Cannot locate Dockerfile: Dockerfile" 
docker: "run" requires a minimum of 1 argument. See 'docker run --help'.

On executing command: python train.py --batch-size 2 ,training freezes at last step of first epoch.

All the libraries are according to the requirement.txt file

Hi @nicolov I was wondering whether in your model, you wouldn't need to have a Conv2DTranspose or Upsample layer to compensate for the maxpool and obtain predictions with the same size as your input image?

Hello @nicolov ,

let me first express my appreciation to your work in image segmentation its great (Y)

small suggestion , i just want to notify you that there is a missing -- in input parsing . very minor change

parser.add_argument('--input_path', nargs='?', default='images/cat.jpg',
                        help='Required path to input image') 

I'm hoping you can help me in understanding how to handle high res images as 1028 and 4k ,

also in the code i found you set input_width, input_height = 900, 900 and label_margin = 186 can you please illustrate what is the reason for this static number and how they effect on the output high and width

output_height = input_height - 2 * label_margin
output_width = input_width - 2 * label_margin

Thanks for creating this implementation. Do you have any plans to implement training of the context module (to allow producing full resolution segmentation maps)?

https://github.com/nicolov/segmentation_keras/blob/master/convert_masks.py isn't necessary.

Just use Pillow and you can load the classes separately from the color palette, which means it will already be in the format you want!

from https://github.com/aurora95/Keras-FCN/blob/master/utils/SegDataGenerator.py#L203

                from PIL import Image
                label = Image.open(label_filepath)
                if self.save_to_dir and self.palette is None:
                    self.palette = label.palette

cool right?