pmeletis / iv2018-hierarchical-semantic-segmentation-for-heterogeneous-datasets Goto Github PK

Hi,

I love your paper and am intrigued by the possibilities this approach brings. Now I run into a problem when simply wanting to try out your some of your code.

Firstly I tried to run predict.py.

I have installed all the packages as described in the README. But when loading the model checkpoint you provide here I get the following error:

INFO:tensorflow:Restoring parameters from checkpoint/model.ckpt-153000
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
2019-02-20 09:52:06.615188: E tensorflow/stream_executor/cuda/cuda_dnn.cc:378] Loaded runtime
CuDNN library: 7301 (compatibility version 7300) but source was compiled with 7102 (compatibility 
version 7100).  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.
2019-02-20 09:52:06.616217: F tensorflow/core/kernels/conv_ops.cc:717] Check failed: stream-
>parent()->GetConvolveAlgorithms( conv_parameters.ShouldIncludeWinogradNonfusedAlgo<T>(), 
&algorithms) 
Aborted (core dumped)

I have the following TF packages and cuda driver:

- tensorboard               1.6.0            py36hf484d3e_1  
- tensorflow-gpu            1.6.0                         0  
- tensorflow-gpu-base       1.6.0                      0
- cudatoolkit               9.0                  h13b8566_0  
- cudnn                     7.3.1                 cuda9.0_0 
- $LD_LIBRARY_PATH = /usr/local/cuda-9.0
- $CUDA_HOME = /usr/local/cuda-9.0
- $PATH =  /usr/local/cuda-9.0:$PATH

What am I missing? Any advice?

Secondly, I would like to try out train.py on the cityscapes or my own dataset. But its hard to grasp what is needed to be done with data before hand. How do we define the hierarchical structure? What format does the data need to be presented? A guide to get the most simple training working on a single dataset would be very helpful to get started!

Keep up the good work,

Marc

Hi,

Thanks for the repo, very interesting.

I uploaded the repo in Google Drive as I am working with Colab. I just want to play with the inference so I downloaded the checkpoint provided on Dropbox.

I created a folder called log to host the checkpoint files, the folder with the images is the same provided in the repo (samples). When running the code to do inference:

!python predict.py /content/drive/Colab Notebooks/hiercala/hiercala/semantic/log /content/drive/Colab Notebooks/hiercala/hiercala/semantic/samples --restore_emas --plotting --export_color_images --results_dir samples/results

I get this error:

usage: predict.py [-h] [--stride_system STRIDE_SYSTEM]
[--stride_network STRIDE_NETWORK]
[--stride_feature_extractor STRIDE_FEATURE_EXTRACTOR]
[--name_feature_extractor {resnet_v1_50,resnet_v1_101}]
[--height_system HEIGHT_SYSTEM]
[--width_system WIDTH_SYSTEM]
[--height_network HEIGHT_NETWORK]
[--width_network WIDTH_NETWORK]
[--height_feature_extractor HEIGHT_FEATURE_EXTRACTOR]
[--width_feature_extractor WIDTH_FEATURE_EXTRACTOR]
[--feature_dims_decreased FEATURE_DIMS_DECREASED]
[--fov_expansion_kernel_size FOV_EXPANSION_KERNEL_SIZE]
[--fov_expansion_kernel_rate FOV_EXPANSION_KERNEL_RATE]
[--upsampling_method {no,bilinear,hybrid}] [--enable_xla]
[--ckpt_path CKPT_PATH]
[--training_problem_def_path TRAINING_PROBLEM_DEF_PATH]
[--results_dir RESULTS_DIR]
[--inference_problem_def_path INFERENCE_PROBLEM_DEF_PATH]
[--plotting] [--timeout TIMEOUT] [--export_color_images]
[--export_lids_images] [--replace_void_decisions] [--Nb NB]
[--restore_emas]
log_dir predict_dir
predict.py: error: unrecognized arguments: /content/drive/Colab Notebooks/hiercala/hiercala/semantic/samples

It seems like the folder samples (with the images to predict) are recognised as a argument.
I tried changing the name of the folder but without succeed.

Many Thanks

Hi,
I managed to do inference correctly. The output generated is a .png image similar to groundtruth images.

Some thoughts below

• would it be possible to know how many instances has been detected (i.e. how many people, kerbs, signs,...)
• would it be possible to overlap masks produced and bounding boxes to the original image?, Basically I would like to get an output as attached.

Thanks,
Alberto.