hobbitlong / shape2prog Goto Github PK

For the Rendernet, I assumed each part of the program would go through the LSTM sequentially, but it looks each batch only goes through the LSTM once:

Plus, the hidden state is reset with each forward pass.
Maybe I'm missing something, but why use an LSTM in a nonrecurrent manner? I guess the gates could still work but I feel like it would be less effective that way.

Edit: Nevermind, for some reason my only experience with nn.LSTM has been passing the sequence in one part at a time in a for loop. But the documentation indicates you can pass the sequence in all at once.

Hi,

Are there some ways to visualize the generate shape in h5 format or the program after running the testing script?

Hi,

Very good work! I am now trying to run the test script as suggested, but it failed. Below are some steps to run, could you please help advise?

1. Install
   a. I use the docker with Ubuntu 16.04, CUDA 9.0, CUDNN 7.0
   b. I use conda to create the environment:
   conda create -n shapeprog python=3.5
   activate env
   pip install -r requirements.txt

2. Run the test script
   a. download data as suggested
   b. run the script: CUDA_VISIBLE_DEVICES=0 python test.py --model ./model/program_generator_GA_chair.t7 --data ./data/chair_testing.h5 --save_path ./output/chair/ --save_prog --save_img

3. Error message
   ========= arguments =========
   Test: epoch 0 batch 0/32, time=0.392
   Test: epoch 0 batch 10/32, time=0.219
   Test: epoch 0 batch 20/32, time=0.218
   Test: epoch 0 batch 30/32, time=0.219
   Mean IoU: 0.560
   Traceback (most recent call last):
   File "test.py", line 177, in
   run()
   File "test.py", line 160, in run
   out.write(program)
   UnicodeEncodeError: 'ascii' codec can't encode character '\xb0' in position 186: ordinal not in range

I have tried python 3.6, it gives the same error.

Hi, I'm trying to use your synthesized data, and I'm confused what does param d means here? I didn't find its appearance in function body.

In train_program_generator.py, during validation, if we set gen_shape=True, we actually render shapes on step level, rather than block level, I am wondering if this is a bug?

When I try to run download.sh, the connection times out with the wget commands. Trying to access the project page via web browser results in a timeout as well. Any idea if that can be fixed?

Looks like page and downladables are back up. Thanks!