hubert0527 / infinitygan Goto Github PK

1. Is this a mistake in L365-L366?
2. NAN in my training.
   logs:
   WARNING:root:NaN or Inf found in input tensor.
   [*] EXP: InfinityGAN; d: nan; g: 2.42; r1: 0.01; ; mean path: 0.01:
   and i found diversity_z_loss = nan

Dear Authors:

Thank you for your great work.

I'm trying to train InfinityGAN with my own dataset which consistes of ~80,000 256 x256 images. I am using a batch size of 32 training on 4 x 3090 GPUs, and I have trained for about 150,000 iterations.

However, the FIDs are still high and I am not getting any visible results. I have attached some sampled results and training stats below, may I ask you if these look normal to you? Does these indicate I might have made some mistakes somewhere? Thank you so much for you help!

Hello,

I am training to train the small dataset from the paper, but I am getting a mistake:

[*] Found lmdb root on local hard drive: /home/anya/Programs/infinityGAN/data/lmdb
Traceback (most recent call last):
  File "train.py", line 798, in <module>
    train_set = MultiResolutionDataset(
  File "/home/anya/Programs/infinityGAN/dataset.py", line 262, in __init__
    self.env = lmdb.open(
lmdb.Error: /home/anya/Programs/infinityGAN/data/lmdb/flickr-landscape-small/train: No such file or directory

I have downloaded the data from Google drive and saved them in /home/anya/Programs/infinityGAN/data/lmdb/flickr-landscape-small/train directory, as lock.mdb and data-001.mdb. Or is the configuration wrong?
Thank you for your answer!

I cannot download the datasets, could you please help me check whether there is a resource problem?

I'm looking for a high-resolution image synthesizer (4k) and wonder if infinityGan suits this task.
I want to train it on my own data and potentially use it as an super-resolution model by finding the latent code of low-resolution images.

How come ultra-high resolution image synthesizer is not listed as an application of infinityGan? especially when I failed to find methods that synthesize images in this high resolution.

Hey, thanks for the open source code. Do you have benchmarks for various training setups, for instance, what is the iter/s for InfinityGAN.yaml (using patchsize of 101).

When i try to do outpainting. CUDA_VISIBLE_DEVICES="0" python test.py --model-config="./configs/model/InfinityGAN.yaml" --test-config="./configs/test/inversion_256x256_L2R.yaml" CUDA_VISIBLE_DEVICES="0" python test.py --model-config="./configs/model/InfinityGAN.yaml" --test-config="./configs/test/outpaint_with_fused_gen_256x256.yaml" --inv-records="./logs/InfinityGAN/test/outpaint_with_fused_gen_256x256/stats/.pkl" --inv-placements=0.5,0.25 When I submit the second instruction, there is no pkl in the folder ./logs/InfinityGAN/test/outpaint_with_fused_gen_256x256/stats/. I don't know where is the pkl file. I find there is a pkl file ("./logs/Infi nityGAN/test/inversion_256x256_L2R/stats/000000.pkl") in another folder, Is this the expected file? I copy the file from /logs/Infi nityGAN/test/inversion_256x256_L2R/stats/ to "./logs/InfinityGAN/test/outpaint_with_fused_gen_256x256/stats/. But finally, the bug is as follows. File "/home/hanzhifan/infinityGAN/test_managers/testing_vars_wrapper.py", line 271, in replace_by_records assert inv_img_st_loc_x >= 0 and inv_img_st_loc_y >= 0, AssertionError: Top-left corner of intended image exceeds image boundary. Got (-25, -1). Could u help me solve this problem? where is the correct pkl file? if the pkl file is correct? how to solve the second question?

Hi. I want to test outpainting with an .png image. I understood you used lmdb data. I want to use prepare_data.py to convert ans png image to lmdb format but I got error. How I can do that?
You mentioned that, we must first # Run inversion first and # Then outpaint
@hubert0527
parser.add_argument("--model-config", default="\infinityGAN-main\configs\model\InfinityGAN-UR.yaml", type=str)

parser.add_argument("--test-config", default="\infinityGAN-main\configs\test\inversion_256x256_L2R.yaml", type=str)

Howdy, when I enter the following command I am returned an error, listed below:

python prepare_data.py ./configs/dataset/flickr-landscape-small.yaml --train_only

Traceback (most recent call last):
File "prepare_data.py", line 200, in
cur_lmdb_root = config.data_params.lmdb_root
NameError: name 'config' is not defined

I'm not sure what the issue is here. Any sugestions?

1. Why do you use self.ts_spatial_size instead of self.ss_spatial_size in the denominator? Is it a mistake or a special design?
2. If it is a special design, could you please explain it? Thanks.

What is the timetable for the code being released?

When i try to do outpainting.

1. CUDA_VISIBLE_DEVICES="0" python test.py --model-config="./configs/model/InfinityGAN.yaml" --test-config="./configs/test/inversion_256x256_L2R.yaml"
2. CUDA_VISIBLE_DEVICES="0" python test.py --model-config="./configs/model/InfinityGAN.yaml" --test-config="./configs/test/outpaint_with_fused_gen_256x256.yaml" --inv-records="./logs/InfinityGAN/test/outpaint_with_fused_gen_256x256/stats/.pkl" --inv-placements=0.5,0.25
   When I submit the second instruction, there is no pkl in the folder ./logs/InfinityGAN/test/outpaint_with_fused_gen_256x256/stats/.
   I don't know where is the pkl file. I find that there is a pkl file ("./logs/Infi nityGAN/test/inversion_256x256_L2R/stats/000000.pkl") in another folder, Is this the expected file?
   I copy the file from /logs/Infi nityGAN/test/inversion_256x256_L2R/stats/ to "./logs/InfinityGAN/test/outpaint_with_fused_gen_256x256/stats/.
   But finally, the bug is as follows. File"/home/hanzhifan/infinityGAN/test_managers/testing_vars_wrapper.py", line 271, in replace_by_records assert inv_img_st_loc_x >= 0 and inv_img_st_loc_y >= 0,
   AssertionError: Top-left corner of intended image exceeds image boundary. Got (-25, -1).

Could u help me solve this problem?
1.where is the correct pkl file?
2.if the pkl file is correct? how to solve the second question?

Oh, here it is.

Dear Authors,

Thanks a lot for your great work! I was trying to reproduce your work but I struggled a lot to download the Flickr-lands-large dataset from Google Drive, which is more than 700GB. Sometimes the downloading was cut off due to network issues, while re-downloading is also not possible with Google Drive complains "too many users have downloaded this file". Basically I need to wait for 24 hours and repeat the process again. (so far failed 3-4 times)

I wonder if it's possible to upload the dataset to more user-friendly platforms, e.g. Dropbox? Other platforms are also fine as long as a successful download is achievable.

Many thanks in advance for your help!

When using the cyclic coordinate, some of the generated images are flipped horizontally. Is there a way to preserve the orientation specifically to the same as the training dataset?

When doing the fused generation with RGB image, I get a continuous image in one channel, but discontinuous in another channel. Is there a control on what channels are used in the training/testing?