蒲黎明

> > Please refer to my previous [repo](https://github.com/Lotayou/SMPL) for more details. > > Basically, you need to download SMPL official model, and convert necessary components into a new pickle. You...

> [This link](https://github.com/akanazawa/hmr/blob/master/doc/train.md) could be helpful, make sure you use the `neutral_smpl_with_cocoplus_reg.pkl`, since it contains `cocoplus_regessor` Thanks a lot, and I have solved this problem by your advice.

I tried. I updated tools/valid.py to inference my images. `save_valid_image(image, final_results, result_image_path)` But the results looks bad. See results in [Issue17](https://github.com/HRNet/Higher-HRNet-Human-Pose-Estimation/issues/17)

> Hi, > I believe this is due to drawing all found keypoints, even with low score. Try to draw keypoints, in which network more or less confident, e.g. which...

我使用pandoc进行渲染

我写了一个pandoc的简易使用教程：[pandoc配置数学公式](https://plmsmile.github.io/2017/10/16/24-hexo-problems/#%E9%85%8D%E7%BD%AE%E6%95%B0%E5%AD%A6%E5%85%AC%E5%BC%8F)。可以作为一个参考。

> accelerate launch --config_file $ACCELERATE_CONFIG_FILE --num_processes $NUM_PROCESSES --num_machines $WORLD_SIZE --machine_rank $RANK --main_process_ip $MASTER_ADDR --main_process_port $MASTER_PORT src/train_bash.py --stage sft --do_train --model_name_or_path $MODEL_PATH --dataset alpaca_zh --template qwen --finetuning_type full --output_dir $OUTPUT_PATH --overwrite_cache...

Thanks for the sharing~ I also met the problem. I trained 30k person images and generated 2 black images. I found the loss is NAN. 

same cuda 12.4 [Hint: cudnn_dso_handle should not be null.] (at /paddle/paddle/phi/backends/dynload/cudnn.cc:64)

> > > > Thanks @Gokulnath-V. Your advice resolved my issues with paddle running on gpu. > > > > > > > > > I've fix the previous bugs...