Anyone manages to run more than 256?

facebookresearch / pifuhd

High-Resolution 3D Human Digitization from A Single Image.

Other

9.5k stars 1.44k forks source link

Anyone manages to run more than 256? #87

Open FahriBilici opened 3 years ago

FahriBilici commented 3 years ago

I have rtx 2070S and i cant run more than 256x256. which type of gpu or specs i need for running more than 256x256 images?

Ceztio commented 3 years ago

I have a rtx2070 and i can run 512x512 no problem. What error do you get?

FahriBilici commented 3 years ago

I have a rtx2070 and i can run 512x512 no problem. What error do you get?

I got tcmalloc: large alloc. I was following colabs code in my pc

archana1998 commented 3 years ago

I was able to run 512*512 without an issue on a 2080S, and I reproduced the colab code locally as well. It worked fine, maybe show the exact code you ran and the entire error and traceback?

manncodes commented 3 years ago

To run locally for 512x512 on GTX 1650 takes 3 mins and 1024x1024 takes 10mins on CPU, as GPU ran out of memory. To do so just change a few things, as follows.

In recon.py:

add following line after imports:

os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"

after importing torch add line:
```
torch.cuda.empty_cache()
```
for switching to CPU change line 145 to:
```
cuda = torch.device('cpu')
```

lemme know if you need more help! :)

narin26 commented 3 years ago

I am getting not so good results myself on a Intel 4 gen core is 7 with 8 gb ram and nvidia 780. Have tried render after adding the lines as suggested. Result is not as good as compared to online using google colab. getting unable to allocate error if I increase the sample rate.