14:18 Dutch guy here. You were very close, actually the best non-dutch pronunciation of Van Goghs name I've heard so far. You nailed the pronunciation of the "hard G" sound at the end (goGH), but the first G in his name is pronounced the same way as the last. Great content, btw👍
Absolutely excellent intro to Neural Style Transfer. Got problems with cuda but hopefully I will resolve by building from scratch by removing anonconda
seriously awesome bro.. u r terirbly hardworking... one question... i tried to read the paper... but really got confused about those equations... how to interpret the implementation from papers.. can u guide me a bit... i find it really difficult
3:51 PackagesNotFoundError: The following packages are not available from current channels: - torchvission torchvission 0.5.0 is not available to install , can someone please help me out?
Okay I just figured it out, For people who faces same issue, I just did removed the torchvision dependency line from environment.yml file and created the environment and all , the in the IDE's command prompt i just "pip install torchvision==0.5.0" and it worked just fine .
For me, when using L-BGFS if I run out of gpu memory (also at 8gb) it buffers to normal ram (very slow, but will allow it to continue), at least that is my experience with tensorflow, maybe torch is different. I have spent the last 2 months rendering just 10,000 frames of style transfer video, I need a fast feedforward system in my life. Also there are neuralnet image scalers that are rather fast these days, it might be possible to implement rendering the video at lower resolution and then using a scaler to increase the output size and maintain smooth edges, much cleaner looking than a bicubic scaling. Keep the input to 512 and setting and output to 2048 and cropping to get fast 1080 output is my first thought, but its two steps per frame. My current setup uses tensorflow to style-transfer 960x540 and use a scaler to get 1920x1080, it is very slow. (also kicking myself for deleting coco set to free up space right now) I have your code open in my editor just waiting for another script to finish up so I can run it.
Sorry I took a short vacation finally so I'm slow with replies. There are ways to do it - Gatys suggested some methods also but I didn't have the time to implement it all.
Got the GPU working - I essentially built the environment manually using conda and removed the environment.yml file (which was loading a CPU only version of Torch).
Hey Aleksa, thanks for the awesome tutorials! You are a legend. I'm having a problem with my GPU. I am able to get PyTorch to connect to GPU within my Miniconda prompt, but for some reason I get the "not compiled with CUDA enabled" error when trying to run torch functions inside of PyCharm with this project loaded up...any chance you can help?
hello man, first thing first, I'd like to thank you, I've been looking kind of tutorials, and now I've some problem and i donno how to fixt it, I've been tried to follow you step by step, but I've stuck activate pytorch.nst, on my cmd is says "could not find conda environment: pytorch.nst" . could you helping me ?? thx man
hello and thank you very much for your share. I have errors like this, about loading images Y:\Programlar\Anaconda\envs\pytorch-nst\python.exe Y:/İndirilenler/pytorch-neural-style-transfer-master/neural_style_transfer.py Traceback (most recent call last): File "Y:/İndirilenler/pytorch-neural-style-transfer-master/neural_style_transfer.py", line 178, in results_path = neural_style_transfer(optimization_config) File "Y:/İndirilenler/pytorch-neural-style-transfer-master/neural_style_transfer.py", line 59, in neural_style_transfer content_img = utils.prepare_img(content_img_path, config['height'], device) File "Y:\İndirilenler\pytorch-neural-style-transfer-master\utils\utils.py", line 40, in prepare_img img = load_image(img_path, target_shape=target_shape) File "Y:\İndirilenler\pytorch-neural-style-transfer-master\utils\utils.py", line 23, in load_image img = cv.imread(img_path,0)[:, :, ::-1] # [:, :, ::-1] converts BGR (opencv format...) into RGB TypeError: 'NoneType' object is not subscriptable do you have any idea ?
You're welcome Gökhan! Try it out with placing your images in the data folder content and style images and using only names, check out the README file it's all in there!
14:18
Dutch guy here. You were very close, actually the best non-dutch pronunciation of Van Goghs name I've heard so far. You nailed the pronunciation of the "hard G" sound at the end (goGH), but the first G in his name is pronounced the same way as the last.
Great content, btw👍
Hahaha thanks man!
Thank you very much! I really learned a lot from this tutorial, and NST is so cool!
I learn a lot of thing in this video. you are a good teacher.
Absolutely excellent intro to Neural Style Transfer. Got problems with cuda but hopefully I will resolve by building from scratch by removing anonconda
Thats a pretty neat and on point tutorial. Thanks a lot man.
seriously awesome bro.. u r terirbly hardworking... one question... i tried to read the paper... but really got confused about those equations... how to interpret the implementation from papers.. can u guide me a bit... i find it really difficult
3:51
PackagesNotFoundError: The following packages are not available from current channels:
- torchvission
torchvission 0.5.0 is not available to install , can someone please help me out?
Okay I just figured it out, For people who faces same issue, I just did removed the torchvision dependency line from environment.yml file and created the environment and all , the in the IDE's command prompt i just "pip install torchvision==0.5.0" and it worked just fine .
@@mugunthan2357 thank you
thank you very much! can i train the model using AMD GPU?
Hi, how were you able to avoid retain_graph=True ? I have always this error and when I do it, I'm out of memory. Can you help me please ?
How does it handle the different resolutions of the images?
For me, when using L-BGFS if I run out of gpu memory (also at 8gb) it buffers to normal ram (very slow, but will allow it to continue), at least that is my experience with tensorflow, maybe torch is different. I have spent the last 2 months rendering just 10,000 frames of style transfer video, I need a fast feedforward system in my life. Also there are neuralnet image scalers that are rather fast these days, it might be possible to implement rendering the video at lower resolution and then using a scaler to increase the output size and maintain smooth edges, much cleaner looking than a bicubic scaling. Keep the input to 512 and setting and output to 2048 and cropping to get fast 1080 output is my first thought, but its two steps per frame. My current setup uses tensorflow to style-transfer 960x540 and use a scaler to get 1920x1080, it is very slow. (also kicking myself for deleting coco set to free up space right now) I have your code open in my editor just waiting for another script to finish up so I can run it.
Sorry I took a short vacation finally so I'm slow with replies. There are ways to do it - Gatys suggested some methods also but I didn't have the time to implement it all.
I do not know what I'm doing wrong, but whether running it through Conda or through PyCharm I can't seem to get the script to actually use my GPU
Also can't seem to get it to run off my GPU which sucks because I'm capping out my normal RAM and I have a GTX 3080
@@floataimlessly7407 Same here also have a 3080 ... did you get a solution
Got the GPU working - I essentially built the environment manually using conda and removed the environment.yml file (which was loading a CPU only version of Torch).
Hey Aleksa, thanks for the awesome tutorials! You are a legend.
I'm having a problem with my GPU. I am able to get PyTorch to connect to GPU within my Miniconda prompt, but for some reason I get the "not compiled with CUDA enabled" error when trying to run torch functions inside of PyCharm with this project loaded up...any chance you can help?
hello man, first thing first, I'd like to thank you, I've been looking kind of tutorials, and now I've some problem and i donno how to fixt it, I've been tried to follow you step by step, but I've stuck activate pytorch.nst, on my cmd is says "could not find conda environment: pytorch.nst" . could you helping me ?? thx man
I think you misspelled it it's pytorch-nst just look for the env name on top of the environment.yml file. You're welcome!
@@TheAIEpiphany nailed it
@@edwopicii1670 awesome! Glad it works!
@@TheAIEpiphany but you don't used visual studio code, are you ??
@@edwopicii1670 Noup - I use PyCharm.
hello and thank you very much for your share. I have errors like this, about loading images
Y:\Programlar\Anaconda\envs\pytorch-nst\python.exe Y:/İndirilenler/pytorch-neural-style-transfer-master/neural_style_transfer.py
Traceback (most recent call last):
File "Y:/İndirilenler/pytorch-neural-style-transfer-master/neural_style_transfer.py", line 178, in
results_path = neural_style_transfer(optimization_config)
File "Y:/İndirilenler/pytorch-neural-style-transfer-master/neural_style_transfer.py", line 59, in neural_style_transfer
content_img = utils.prepare_img(content_img_path, config['height'], device)
File "Y:\İndirilenler\pytorch-neural-style-transfer-master\utils\utils.py", line 40, in prepare_img
img = load_image(img_path, target_shape=target_shape)
File "Y:\İndirilenler\pytorch-neural-style-transfer-master\utils\utils.py", line 23, in load_image
img = cv.imread(img_path,0)[:, :, ::-1] # [:, :, ::-1] converts BGR (opencv format...) into RGB
TypeError: 'NoneType' object is not subscriptable
do you have any idea ?
You're welcome Gökhan! Try it out with placing your images in the data folder content and style images and using only names, check out the README file it's all in there!
samo cepaj aleksa
Hvala! Sportski pozdrav!