lnd212
/
MinDalle_StableDiff
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Support for 6gb cards

This commit is contained in:
Морозов Андрей 2022-09-24 12:03:40 +04:00
parent ed4d5f6c1b
commit dafa2c16bf
2 changed files with 95 additions and 93 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
README.md
View File

@ -24,8 +24,9 @@ c. Create a write token/Copy an existing Token key and enter in the cmd window.
5. You can find output files in .\\data\\out\\%Timestamp% folder. Jpeg contain commentaries in metadata with some settings used while generating. 5. You can find output files in .\\data\\out\\%Timestamp% folder. Jpeg contain commentaries in metadata with some settings used while generating.
## Requirements ## Requirements
* 11GB of free space. * 10GB of free space.
* Nvidia card with 8GB+ video memory. * Nvidia card with 6GB+ video memory.
* Currently min dalle can run in 6gb. Testing purpose
## Additional info ## Tests
Tested on RTX3070. One picture was making 12 - 14 seconds. RTX3070: 12 - 14 seconds per picture.
RTX2060_Laptop: 21 - 25 seconds per picture.

177
scripts/stable_diff.py
View File

@ -20,113 +20,114 @@ from diffusers import (
from diffusers import StableDiffusionImg2ImgPipeline from diffusers import StableDiffusionImg2ImgPipeline
t = torch.cuda.get_device_properties(0).total_memory t = torch.cuda.get_device_properties(0).total_memory
if t <=8500000000: if t<=6400000000:
print("Not enough GPU memory to generate pictures") print("Running with less than 6gb memory. Working is not garanted")
else:
file1 = open("prompt.txt","r+")
text = file1.read()
print(text)
##Generating with stable diffusion file1 = open("prompt.txt","r+")
device = "cuda" text = file1.read()
model_path = "CompVis/stable-diffusion-v1-4" print(text)
# Using DDIMScheduler as anexample,this also works with PNDMScheduler ##Generating with stable diffusion
device = "cuda"
model_path = "CompVis/stable-diffusion-v1-4"
# Using DDIMScheduler as an example,this also works with PNDMScheduler
# uncomment this line if you want to use it. # uncomment this line if you want to use it.
#scheduler = PNDMScheduler.from_config(model_path, subfolder="scheduler", use_auth_token=True) #scheduler = PNDMScheduler.from_config(model_path, subfolder="scheduler", use_auth_token=True)
scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False) scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False)
pipe = StableDiffusionImg2ImgPipeline.from_pretrained( pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
model_path, model_path,
scheduler=scheduler, scheduler=scheduler,
revision="fp16", revision="fp16",
torch_dtype=torch.float16, torch_dtype=torch.float16,
use_auth_token=True use_auth_token=True
).to(device) ).to(device)
if t <= 10500000000: if t <= 11000000000:
pipe.enable_attention_slicing() print("Less then 11gb video memory. Running with attention slicing.")
pipe.enable_attention_slicing()
def dummy_checker(images, **kwargs): def dummy_checker(images, **kwargs):
return images, False return images, False
pipe.safety_checker = dummy_checker pipe.safety_checker = dummy_checker
def preprocess(image): def preprocess(image):
w, h = image.size w, h = image.size
w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32
image = image.resize((w, h), Image.Resampling.LANCZOS) image = image.resize((w, h), Image.Resampling.LANCZOS)
image = np.array(image).astype(np.float32) / 255.0 image = np.array(image).astype(np.float32) / 255.0
image = image[None].transpose(0, 3, 1, 2) image = image[None].transpose(0, 3, 1, 2)
image = torch.from_numpy(image) image = torch.from_numpy(image)
return 2.*image - 1. return 2.*image - 1.
path_img = [] path_img = []
startStrength = 0.86 startStrength = 0.86
deltaStrength = 0.02 deltaStrength = 0.02
endStrength = 0.941 endStrength = 0.941
startScale = 7.5 startScale = 7.5
deltaScale = 2.5 deltaScale = 2.5
endScale = 15.0 endScale = 15.0
startSeed = 1022 startSeed = 1022
endSeed = 1024 endSeed = 1024
directory_in = "./data/input" directory_in = "./data/input"
if not os.path.exists(directory_in): if not os.path.exists(directory_in):
os.makedirs(directory_in) os.makedirs(directory_in)
for root, subdirectories, files in os.walk(directory_in): for root, subdirectories, files in os.walk(directory_in):
for filename in files: for filename in files:
if filename.endswith(".png"): if filename.endswith(".png"):
path_img.append(os.path.join(root, filename)) path_img.append(os.path.join(root, filename))
print("Found " +str(len(path_img)) +" pictures") print("Found " +str(len(path_img)) +" pictures")
start_time = time.time() start_time = time.time()
counterr=0 counterr=0
allwork=0 allwork=0
directory="./data/out/" + strftime("%Y-%m-%d_%H-%M-%S", gmtime()) + "/" directory="./data/out/" + strftime("%Y-%m-%d_%H-%M-%S") + "/"
if not os.path.exists(directory): if not os.path.exists(directory):
os.makedirs(directory) os.makedirs(directory)
with open(directory+"prompt.txt", 'w') as f: with open(directory+"prompt.txt", 'w') as f:
f.write(text) f.write(text)
shutil.copytree(directory_in, directory+"input") shutil.copytree(directory_in, directory+"input")
for i in path_img: for i in path_img:
print(i) print(i)
if not os.path.exists(directory+str(counterr)): if not os.path.exists(directory+str(counterr)):
os.makedirs(directory+str(counterr)) os.makedirs(directory+str(counterr))
init_img = Image.open(i) init_img = Image.open(i)
init_img = init_img.resize((768, 512)) init_img = init_img.resize((768, 512))
init_image = preprocess(init_img) init_image = preprocess(init_img)
prompt = text prompt = text
for seed in range(endSeed-startSeed+1): for seed in range(endSeed-startSeed+1):
generator = torch.Generator(device=device).manual_seed(startSeed+seed) generator = torch.Generator(device=device).manual_seed(startSeed+seed)
guidance_scale = startScale guidance_scale = startScale
while guidance_scale <= endScale: while guidance_scale <= endScale:
strenght = startStrength strenght = startStrength
while strenght <= endStrength: while strenght <= endStrength:
with autocast("cuda"): with autocast("cuda"):
image = pipe(prompt=prompt, init_image=init_image, strength=strenght, guidance_scale=guidance_scale, generator=generator)["sample"][0] image = pipe(prompt=prompt, init_image=init_image, strength=strenght, guidance_scale=guidance_scale, generator=generator)["sample"][0]
image.save(directory+str(counterr)+"/" + str(allwork) +".jpg") image.save(directory+str(counterr)+"/" + str(allwork) +".jpg")
exif_dict = piexif.load(directory+str(counterr)+"/" + str(allwork) +".jpg") exif_dict = piexif.load(directory+str(counterr)+"/" + str(allwork) +".jpg")
userCommentAsDict = {} userCommentAsDict = {}
userCommentAsDict['Seed']=str(seed+startSeed) userCommentAsDict['Seed']=str(seed+startSeed)
userCommentAsDict['Strength']=str(strenght) userCommentAsDict['Strength']=str(strenght)
userCommentAsDict['Guidance_scale']=str(guidance_scale) userCommentAsDict['Guidance_scale']=str(guidance_scale)
user_comment = piexif.helper.UserComment.dump(json.dumps(userCommentAsDict)) user_comment = piexif.helper.UserComment.dump(json.dumps(userCommentAsDict))
exif_dict["Exif"][piexif.ExifIFD.UserComment] = user_comment exif_dict["Exif"][piexif.ExifIFD.UserComment] = user_comment
exif_bytes = piexif.dump(exif_dict) exif_bytes = piexif.dump(exif_dict)
piexif.insert(exif_bytes, directory+str(counterr)+"/" + str(allwork) +".jpg") piexif.insert(exif_bytes, directory+str(counterr)+"/" + str(allwork) +".jpg")
allwork+=1 allwork+=1
strenght+=deltaStrength strenght+=deltaStrength
guidance_scale+=deltaScale guidance_scale+=deltaScale
counterr+=1 counterr+=1
print("Made " + str(allwork) + " pictures in " + str(time.time()-start_time) + " seconds") print("Made " + str(allwork) + " pictures in " + str(time.time()-start_time) + " seconds")