pengdaqian commited on
Commit
badc295
1 Parent(s): 83c3831

warm up model

Browse files
Files changed (2) hide show
  1. Dockerfile +3 -0
  2. img_nsfw.py +11 -10
Dockerfile CHANGED
@@ -3,6 +3,9 @@ FROM python:3.8
3
  COPY requirements.txt requirements.txt
4
  RUN pip3 install -r requirements.txt
5
 
 
 
 
6
  WORKDIR $HOME/app
7
 
8
  COPY . $HOME/app
 
3
  COPY requirements.txt requirements.txt
4
  RUN pip3 install -r requirements.txt
5
 
6
+ ENV TRANSFORMERS_CACHE /var/cache/transformers
7
+ RUN transformers-cli download CompVis/stable-diffusion-v1-4
8
+
9
  WORKDIR $HOME/app
10
 
11
  COPY . $HOME/app
img_nsfw.py CHANGED
@@ -11,12 +11,13 @@ def init_nsfw_pipe():
11
  from torch import nn
12
 
13
  # make sure you're logged in with `huggingface-cli login`
14
- # if torch.cuda.is_available():
15
- # pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", revision="fp16",
16
- # torch_dtype=torch.float16)
17
- # pipe = pipe.to('cuda')
18
- # else:
19
- pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_dtype=torch.float32)
 
20
 
21
  def cosine_distance(image_embeds, text_embeds):
22
  normalized_image_embeds = nn.functional.normalize(image_embeds)
@@ -78,10 +79,10 @@ def check_nsfw(img, pipe):
78
  if isinstance(img, str):
79
  img = dbimutils.read_img_from_url(img)
80
  safety_checker_input = pipe.feature_extractor(images=img, return_tensors="pt")
81
- # if torch.cuda.is_available():
82
- # safety_checker_input = safety_checker_input.to('cuda')
83
- # else:
84
- safety_checker_input = safety_checker_input.to('cpu')
85
 
86
  from torch.cuda.amp import autocast
87
  with autocast():
 
11
  from torch import nn
12
 
13
  # make sure you're logged in with `huggingface-cli login`
14
+ if torch.cuda.is_available():
15
+ pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", revision="fp16",
16
+ torch_dtype=torch.float16)
17
+ pipe = pipe.to('cuda')
18
+ else:
19
+ pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4",
20
+ torch_dtype=torch.float32)
21
 
22
  def cosine_distance(image_embeds, text_embeds):
23
  normalized_image_embeds = nn.functional.normalize(image_embeds)
 
79
  if isinstance(img, str):
80
  img = dbimutils.read_img_from_url(img)
81
  safety_checker_input = pipe.feature_extractor(images=img, return_tensors="pt")
82
+ if torch.cuda.is_available():
83
+ safety_checker_input = safety_checker_input.to('cuda')
84
+ else:
85
+ safety_checker_input = safety_checker_input.to('cpu')
86
 
87
  from torch.cuda.amp import autocast
88
  with autocast():