Spaces:
Runtime error
Runtime error
Upload safety_checker.py
Browse files- safety_checker.py +80 -0
safety_checker.py
ADDED
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
|
5 |
+
from transformers import CLIPConfig, CLIPVisionModel, PreTrainedModel
|
6 |
+
|
7 |
+
from diffusers.utils import logging
|
8 |
+
|
9 |
+
|
10 |
+
logger = logging.get_logger(__name__)
|
11 |
+
|
12 |
+
|
13 |
+
def cosine_distance(image_embeds, text_embeds):
|
14 |
+
normalized_image_embeds = nn.functional.normalize(image_embeds)
|
15 |
+
normalized_text_embeds = nn.functional.normalize(text_embeds)
|
16 |
+
return torch.mm(normalized_image_embeds, normalized_text_embeds.T)
|
17 |
+
|
18 |
+
|
19 |
+
class StableDiffusionSafetyChecker(PreTrainedModel):
|
20 |
+
config_class = CLIPConfig
|
21 |
+
|
22 |
+
def __init__(self, config: CLIPConfig):
|
23 |
+
super().__init__(config)
|
24 |
+
|
25 |
+
self.vision_model = CLIPVisionModel(config.vision_config)
|
26 |
+
self.visual_projection = nn.Linear(config.vision_config.hidden_size, config.projection_dim, bias=False)
|
27 |
+
|
28 |
+
self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False)
|
29 |
+
self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False)
|
30 |
+
|
31 |
+
self.register_buffer("concept_embeds_weights", torch.ones(17))
|
32 |
+
self.register_buffer("special_care_embeds_weights", torch.ones(3))
|
33 |
+
|
34 |
+
@torch.no_grad()
|
35 |
+
def forward(self, clip_input, images):
|
36 |
+
pooled_output = self.vision_model(clip_input)[1] # pooled_output
|
37 |
+
image_embeds = self.visual_projection(pooled_output)
|
38 |
+
|
39 |
+
special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().numpy()
|
40 |
+
cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().numpy()
|
41 |
+
|
42 |
+
result = []
|
43 |
+
batch_size = image_embeds.shape[0]
|
44 |
+
for i in range(batch_size):
|
45 |
+
result_img = {"special_scores": {}, "special_care": [], "concept_scores": {}, "bad_concepts": []}
|
46 |
+
|
47 |
+
# increase this value to create a stronger `nfsw` filter
|
48 |
+
# at the cost of increasing the possibility of filtering benign images
|
49 |
+
adjustment = 0.0
|
50 |
+
|
51 |
+
for concet_idx in range(len(special_cos_dist[0])):
|
52 |
+
concept_cos = special_cos_dist[i][concet_idx]
|
53 |
+
concept_threshold = self.special_care_embeds_weights[concet_idx].item()
|
54 |
+
result_img["special_scores"][concet_idx] = round(concept_cos - concept_threshold + adjustment, 3)
|
55 |
+
if result_img["special_scores"][concet_idx] > 0:
|
56 |
+
result_img["special_care"].append({concet_idx, result_img["special_scores"][concet_idx]})
|
57 |
+
adjustment = 0.01
|
58 |
+
|
59 |
+
for concet_idx in range(len(cos_dist[0])):
|
60 |
+
concept_cos = cos_dist[i][concet_idx]
|
61 |
+
concept_threshold = self.concept_embeds_weights[concet_idx].item()
|
62 |
+
result_img["concept_scores"][concet_idx] = round(concept_cos - concept_threshold + adjustment, 3)
|
63 |
+
if result_img["concept_scores"][concet_idx] > 0:
|
64 |
+
result_img["bad_concepts"].append(concet_idx)
|
65 |
+
|
66 |
+
result.append(result_img)
|
67 |
+
|
68 |
+
has_nsfw_concepts = [len(res["bad_concepts"]) > 0 for res in result]
|
69 |
+
|
70 |
+
for idx, has_nsfw_concept in enumerate(has_nsfw_concepts):
|
71 |
+
if has_nsfw_concept:
|
72 |
+
images[idx] = np.zeros(images[idx].shape) # black image
|
73 |
+
|
74 |
+
if any(has_nsfw_concepts):
|
75 |
+
logger.warning(
|
76 |
+
"Potential NSFW content was detected in one or more images. A black image will be returned instead."
|
77 |
+
" Try again with a different prompt and/or seed."
|
78 |
+
)
|
79 |
+
|
80 |
+
return images, has_nsfw_concepts
|