Spaces:
Runtime error
Runtime error
File size: 3,677 Bytes
27c35a5 0773c5a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 |
"""
file copy from diffusion library from Huggingface: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/safety_checker.py
"""
import numpy as np
import torch
import torch.nn as nn
from transformers import CLIPConfig, CLIPVisionModel, PreTrainedModel
from diffusers.utils import logging
logger = logging.get_logger(__name__)
def cosine_distance(image_embeds, text_embeds):
normalized_image_embeds = nn.functional.normalize(image_embeds)
normalized_text_embeds = nn.functional.normalize(text_embeds)
return torch.mm(normalized_image_embeds, normalized_text_embeds.T)
class StableDiffusionSafetyChecker(PreTrainedModel):
config_class = CLIPConfig
def __init__(self, config: CLIPConfig):
super().__init__(config)
self.vision_model = CLIPVisionModel(config.vision_config)
self.visual_projection = nn.Linear(config.vision_config.hidden_size, config.projection_dim, bias=False)
self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False)
self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False)
self.register_buffer("concept_embeds_weights", torch.ones(17))
self.register_buffer("special_care_embeds_weights", torch.ones(3))
@torch.no_grad()
def forward(self, clip_input, images):
pooled_output = self.vision_model(clip_input)[1] # pooled_output
image_embeds = self.visual_projection(pooled_output)
special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().numpy()
cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().numpy()
result = []
batch_size = image_embeds.shape[0]
for i in range(batch_size):
result_img = {"special_scores": {}, "special_care": [], "concept_scores": {}, "bad_concepts": []}
# increase this value to create a stronger `nfsw` filter
# at the cost of increasing the possibility of filtering benign images
adjustment = 0.0
for concet_idx in range(len(special_cos_dist[0])):
concept_cos = special_cos_dist[i][concet_idx]
concept_threshold = self.special_care_embeds_weights[concet_idx].item()
result_img["special_scores"][concet_idx] = round(concept_cos - concept_threshold + adjustment, 3)
if result_img["special_scores"][concet_idx] > 0:
result_img["special_care"].append({concet_idx, result_img["special_scores"][concet_idx]})
adjustment = 0.01
for concet_idx in range(len(cos_dist[0])):
concept_cos = cos_dist[i][concet_idx]
concept_threshold = self.concept_embeds_weights[concet_idx].item()
result_img["concept_scores"][concet_idx] = round(concept_cos - concept_threshold + adjustment, 3)
if result_img["concept_scores"][concet_idx] > 0:
result_img["bad_concepts"].append(concet_idx)
result.append(result_img)
has_nsfw_concepts = [len(res["bad_concepts"]) > 0 for res in result]
for idx, has_nsfw_concept in enumerate(has_nsfw_concepts):
if has_nsfw_concept:
images[idx] = np.zeros(images[idx].shape) # black image
if any(has_nsfw_concepts):
logger.warning(
"Potential NSFW content was detected in one or more images. A black image will be returned instead."
" Try again with a different prompt and/or seed."
)
return images, has_nsfw_concepts
|