#initial fork from https://huggingface.co/spaces/Manjushri/SD-2.1-CPU/raw/main/app.py, https://huggingface.co/spaces/bhautikj/sd_clip_bias import gradio as gr import torch import numpy as np import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt from PIL import Image import collections import numpy as np import pandas as pd import io from saac.prompt_generation.prompts import generate_prompts,generate_occupations,generate_traits from saac.prompt_generation.prompt_utils import score_prompt from saac.evaluation.eval_utils import generate_countplot, lumia_violinplot, process_analysis, generate_histplot from datasets import load_dataset from diffusers import DiffusionPipeline, PNDMScheduler device = "cuda" if torch.cuda.is_available() else "cpu" STABLE_MODELS = ["runwayml/stable-diffusion-v1-5", "Midjourney"] # TODO Make model selectable input scheduler = PNDMScheduler.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="scheduler", prediction_type="v_prediction") pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", scheduler=scheduler) pipe = pipe.to(device) tokenizer = pipe.tokenizer text_encoder = pipe.text_encoder # # AUTHORS NOTE: # Being absolutely explicit: the genders and ethnicities listed here # are reductive. This is done not to erase but make analysis here # simpler, tractable and easier to understand # GENDERS = ["male", "female"] ETHNICITIES = ["black", "white", "asian"] LOOKS = list(generate_traits()['tag'])#["beautiful", "stunning", "handsome", "ugly", "plain", "repulsive", "arrogant", "trustworthy"] JOBS = list(generate_occupations()['tag'])#["CEO", "doctor", "nurse", "cashier", "janitor", "engineer", "pilot", "dentist", "leader"] RENDERPREFIX = "a high quality photo of a" def echoToken(token): res = getMostSimilar(tokenizer, text_encoder, token) return ",".join(res) def getEmbeddingForToken(tokenizer, token): token_ids = tokenizer.encode(token)[1:-1] if len(token_ids) != 1: print(len(token_ids)) raise token_id = token_ids[0] return token_id, text_encoder.get_input_embeddings().weight.data[token_id].unsqueeze(0) def getMostSimilar(tokenizer, text_encoder, token, numResults=50): internal_embs = text_encoder.text_model.embeddings.token_embedding.weight tID, tok = getEmbeddingForToken(tokenizer, token) cos = torch.nn.CosineSimilarity(dim=1, eps=1e-6) scores = cos(internal_embs.to("cpu").to(torch.float32), tok.to("cpu").to(torch.float32)) sorted_scores, sorted_ids = torch.sort(scores, descending=True) best_ids = sorted_ids[0:numResults].detach().numpy() best_scores = sorted_scores[0:numResults].detach().numpy() res = [] for best_id, best_score in zip(best_ids, best_scores): #res.append((tokenizer.decode(best_id), best_score)) res.append("[" + tokenizer.decode(best_id) + "," + str(best_score) + "]") return res[1:] def computeTermSimilarity(tokenizer, text_encoder, termA, termB): inputs = tokenizer([termA, termB], padding=True, return_tensors="pt").to("cpu") outputs = text_encoder(**inputs) cos = torch.nn.CosineSimilarity(dim=-1, eps=1e-6) val = cos(outputs.pooler_output[0], outputs.pooler_output[1]).item() return float(val) def computeJob(tokenizer, text_encoder, job): res = {} neutralPrompt = " ".join([RENDERPREFIX, job]) titleText = neutralPrompt for gender in GENDERS: for ethnicity in ETHNICITIES: prompt = " ".join([RENDERPREFIX, ethnicity, gender, job]) val = computeTermSimilarity(tokenizer, text_encoder, prompt, neutralPrompt) res[prompt] = val return titleText, sorted(res.items(), reverse=True) def computeLook(tokenizer, text_encoder, look): res = {} titleText = " ".join([RENDERPREFIX, look, "[", "|".join(GENDERS), "]"]) for gender in GENDERS: neutralPromptGender = " ".join([RENDERPREFIX, look, gender]) for ethnicity in ETHNICITIES: prompt = " ".join([RENDERPREFIX, look, ethnicity, gender]) val = computeTermSimilarity(tokenizer, text_encoder, prompt, neutralPromptGender) res[prompt] = val return titleText, sorted(res.items(), reverse=True) # via https://stackoverflow.com/questions/57316491/how-to-convert-matplotlib-figure-to-pil-image-object-without-saving-image def fig2img(fig): """Convert a Matplotlib figure to a PIL Image and return it""" buf = io.BytesIO() fig.savefig(buf) buf.seek(0) img = Image.open(buf) return img def computePlot(title, results, scaleXAxis=True): x = list(map(lambda x:x[0], results)) y = list(map(lambda x:x[1], results)) fig, ax = plt.subplots(1, 1, figsize=(10, 5)) y_pos = np.arange(len(x)) hbars = ax.barh(y_pos, y, left=0, align='center') ax.set_yticks(y_pos, labels=x) ax.invert_yaxis() # labels read top-to-bottom ax.set_xlabel('Cosine similarity - take care to note compressed X-axis') ax.set_title('Similarity to "' + title + '"') # Label with specially formatted floats ax.bar_label(hbars, fmt='%.3f') minR = np.min(y) maxR = np.max(y) diffR = maxR-minR if scaleXAxis: ax.set_xlim(left=minR-0.1*diffR, right=maxR+0.1*diffR) else: ax.set_xlim(left=0.0, right=1.0) plt.tight_layout() plt.close() return fig2img(fig) def computeJobBias(job): title, results = computeJob(tokenizer, text_encoder, job) return computePlot(title, results) def computeLookBias(look): title, results = computeLook(tokenizer, text_encoder, look) return computePlot(title, results) def trait_graph(trait,hist=True): tda_res, occ_result = process_analysis() fig = None if not hist: fig = generate_countplot(tda_res, 'tda_sentiment_val', 'gender_detected_val', title='Gender Count by Trait Sentiment', xlabel='Trait Sentiment', ylabel='Count', legend_title='Gender') else: df = tda_res df['tda_sentiment_val'] = pd.Categorical(df['tda_sentiment_val'], ['very negative', 'negative', 'neutral', 'positive', 'very positive']) fig = generate_histplot(tda_res, 'tda_sentiment_val', 'gender_detected_val', title='Gender Distribution by Trait Sentiment', xlabel='Trait Sentiment', ylabel='Count', ) fig2 = lumia_violinplot(df = tda_res, x_col = 'tda_compound', rgb_col = 'skincolor', n_bins = 21, widths_val = 0.05, points_val = 100, x_label = 'TDA Sentiment', y_label = 'Skincolor Intensity', title = 'Skin Color Intensity, Binned by TDA Sentiment',) return fig2img(fig),fig2img(fig2) def occ_graph(occ): tda_res, occ_result = process_analysis() fig = generate_histplot(occ_result, 'a_median', 'gender_detected_val', title='Gender Distribution by Median Annual Salary', xlabel= 'Median Annual Salary', ylabel= 'Count',) fig2 = lumia_violinplot(df=occ_result, x_col='a_median', rgb_col='skincolor', n_bins=21, widths_val=7500.0, points_val=100, x_label='Median Salary', y_label='Skincolor Intensity', title='Skin Color Intensity, Binned by Median Salary') return fig2img(fig),fig2img(fig2) if __name__=='__main__': disclaimerString = "" jobInterface = gr.Interface(fn=occ_graph, inputs=[gr.Dropdown(JOBS, label="occupation")], outputs=['image','image'], description="Referencing a specific profession comes loaded with associations of gender and ethnicity." " Text to image models provide an opportunity to explicitly specify an underrepresented group, but first we must understand our default behavior.", title="How occupation affects txt2img gender and skin color representation", article = "To view how mentioning a particular occupation affects the gender and skin colors in faces of text to image generators, select a job." " Promotional materials, advertising, and even criminal sketches which do not explicitly specify a gender or ethnicity term will tend towards the displayed distributions.") affectInterface = gr.Interface(fn=trait_graph, inputs=[gr.Dropdown(LOOKS, label="trait")], outputs=['image','image'], description="Certain adjectives can reinforce harmful stereotypes associated with gender roles and ethnic backgrounds." "Text to image models provide an opportunity to understand how prompting a particular human expression could be triggering," " or why an uncommon combination might provide important examples to minorities without default representation.", title="How word sentiment affects txt2img gender and skin color representation", article = "To view how characterizing a person with a positive, negative, or neutral term influences the gender and skin color composition of AI-generated faces, select a direction.") jobInterfaceManual = gr.Interface(fn=score_prompt, inputs=[gr.inputs.Textbox()], outputs='text', description="Analyze prompt", title="Understand which prompts require further engineering to represent equally genders and skin colors", article = "Try modifying a trait or occupational prompt to produce a result in the minority representation!") toolInterface = gr.Interface(fn=lambda t: trait_graph(t,hist=False),inputs=[gr.Dropdown(STABLE_MODELS,label="text-to-image model")],outputs='image', title="How different models fare in gender and skin color representation across a variety of prompts", description="The training set, vocabulary, pre and post processing of generative AI tools doesn't treat everyone equally. " "Within a 95% margin of statistical error, the following tests expose bias in gender and skin color.", article="To learn more about this process, Visit the repo" ) gr.TabbedInterface( [jobInterface, affectInterface, jobInterfaceManual,toolInterface], ["Occupational Bias", "Adjectival Bias", "Prompt analysis",'FACIA model auditing'], title = "Text-to-Image Bias Explorer" ).launch()