add sd2 and model passing to bias eval

app.py

@@ -23,10 +23,14 @@ from datasets import load_dataset
 from diffusers import DiffusionPipeline, PNDMScheduler,EulerDiscreteScheduler
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
-STABLE_MODELS = ["Stable Diffusion v1.5", "Midjourney"]
+STABLE_MODELS = {"Stable Diffusion v1.5":'stable', "Midjourney":'midjourney',"Stable Diffusion v2":"sd2"}
 results = dict()
-results[STABLE_MODELS[0]] = process_analysis(os.path.join(EVAL_DATA_DIRECTORY,'raw',"stable_diffusion_raw_processed.csv"),filtered=True)
-results[STABLE_MODELS[1]] = process_analysis(os.path.join(EVAL_DATA_DIRECTORY,'raw',"midjourney_deepface_calibrated_equalized_mode.csv"),filtered=True)
+results["stable"] = process_analysis(os.path.join(EVAL_DATA_DIRECTORY,'raw',"stable_diffusion_raw_processed.csv"),filtered=True,model="stable")
+results["midjourney"] = process_analysis(os.path.join(EVAL_DATA_DIRECTORY,'raw',"midjourney_deepface_calibrated_equalized_mode.csv"),filtered=True,model='midjourney')
+results["sd2"] = process_analysis(os.path.join(EVAL_DATA_DIRECTORY,'raw',"sd2_analysis.csv"),filtered=True,model='sd2')
+for m in results:
+    t,o = results[m]
+    print(m,len(t.index),len(o.index))
 
 # scheduler = PNDMScheduler.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="scheduler", prediction_type="v_prediction",revision="fp16",torch_dtype=torch.float16)
 esched = EulerDiscreteScheduler.from_pretrained("stabilityai/stable-diffusion-2-base",subfolder="scheduler")
@@ -48,7 +52,8 @@ def fig2img(fig):
 
 def trait_graph(model,hist=True):
     tda_res,occ_res = results[model]
-    pass_gen,pass_skin = evaluate_by_adjectives(adjective_df=tda_res)
+    # print(tda_res)
+    pass_gen,pass_skin = evaluate_by_adjectives(adjective_df=tda_res,model=model)
     fig = None
     if not hist:
         fig = generate_countplot(tda_res, 'tda_sentiment_val', 'gender_detected_val',
@@ -77,7 +82,7 @@ def trait_graph(model,hist=True):
     return pass_skin,pass_gen,fig2img(fig2),fig2img(fig)
 def occ_graph(model):
     tda_res,occ_result = results[model]
-    pass_gen,pass_skin = evaluate_by_occupation(occupation_df=occ_result)
+    pass_gen,pass_skin = evaluate_by_occupation(occupation_df=occ_result,model=model)
     fig = generate_histplot(occ_result, 'a_median', 'gender_detected_val',
                    title='Gender Distribution by Median Annualg Salary',
                    xlabel= 'Median Annual Salary',
@@ -100,20 +105,22 @@ def adj_example(adj):
     return example_analysis(prompt)
 def example_analysis(prompt):
     pil_img = pipe(prompt,num_inference_steps=20).images[0]
-    # pil_img = Image.open('./this-is-fine.0.jpg')
+    # pil_img = Image.open('./a_anesthesiologist_photorealistic.png')
     df = process_image_pil(pil_img,prompt)
     rgb_tup = (128,128,128)
     if "skin color" in df:
         rgb_tup = df["skin color"][0]
-
+    print('RGB tup',rgb_tup)
     def clamp(x):
         return max(0, min(int(x), 255))
     def hex_from_tup(in_tup):
         return "#{0:02x}{1:02x}{2:02x}".format(clamp(in_tup[0]), clamp(in_tup[1]), clamp(in_tup[2]))
     rgb_hex = hex_from_tup(rgb_tup)
     intensity_val = rgb_intensity(rgb_tup)
+    print('intensity_val',intensity_val)
     intense_hex = str(hex(int(intensity_val)))
     intense_hex = f"#{intense_hex}{intense_hex}{intense_hex}"
+    print(intense_hex)
     gender_w = float(df["gender.Woman"][0]) if "gender.Woman" in df else -1
     gender_m = float(df["gender.Man"][0]) if "gender.Man" in df else -1
     gender_str = f"Male ({gender_m}%)" if gender_m>gender_w else f"Female({gender_w}%)"
@@ -131,11 +138,16 @@ def bias_assessment(model):
            [(f"Skin color {'unbiased' if os else 'biased'} by Income/Occupation",boo_to_str(os))], \
            [(f"Gender {'unbiased' if og else 'biased'} by Income/Occupation",boo_to_str(og))],\
            osgraph,oggraph
-mj_analysis = bias_assessment("Midjourney")
-sd_analysis = bias_assessment("Stable Diffusion v1.5")
+
+mj_analysis = bias_assessment("midjourney")
+sd_analysis = bias_assessment("stable")
+sd2_analysis = bias_assessment("sd2")
 def cached_results(model):
-    if model=="Midjourney":
+    model = STABLE_MODELS[model]
+    if model=="midjourney":
         return mj_analysis
+    elif model=="sd2":
+        return sd2_analysis
     else:
         return sd_analysis
 
@@ -148,7 +160,7 @@ if __name__=='__main__':
         with gr.Tab("Model Audit"):
             with gr.Row():
                 with gr.Column():
-                    model = gr.Dropdown(STABLE_MODELS,label="Text-to-Image Model")
+                    model = gr.Dropdown(list(STABLE_MODELS.keys()),label="Text-to-Image Model",value="Midjourney")
                     btn = gr.Button("Assess Model Bias")
                     gr.Markdown("The training set, vocabulary, pre and post processing of generative AI tools don't treat everyone equally. "
                                 "Within a 95% margin of statistical error, the following tests expose bias in gender and skin color. To learn more about this process, <a href=\"http://github.com/TRSS-Research/SAAC.git\"/> Visit the repo</a>")