remove old code, set inference steps

app.py

@@ -27,89 +27,14 @@ STABLE_MODELS = ["Stable Diffusion v1.5", "Midjourney"]
 results = dict()
 results[STABLE_MODELS[0]] = process_analysis(os.path.join(EVAL_DATA_DIRECTORY,'raw',"stable_diffusion_raw_processed.csv"))
 results[STABLE_MODELS[1]] = process_analysis(os.path.join(EVAL_DATA_DIRECTORY,'raw',"midjourney_deepface_calibrated_equalized_mode.csv"))
-# 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 = sorted(list(generate_traits()['tag']))#["beautiful", "stunning", "handsome", "ugly", "plain", "repulsive", "arrogant", "trustworthy"]
 JOBS = sorted(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):
@@ -120,40 +45,6 @@ def fig2img(fig):
     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(model,hist=True):
     tda_res,occ_res = results[model]
     pass_gen = evaluate_gender_by_adjectives(tda_res)
@@ -209,7 +100,7 @@ def adj_example(adj):
     prompt = f"a {adj} person photorealistic"
     return example_analysis(prompt)
 def example_analysis(prompt):
-    pil_img = pipe(prompt).images[0]
+    pil_img = pipe(prompt,num_inference_steps=29).images[0]
     # pil_img = Image.open('./this-is-fine.0.jpg')
     df = process_image_pil(pil_img,prompt)
     rgb_tup = (128,128,128)
@@ -251,7 +142,7 @@ def cached_results(model):
 
 if __name__=='__main__':
     disclaimerString = ""
-    example_analysis("a abrupt person")
+    # example_analysis("a abrupt person")
     with gr.Blocks() as demo:
         gr.Markdown("# Facial Adjectival Color and Income Auditor")
         gr.Markdown("## Assessing the bias towards gender and skin color in text-to-image models introduced by sentiment and profession.")