set lower precision

app.py

@@ -28,7 +28,8 @@ 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"))
 
-scheduler = PNDMScheduler.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="scheduler", prediction_type="v_prediction")
+scheduler = PNDMScheduler.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="scheduler", prediction_type="v_prediction",revision="fp16",
+                    torch_dtype=torch.float16)
 pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", scheduler=scheduler)
 pipe = pipe.to(device)
 
@@ -100,7 +101,7 @@ def adj_example(adj):
     prompt = f"a {adj} person photorealistic"
     return example_analysis(prompt)
 def example_analysis(prompt):
-    pil_img = pipe(prompt,num_inference_steps=29).images[0]
+    pil_img = pipe(prompt,num_inference_steps=20).images[0]
     # pil_img = Image.open('./this-is-fine.0.jpg')
     df = process_image_pil(pil_img,prompt)
     rgb_tup = (128,128,128)
@@ -181,18 +182,20 @@ if __name__=='__main__':
 
                     inp = gr.Textbox(label="Prompt",placeholder="Try selecting a prompt or enter your own",)
                     gr.Markdown("If the above component is stuck, try switching between the dropdown options.")
-                    sent = gr.Dropdown(LOOKS,label="Trait")
-                    with gr.Accordion("Details",open=False):
+                    with gr.Tab("Trait/Sentiment"):
+                        sent = gr.Dropdown(LOOKS,label="Trait")
                         gr.Markdown("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. "
                                     "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 distributions in the Model Audit tab.")
-                    occs = gr.Dropdown(JOBS,label="Occupation")
-                    with gr.Accordion("Details",open=False):
+                        sent.change(fn=lambda k: f"a {k} person photorealistic", inputs=sent, outputs=inp)
+                    with gr.Tab("Occupation/Income"):
+                        occs = gr.Dropdown(JOBS,label="Occupation")
                         gr.Markdown("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."
                                     "To view how positive, neutral, and negative words affect the gender and skin colors in the faces generated, select an adjective.")
+                        occs.change(fn=lambda k: f"a {k} photorealistic", inputs=occs, outputs=inp, )
                     btn = gr.Button("Generate and Analyze")
                 with gr.Column():
 
@@ -202,8 +205,8 @@ if __name__=='__main__':
                         inten = gr.ColorPicker(label="Grayscale intensity")
                     img = gr.Image(label="Stable Diffusion v1.5")
                     sentscore = gr.Text(label="VADER sentiment score",interactive=False)
-            sent.change(fn=lambda k: f"a {k} person photorealistic", inputs=sent, outputs=inp)
-            occs.change(fn=lambda k: f"a {k} photorealistic", inputs=occs, outputs=inp,)
+
+
             btn.click(fn=example_analysis,inputs=inp,outputs=[img,gender,skin,inten,sentscore])
                 # inp.submit(fn=example_analysis, outputs=[img,gender,skin,inten])
 
@@ -229,4 +232,4 @@ if __name__=='__main__':
     #     ["Occupational Bias", "Adjectival Bias", "Prompt analysis",'FACIA model auditing'],
     #     title = "Text-to-Image Bias Explorer"
     # ).launch()
-    demo.launch()
+    demo.launch(enable_queue=True,)