Update app.py

app.py

@@ -13,6 +13,8 @@ from peft import PeftModel
 import re
 import spaces
 from diffusers import StableDiffusionXLPipeline, DiffusionPipeline
+import anthropic
+import base64
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 print(device)
@@ -55,17 +57,14 @@ css="""
     margin: 0 auto;
     max-width: 520px;
 }
-
 #gen-container {
     margin: 0 auto;
     max-width: 640px;
 }
-
 #title-container {
     margin: 0 auto;
     max-width: 1340px;
 }
-
 #main-container {
     margin: 0 auto;
     max-width: 1340px;
@@ -141,7 +140,7 @@ comment_images = [
 ]
 
 comments = {'test.png': "Not sure about the concept, it's too straightforward. Though the boy looks kinda creepy which makes it exciting. the art style is pretty to look at. I like that the colors are muted, but wish they were a bit darker to make it more eerie and add depth.", 'comment_images/0.png': "Hate this with a passion. The colors are too vibrant and don't match at all. I hate these colors in general. The patterns are too abstract and contemporary. a 5-year-old could draw this. pass.", 'comment_images/1.png': "Woah I love the art style. The texture feels like old paper which is oh so beautiful. There are so many details to focus on. I love the expressive lines and how busy the composition is. Even though orange isn't my favorite, the greenish blue color of the water is so gorgeous.", 'comment_images/2.png': "I don't like how monochromatic and muted this one is. but the paperish texture is nice and the details are so intricate.", 'comment_images/3.png': "Oh super pretty! Looks so smooth and wet. Love the details and loose lines too. Feels mystical and magical and eerie. Also dark purples and blues? deep indigo? My fav ever. I'm here for it.", 'comment_images/4.png': "Love the art style. The uncanny vibe and nightmarish horror is so cool. Like its horror but if you squint you can't tell? Love the strange. wish it had more colors though. not a fan of greyscale.", 'comment_images/5.png': 'omg I hate this haha. what the hell. everything about it disgusts me so boring and childish ew.', 'comment_images/6.png': 'yessss. give it to the texture give it to the brushstrokes give it to the style. perfect. just wish the colors were less beige and more bold. I want an active nightmare. but kisses to the surrealism.'}
-comments = dict()
+#comments = dict()
 
 image_index = 0
 
@@ -205,23 +204,90 @@ def clear_comments():
     extract_vp_botton = gr.Button(f"Extract visual preference from {len(comments)} comments", interactive=len(comments) != 0)
     clear_botton = gr.Button("Clear comments", interactive=len(comments) != 0)
     return extract_vp_botton, clear_botton
+
+@spaces.GPU(duration=120)
+def extract_vp_from_vpe():
+    global comments
+        
+    prompt = """I will provide a set of artworks along with accompanying comments from a person. Analyze these artworks and the comments on them and identify artistic features such as present or mentioned colors, style, composition, mood, medium, texture, brushwork, lighting, shadow effects, perspective, and other noteworthy elements.
+Your task is to extract the artistic features the person likes and dislikes based on both the artworks' features and the person's comments. Focus solely on artistic aspects and refrain from considering subject matter.
+If the person expresses a preference for a specific aspect without clearly stating its category (e.g., appreciating the colors without specifying which colors), identify these specific features from the images directly to make the person's preference understandable without needing to see the artwork.
+Your output should consist of two concise lists of keywords: one listing the specific art features the person likes and another listing the specific features they dislike (specified in keyword format without using sentences).
+Here are the images and their corresponding comments:
+"""
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", 
+                "text": prompt},
+            ]
+        }
+    ]
+    images = []
+    comment_number = 1
+    for image in comments:
+        comment = comments[image]
+        image = Image.open(image)
+        images.append(image)
     
+        messages[0]["content"].append(
+            {"type": "image"}
+        )
+
+        messages[0]["content"].append(
+            {"type": "text", 
+             "text": f"Comment {comment_number}: {comment}"}
+        )
+        comment_number = comment_number + 1
+    
+    prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
+    inputs = processor(text=prompt, images=images, return_tensors="pt")
+    inputs = {k: v.to(device) for k, v in inputs.items()}
+    
+    generated_ids = vpe_model.generate(**inputs, max_new_tokens=2000, repetition_penalty=0.99, do_sample=False)
+    generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+    positive_vp, negative_vp = re.search('.* \nAssistant: Liked Art Features: (.*)\nDisliked Art Features: (.*)', generated_texts).groups()
+    return positive_vp, negative_vp
 
-@spaces.GPU(duration=200)
 def extract_vp():
     if valid_api == "":
-    
-        global comments
+        positive_vp, negative_vp = extract_vp_from_vpe()
         
-        prompt = """I will provide a set of artworks along with accompanying comments from a person. Analyze these artworks and the comments on them and identify artistic features such as present or mentioned colors, style, composition, mood, medium, texture, brushwork, lighting, shadow effects, perspective, and other noteworthy elements.
 
-Your task is to extract the artistic features the person likes and dislikes based on both the artworks' features and the person's comments. Focus solely on artistic aspects and refrain from considering subject matter.
+    else:
+        client = anthropic.Anthropic(
+            api_key=valid_api,
+        )
 
-If the person expresses a preference for a specific aspect without clearly stating its category (e.g., appreciating the colors without specifying which colors), identify these specific features from the images directly to make the person's preference understandable without needing to see the artwork.
+        prompt = """**Objective:**
+Analyze a set of artworks and accompanying comments from a person to identify artistic features they like and dislike.
 
-Your output should consist of two concise lists of keywords: one listing the specific art features the person likes and another listing the specific features they dislike (specified in keyword format without using sentences).
+**Steps:**
 
-Here are the images and their corresponding comments:
+1. **Analyze Artworks and Comments:**
+   - Examine each artwork for artistic features such as colors, style, composition, mood, medium, texture, brushwork, lighting, shadow effects, perspective, and other noteworthy elements.
+   - Review the accompanying comments to understand the person's preferences and opinions on these features.
+
+2. **Identify Preferences:**
+   - Extract artistic features that the person likes and dislikes based on the artworks' features and the comments.
+   - Focus solely on artistic aspects and ignore the subject matter.
+   - Convert the art features mentioned in the comments to well-known synonyms if needed.
+
+3. **Resolve Ambiguous Preferences:**
+   - If the person expresses a preference without clearly stating its category (e.g., "I like the style" without specifying which style), identify these specific features from the images directly.
+   - Make the person's preference understandable and independednt of the artworks.
+
+4. **Output Format:**
+   - Create two concise lists of keywords: one for features the person likes and another for features they dislike.
+   - Ensure the lists are in keyword format, divided by commas, without using sentences.
+   - Maintain detail and accuracy for all comments and images.
+
+**Your Task:**
+
+Follow the example format and ensure that your output consists of two lists of keywords summarizing the person's preferences based on the artworks and comments provided. Consider all comments and images comprehensively.
+
+**Example**: example START:
 """
         messages = [
             {
@@ -236,11 +302,30 @@ Here are the images and their corresponding comments:
         comment_number = 1
         for image in comments:
             comment = comments[image]
-            image = Image.open(image)
-            images.append(image)
+            if not image.lower().endswith(".jpg"):
+                jpg_image_path = image.replace("png", "jpg")
+                image = Image.open(image)
+                rgb_img = image.convert("RGB")
+                rgb_img.save(jpg_image_path, format="JPEG")
+                with open(jpg_image_path, "rb") as image_file:
+                    image = base64.b64encode(image_file.read()).decode("utf-8")
+            
+            else:
+                with open(image_path, "rb") as image_file:
+                    image = base64.b64encode(image_file.read()).decode("utf-8")
+
+            messages[0]["content"].append(
+                {"type": "text",
+                "text": f"Image {comment_number}:"}
+            )
     
             messages[0]["content"].append(
-                {"type": "image"}
+                {"type": "image",
+                "source": {
+                        "type": "base64",
+                        "media_type": "image/jpeg",
+                        "data": image,
+                    },}
             )
 
             messages[0]["content"].append(
@@ -248,17 +333,15 @@ Here are the images and their corresponding comments:
                  "text": f"Comment {comment_number}: {comment}"}
             )
             comment_number = comment_number + 1
-    
-        prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
-        inputs = processor(text=prompt, images=images, return_tensors="pt")
-        inputs = {k: v.to(device) for k, v in inputs.items()}
-        
-        generated_ids = vpe_model.generate(**inputs, max_new_tokens=2000, repetition_penalty=0.99, do_sample=False)
-        generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
-        positive_vp, negative_vp = re.search('.* \nAssistant: Liked Art Features: (.*)\nDisliked Art Features: (.*)', generated_texts).groups()
 
-    else:
-        PRINT(valid_api)
+        message = client.messages.create(
+            model="claude-3-5-sonnet-20240620",
+            max_tokens=1024,
+            messages=messages
+        )
+
+        generated_text = message.content[0].text
+        positive_vp, negative_vp = re.search('.*Like.*:\n(.*)\n*Dislike.*:\n(.*)', generated_text).groups()
     
     gr.Info("Visual preference successfully extracted.")
     
@@ -266,13 +349,27 @@ Here are the images and their corresponding comments:
 
 def api_fn(api):
     global valid_api
-    if api != "correct":
-        gr.Warning("Invalid API!")
-        valid_api = ""
-    else:
+    client = anthropic.Anthropic(
+        api_key=api,
+    )
+    try:
+        message = client.messages.create(
+            model="claude-3-5-sonnet-20240620",
+            max_tokens=1024,
+            messages=[
+                {"role": "user", "content": "Hello, Claude"}
+            ]
+        )
         gr.Info("Valid API")
+        print("correct")
         valid_api = api
 
+    except anthropic.AuthenticationError:
+        gr.Warning("Invalid API!")
+        valid_api = ""
+        
+        
+
 @spaces.GPU(duration=45)
 def generate(prompt, vp_pos, vp_neg, slider):
     if vp_pos == "" and vp_neg == "":
@@ -306,7 +403,6 @@ with gr.Blocks(css=css, title="ViPer Demo", theme=gr.themes.Base()) as demo:
                 \n
                 \n
                 \n
-
                 """)
     with gr.Row(elem_id="main-container"):
             
@@ -355,7 +451,7 @@ with gr.Blocks(css=css, title="ViPer Demo", theme=gr.themes.Base()) as demo:
             
             clear_botton = gr.Button("Clear comments", interactive=len(comments) != 0)
 
-            with gr.Accordion("Enter GPT API for Better Results (optional)", open=False):
+            with gr.Accordion("Enter Cluade API for Better Results (optional)", open=False):
                 with gr.Row():
                     api = gr.Text(
                         max_lines=1,
@@ -490,7 +586,4 @@ with gr.Blocks(css=css, title="ViPer Demo", theme=gr.themes.Base()) as demo:
         [comment_image, comment]
     )
 
-demo.launch(share=True)
-
-
-
+demo.launch(share=True)