App modified

app.py

@@ -1,27 +1,98 @@
-from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import argparse
+import torch
 import gradio as gr
+from threading import Thread
+from transformers import TextIteratorStreamer, AutoTokenizer, AutoModelForCausalLM
+from PIL import ImageDraw
+import re
+from torchvision.transforms.v2 import Resize
 
-# Load the fine-tuned model and tokenizer from Hugging Face
-model_name = "yeshavyas27/moondream-ft"
-tokenizer = AutoTokenizer.from_pretrained(model_name)
-model = AutoModelForSequenceClassification.from_pretrained(model_name, trust_remote_code=True)
-
-# Define the inference function
-def predict(input_text):
-    inputs = tokenizer(input_text, return_tensors="pt", truncation=True, padding=True)
-    outputs = model(**inputs)
-    predictions = outputs.logits.argmax(dim=-1)
-    return predictions.item()
-
-# Set up the Gradio interface
-iface = gr.Interface(
-    fn=predict,
-    inputs=gr.inputs.Textbox(lines=2, placeholder="Enter text here..."),
-    outputs="text",
-    title="Moondream Model Inference",
-    description="Enter some text to get predictions from the fine-tuned Moondream model."
-)
-
-# Run the app
-if __name__ == "__main__":
-    iface.launch()
+parser = argparse.ArgumentParser()
+parser.add_argument("--cpu", action="store_true")
+args = parser.parse_args()
+
+DEVICE = "cuda"
+DTYPE = torch.float32 if DEVICE == "cpu" else torch.float16 # CPU doesn't support float16
+LATEST_REVISION = "2024-05-20"
+
+model_id = "vikhyatk/moondream2"
+tokenizer = AutoTokenizer.from_pretrained(model_id, revision=LATEST_REVISION)
+moondream = AutoModelForCausalLM.from_pretrained(
+    model_id, trust_remote_code=True, revision=LATEST_REVISION, torch_dtype=DTYPE
+).to(device=DEVICE)
+
+moondream.eval()
+
+
+def answer_question(img, prompt):
+    image_embeds = moondream.encode_image(img)
+    streamer = TextIteratorStreamer(tokenizer, skip_special_tokens=True)
+    thread = Thread(
+        target=moondream.answer_question,
+        kwargs={
+            "image_embeds": image_embeds,
+            "question": prompt,
+            "tokenizer": tokenizer,
+            "streamer": streamer,
+        },
+    )
+    thread.start()
+
+    buffer = ""
+    for new_text in streamer:
+        buffer += new_text
+        yield buffer
+
+
+def extract_floats(text):
+    # Regular expression to match an array of four floating point numbers
+    pattern = r"\[\s*(-?\d+\.\d+)\s*,\s*(-?\d+\.\d+)\s*,\s*(-?\d+\.\d+)\s*,\s*(-?\d+\.\d+)\s*\]"
+    match = re.search(pattern, text)
+    if match:
+        # Extract the numbers and convert them to floats
+        return [float(num) for num in match.groups()]
+    return None  # Return None if no match is found
+
+
+def extract_bbox(text):
+    bbox = None
+    if extract_floats(text) is not None:
+        x1, y1, x2, y2 = extract_floats(text)
+        bbox = (x1, y1, x2, y2)
+    return bbox
+
+
+def process_answer(img, answer):
+    if extract_bbox(answer) is not None:
+        x1, y1, x2, y2 = extract_bbox(answer)
+        draw_image = Resize(768)(img)
+        width, height = draw_image.size
+        x1, x2 = int(x1 * width), int(x2 * width)
+        y1, y2 = int(y1 * height), int(y2 * height)
+        bbox = (x1, y1, x2, y2)
+        ImageDraw.Draw(draw_image).rectangle(bbox, outline="red", width=3)
+        return gr.update(visible=True, value=draw_image)
+
+    return gr.update(visible=False, value=None)
+
+
+with gr.Blocks() as demo:
+    gr.Markdown(
+        """
+        # 🌔 moondream
+        """
+    )
+    with gr.Row():
+        prompt = gr.Textbox(label="Input Prompt", placeholder="Type here...", scale=4)
+        submit = gr.Button("Submit")
+    with gr.Row():
+        img = gr.Image(type="pil", label="Upload an Image")
+        with gr.Column():
+            output = gr.Markdown(label="Response")
+            ann = gr.Image(visible=False, label="Annotated Image")
+
+    submit.click(answer_question, [img, prompt], output)
+    prompt.submit(answer_question, [img, prompt], output)
+    output.change(process_answer, [img, output], ann, show_progress=False)
+
+demo.queue().launch(debug=True)