Spaces:
Runtime error
Runtime error
File size: 2,493 Bytes
021fd45 704bddb 9cfec01 6d16e6e 9cfec01 b21f7ff 9cfec01 deb3cb9 9cfec01 deb3cb9 6d16e6e 9cfec01 6d16e6e 9cfec01 6d16e6e 021fd45 9cfec01 6d16e6e 9cfec01 6d16e6e 021fd45 6d16e6e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 |
import gradio as gr
import torch
import importlib
from transformers import AutoConfig, AutoProcessor, AutoModelForCausalLM
from transformers.models.llava.configuration_llava import LlavaConfig
# --- Diagnostic: Load the configuration ---
config = AutoConfig.from_pretrained("lmms-lab/LLaVA-Video-7B-Qwen2", trust_remote_code=True)
print("Configuration type:", type(config))
print("Configuration architectures:", config.architectures)
# Expecting the architecture name to be "LlavaQwenForCausalLM"
arch = config.architectures[0] # This should be "LlavaQwenForCausalLM"
# --- Dynamic Import: Retrieve the model class by name ---
# Import the module that (should) contain the custom model class.
module = importlib.import_module("transformers.models.llava.modeling_llava")
try:
model_cls = getattr(module, arch)
print("Successfully imported model class:", model_cls)
except AttributeError:
raise ImportError(f"Cannot find class {arch} in module transformers.models.llava.modeling_llava")
# --- Register the Custom Model Class ---
# This tells the auto loader that for LlavaConfig, use our dynamically imported model class.
AutoModelForCausalLM.register(LlavaConfig, model_cls)
# --- Load Processor and Model ---
processor = AutoProcessor.from_pretrained(
"lmms-lab/LLaVA-Video-7B-Qwen2",
trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
"lmms-lab/LLaVA-Video-7B-Qwen2",
trust_remote_code=True
)
# Move model to GPU if available
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
def analyze_video(video_path):
prompt = "Analyze this video of a concert and determine the moment when the crowd is most engaged."
# Process the text and video input
inputs = processor(text=prompt, video=video_path, return_tensors="pt")
inputs = {k: v.to(device) for k, v in inputs.items()}
# Generate output (assuming the custom model implements generate)
outputs = model.generate(**inputs, max_new_tokens=100)
answer = processor.decode(outputs[0], skip_special_tokens=True)
return answer
# Create the Gradio Interface
iface = gr.Interface(
fn=analyze_video,
inputs=gr.Video(label="Upload Concert/Event Video", type="filepath"),
outputs=gr.Textbox(label="Engagement Analysis"),
title="Crowd Engagement Analyzer",
description="Upload a video of a concert or event and the model will analyze the moment when the crowd is most engaged."
)
if __name__ == "__main__":
iface.launch()
|