import gradio as gr
import torch

from transformers import BlipForConditionalGeneration, BlipProcessor

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-large")
model_image_captioning = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large").to(device)

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

def inference(raw_image, question, decoding_strategy):    
    inputs = processor(images=raw_image, text=question, return_tensors="pt")

    if decoding_strategy == "Beam search":
        inputs["max_length"] = 20
        inputs["num_beams"] = 5
    elif decoding_strategy == "Nucleus sampling":
        inputs["max_length"] = 20
        inputs["num_beams"] = 1
        inputs["do_sample"] = True
        inputs["top_k"] = 50
        inputs["top_p"] = 0.95

    out = model_image_captioning.generate(**inputs)
    return processor.batch_decode(out, skip_special_tokens=True)[0]

inputs = [
        gr.inputs.Image(type='pil'),
        gr.inputs.Textbox(lines=2, label="Context (optional)"),
        gr.inputs.Radio(choices=['Beam search','Nucleus sampling'], type="value", default="Nucleus sampling", label="Caption Decoding Strategy")
    ]
outputs = gr.outputs.Textbox(label="Output")

title = "BLIP"

description = "Gradio demo for BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation (Salesforce Research). To use it, simply upload your image, or click one of the examples to load them. Read more at the links below."

article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2201.12086' target='_blank'>BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation</a> | <a href='https://github.com/salesforce/BLIP' target='_blank'>Github Repo</a></p>"

gr.Interface(inference, inputs, outputs, title=title, description=description, article=article).launch(enable_queue=True)