app.py

from transformers import Blip2ForConditionalGeneration
from transformers import Blip2Processor
from peft import PeftModel
import streamlit as st
from PIL import Image
#import torch
import os

preprocess_ckp = "Salesforce/blip2-opt-2.7b" #Checkpoint path used for perprocess image
base_model_ckp = "./model/blip2-opt-2.7b-fp16-sharded" #Base model checkpoint path
peft_model_ckp = "./model/blip2_peft" #PEFT model checkpoint path
sample_img_path = "./sample_images/"

#init_model_required = True
#processor = None
#model = None

#def init_model():

    #if init_model_required:

#Preprocess input 
processor = Blip2Processor.from_pretrained(preprocess_ckp)

#Model   
#Inferance on GPU device. Will give error in CPU system, as "load_in_8bit" is an setting of bitsandbytes library and only works for GPU
#model = Blip2ForConditionalGeneration.from_pretrained(base_model_ckp, load_in_8bit = True, device_map = "auto") 

#Inferance on CPU device
model = Blip2ForConditionalGeneration.from_pretrained(base_model_ckp) 

model = PeftModel.from_pretrained(model, peft_model_ckp)

        #init_model_required = False

    

def main():

    st.title("Fashion Image Caption using BLIP2")

    #init_model()

    #Select few sample images for the catagory of cloths
    #option = st.selectbox('Select from sample an images', ('cap', 'tee', 'dress'), index = -1)
    #st.text("OR")
    file_name = st.file_uploader("Upload an image")

    #if file_name is None and option is not None:

    #    file_name = os.path.join(sample_img_path, option)

    if file_name is not None:

        image_col, caption_text = st.columns(2)

        image_col.header("Image")
        image = Image.open(file_name)
        image_col.image(image, use_column_width = True)

        #Preprocess the image
        #Inferance on GPU. When used this on GPU will get errors like: "slow_conv2d_cpu" not implemented for 'Half'" , " Input type (float) and bias type (struct c10::Half)"
        #inputs = processor(images = image, return_tensors = "pt").to('cuda', torch.float16)

        #Inferance on CPU 
        inputs = processor(images = image, return_tensors = "pt")

        pixel_values = inputs.pixel_values

        #Predict the caption for the imahe
        generated_ids = model.generate(pixel_values = pixel_values, max_length = 25)
        generated_caption = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]  

        #Output the predict text
        caption_text.header("Generated Caption")
        caption_text.text(generated_caption)

if __name__ == "__main__":
    main()