app.py

import gradio as gr
import peft
from peft import LoraConfig
from transformers import AutoTokenizer,BitsAndBytesConfig, AutoModelForCausalLM, CLIPVisionModel, AutoProcessor
import torch
from peft import PeftModel
import torch.nn as nn
import whisperx

# Determine the appropriate device
device = "cuda" if torch.cuda.is_available() else "cpu"

# Set compute_type based on device capabilities
if device == "cuda" and torch.cuda.is_bf16_supported():
    compute_type = "float16"
elif device == "cuda":
    compute_type = "float32"
else:
    compute_type = "int8"


clip_model_name = "openai/clip-vit-base-patch32"
phi_model_name  = "microsoft/phi-2"
tokenizer  = AutoTokenizer.from_pretrained(phi_model_name, trust_remote_code=True)
processor  = AutoProcessor.from_pretrained(clip_model_name)
tokenizer.pad_token = tokenizer.eos_token
IMAGE_TOKEN_ID = 23893 # token for word comment
QA_TOKEN_ID = 50295 # token for qa
device = "cuda" if torch.cuda.is_available() else "cpu"
clip_embed = 768
phi_embed  = 2560
compute_type = "float16"
audio_batch_size = 16

class SimpleResBlock(nn.Module):
    def __init__(self, phi_embed):
        super().__init__()
        self.pre_norm = nn.LayerNorm(phi_embed)
        self.proj = nn.Sequential(
            nn.Linear(phi_embed, phi_embed),
            nn.GELU(),
            nn.Linear(phi_embed, phi_embed)
        )
    def forward(self, x):
        x = self.pre_norm(x)
        return x + self.proj(x)



# models
clip_model = CLIPVisionModel.from_pretrained(clip_model_name).to(device)
projection = torch.nn.Linear(clip_embed, phi_embed).to(device)
resblock = SimpleResBlock(phi_embed).to(device)
phi_model = AutoModelForCausalLM.from_pretrained(phi_model_name,trust_remote_code=True).to(device)
# Load the model with the appropriate compute_type
audio_model_size = "tiny"
try:
    audio_model = whisperx.load_model(audio_model_size, device, compute_type=compute_type)
    print(f"Model loaded successfully with compute_type: {compute_type}")
except ValueError as e:
    print(f"Error loading model: {e}")
    print("Falling back to int8 compute type")
    audio_model = whisperx.load_model(audio_model_size, device, compute_type="int8")

# load weights
model_to_merge = PeftModel.from_pretrained(phi_model,'./model_chkpt/lora_adaptor')
merged_model   = model_to_merge.merge_and_unload()
projection.load_state_dict(torch.load('./model_chkpt/step2_projection.pth',map_location=torch.device(device)))
resblock.load_state_dict(torch.load('./model_chkpt/step2_resblock.pth',map_location=torch.device(device)))

def model_generate_ans(img=None,img_audio=None,val_q=None):

    max_generate_length = 100
    val_combined_embeds = []
    
    with torch.no_grad():
    
        # image
        if img is not None:
            image_processed  = processor(images=img, return_tensors="pt").to(device)
            clip_val_outputs = clip_model(**image_processed).last_hidden_state[:,1:,:]
            val_image_embeds = projection(clip_val_outputs)
            val_image_embeds = resblock(val_image_embeds).to(torch.float16)
            
            img_token_tensor = torch.tensor(IMAGE_TOKEN_ID).to(device)
            img_token_embeds = merged_model.model.embed_tokens(img_token_tensor).unsqueeze(0).unsqueeze(0)

            val_combined_embeds.append(val_image_embeds)
            val_combined_embeds.append(img_token_embeds)

        # audio
        if img_audio is not None:
            audio_result = audio_model.transcribe(img_audio)
            audio_text = ''
            for seg in audio_result['segments']:
                audio_text += seg['text']
            audio_text = audio_text.strip()
            audio_tokens = tokenizer(audio_text, return_tensors="pt", return_attention_mask=False)['input_ids'].squeeze(0).to(device)
            audio_embeds    = merged_model.model.embed_tokens(audio_tokens).unsqueeze(0)
            val_combined_embeds.append(audio_embeds)
            
        # text question
        if len(val_q) != 0:
            val_q_tokenised = tokenizer(val_q, return_tensors="pt", return_attention_mask=False)['input_ids'].squeeze(0).to(device)
            val_q_embeds    = merged_model.model.embed_tokens(val_q_tokenised).unsqueeze(0)
            val_combined_embeds.append(val_q_embeds)


        if img_audio is not None or len(val_q) != 0: # add QA Token
            
            QA_token_tensor = torch.tensor(QA_TOKEN_ID).to(device)
            QA_token_embeds = merged_model.model.embed_tokens(QA_token_tensor).unsqueeze(0).unsqueeze(0)
            val_combined_embeds.append(QA_token_embeds)
            
        val_combined_embeds = torch.cat(val_combined_embeds,dim=1)
        predicted_caption = merged_model.generate(inputs_embeds=val_combined_embeds,
                                                  max_new_tokens=max_generate_length,
                                                  return_dict_in_generate = True)
    
        predicted_captions_decoded = tokenizer.batch_decode(predicted_caption.sequences[:, 1:])[0] 
        predicted_captions_decoded = predicted_captions_decoded.replace("<|endoftext|>", "")  
    
    return predicted_captions_decoded
    

with gr.Blocks() as demo:

    gr.Markdown(
    """
    # Chat with MultiModal GPT !
    Build using combining clip model and phi-2 model.
    """
    )

    # app GUI
    with gr.Row():
        with gr.Column():
            img_input    = gr.Image(label='Image',type="pil")
            img_audio    = gr.Audio(label="Audio Query", sources=['microphone', 'upload'], type='filepath')
            img_question = gr.Text(label ='Text Query')
        with gr.Column():
            img_answer   = gr.Text(label ='Answer')

    section_btn = gr.Button("Submit")
    section_btn.click(model_generate_ans, inputs=[img_input,img_audio,img_question], outputs=[img_answer])
    
if __name__ == "__main__":
    demo.launch()