app.py

import os
import gradio as gr
from prediction import run_sequence_prediction
import torch
import torchvision.transforms as T
from celle.utils import process_image
from celle_main import instantiate_from_config
from omegaconf import OmegaConf
from huggingface_hub import hf_hub_download

def bold_predicted_letters(input_string: str, output_string: str) -> str:
    result = []
    i = j = 0
    input_string = input_string.upper()
    output_string = output_string.upper()
    
    while i < len(input_string):
        if input_string[i:i+6] == "<MASK>":
            start_index = i
            end_index = i + 6
            while end_index < len(input_string) and input_string[end_index:end_index+6] == "<MASK>":
                end_index += 6
            
            result.append("**" + output_string[j:j+(end_index-start_index)//6] + "**")
            i = end_index
            j += (end_index-start_index)//6
        else:
            result.append(input_string[i])
            i += 1
            if input_string[i-1] != "<":
                j += 1
    
    return "".join(result)

class model:
    def __init__(self):
        self.model = None
        self.model_name = None

    def gradio_demo(self, model_name, sequence_input, image):
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        if self.model_name != model_name:
            self.model_name = model_name
            model_ckpt_path = hf_hub_download(repo_id=f"HuangLab/{model_name}", filename="model.ckpt")
            model_config_path = hf_hub_download(repo_id=f"HuangLab/{model_name}", filename="config.yaml")
            hf_hub_download(repo_id=f"HuangLab/{model_name}", filename="nucleus_vqgan.yaml")
            hf_hub_download(repo_id=f"HuangLab/{model_name}", filename="threshold_vqgan.yaml")

            # Load model config and set ckpt_path if not provided in config
            config = OmegaConf.load(model_config_path)
            if config["model"]["params"]["ckpt_path"] is None:
                config["model"]["params"]["ckpt_path"] = model_ckpt_path

            # Set condition_model_path and vqgan_model_path to None
            config["model"]["params"]["condition_model_path"] = None
            config["model"]["params"]["vqgan_model_path"] = None
            
            base_path = os.getcwd()

            os.chdir(os.path.dirname(model_ckpt_path))

            # Instantiate model from config and move to device
            self.model = instantiate_from_config(config.model).to(device)
            self.model = torch.compile(self.model,mode='max-autotune')
            
            os.chdir(base_path)
            
            
        if "Finetuned" in model_name:
            dataset = "OpenCell"

        else:
            dataset = "HPA"
            
            
        nucleus_image = image['image'].convert('L')
        protein_image = image['mask'].convert('L')

        to_tensor = T.ToTensor()
        nucleus_image = to_tensor(nucleus_image)
        protein_image = to_tensor(protein_image)
        stacked_images = torch.stack([nucleus_image, protein_image], dim=0)
        processed_images = process_image(stacked_images, dataset)

        nucleus_image = processed_images[0].unsqueeze(0)
        protein_image = processed_images[1].unsqueeze(0)
        protein_image = protein_image/torch.max(protein_image)
        protein_image = 1 - protein_image                

        formatted_predicted_sequence = run_sequence_prediction(
            sequence_input=sequence_input,
            nucleus_image=nucleus_image,
            protein_image=protein_image,
            model=self.model,
            device=device,
        )
        
        formatted_predicted_sequence = formatted_predicted_sequence[0]
        formatted_predicted_sequence = formatted_predicted_sequence.replace("<pad>","")
        formatted_predicted_sequence = formatted_predicted_sequence.replace("<cls>","")
        formatted_predicted_sequence = formatted_predicted_sequence.replace("<eos>","")

        formatted_predicted_sequence = bold_predicted_letters(sequence_input, formatted_predicted_sequence)                                                      
        return T.ToPILImage()(protein_image[0,0]), T.ToPILImage()(nucleus_image[0,0]), formatted_predicted_sequence

base_class = model()

with gr.Blocks(theme='gradio/soft') as demo:
    gr.Markdown("## Inputs")
    gr.Markdown("Select the prediction model. **Note the first run may take ~2-3 minutes, but will take 3-4 seconds afterwards.**")
    gr.Markdown(
        "- ```CELL-E_2_HPA_2560``` is a good general purpose model for various cell types using ICC-IF."
    )
    gr.Markdown(
        "- ```CELL-E_2_OpenCell_2560``` is trained on OpenCell and is good more live-cell predictions on HEK cells."
    )
    with gr.Row():
        model_name = gr.Dropdown(
            ["CELL-E_2_HPA_2560", "CELL-E_2_OpenCell_2560"],
            value="CELL-E_2_HPA_2560",
            label="Model Name",
        )
    with gr.Row():
        gr.Markdown(
            "Input the desired amino acid sequence. GFP is shown below by default. The sequence must include ```<mask>``` for a prediction to be run."
        )

    with gr.Row():
        sequence_input = gr.Textbox(
            value="M<mask><mask><mask><mask><mask>SKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTLVTTFSYGVQCFSRYPDHMKQHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLVNRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLADHYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITHGMDELYK",
            label="Sequence",
        )
    with gr.Row():
        gr.Markdown(
            "Uploading a nucleus image is necessary. A random crop of 256 x 256 will be applied if larger. We provide default images in [images](https://huggingface.co/spaces/HuangLab/CELL-E_2/tree/main/images). Draw the desired localization on top of the nucelus image."
        )

    with gr.Row().style(equal_height=True):
        nucleus_image = gr.Image(
            source="upload", 
            tool="sketch",
            label="Nucleus Image", 
            interactive=True,
            image_mode="L",
            type="pil"
        )
        
    with gr.Row():
        gr.Markdown("## Outputs")
        
    with gr.Row().style(equal_height=True):
        nucleus_crop = gr.Image(
            label="Nucleus Image (Crop)", 
            image_mode="L",
            type="pil"
        )

        mask = gr.Image(
            label="Threshold Image", 
            image_mode="L",
            type="pil"
        )
    with gr.Row():
        gr.Markdown("Sequence predictions are show below.")

    with gr.Row().style(equal_height=True):
        predicted_sequence = gr.Markdown(label='Predicted Sequence')


    with gr.Row():
        button = gr.Button("Run Model")

        inputs = [model_name, sequence_input, nucleus_image]

        outputs = [mask, nucleus_crop, predicted_sequence]

        button.click(base_class.gradio_demo, inputs, outputs)

demo.launch(enable_queue=True)