app.py

import os
import torch
import torch.nn as nn
import pandas as pd
import torch.nn.functional as F
from lavis.models.protein_models.protein_function_opt import Blip2ProteinMistral
from lavis.models.base_model import FAPMConfig
import spaces
import gradio as gr
from esm_scripts.extract import run_demo
from esm import pretrained, FastaBatchedDataset

# from transformers import EsmTokenizer, EsmModel


# Load the model
# model = Blip2ProteinMistral(config=FAPMConfig(), esm_size='3b')
# model.load_checkpoint("model/checkpoint_mf2.pth")
# model.to('cuda')

# model_esm, alphabet = pretrained.load_model_and_alphabet('esm2_t36_3B_UR50D')
# model_esm.to('cuda')
# model_esm.eval()


# tokenizer = EsmTokenizer.from_pretrained("facebook/esm2_t36_3B_UR50D")
# model_esm = EsmModel.from_pretrained("facebook/esm2_t36_3B_UR50D")
# model_esm.to('cuda')
# model_esm.eval()

@spaces.GPU
def generate_caption(protein, prompt):
    # Process the image and the prompt
    # with open('/home/user/app/example.fasta', 'w') as f:
    #     f.write('>{}\n'.format("protein_name"))
    #     f.write('{}\n'.format(protein.strip()))
    # os.system("python esm_scripts/extract.py esm2_t36_3B_UR50D /home/user/app/example.fasta /home/user/app --repr_layers 36 --truncation_seq_length 1024 --include per_tok")
    # esm_emb = run_demo(protein_name='protein_name', protein_seq=protein,
    #                    model=model_esm, alphabet=alphabet,
    #                    include='per_tok', repr_layers=[36], truncation_seq_length=1024)

    protein_name = 'protein_name'
    protein_seq = protein
    include = 'per_tok'
    repr_layers = [36]
    truncation_seq_length = 1024
    toks_per_batch = 4096
    print("start")
    dataset = FastaBatchedDataset([protein_name], [protein_seq])
    print("dataset prepared")
    batches = dataset.get_batch_indices(toks_per_batch, extra_toks_per_seq=1)
    print("batches prepared")

    model_esm, alphabet = pretrained.load_model_and_alphabet('esm2_t36_3B_UR50D')
    model_esm.to('cuda')
    model_esm.eval()

    data_loader = torch.utils.data.DataLoader(
        dataset, collate_fn=alphabet.get_batch_converter(truncation_seq_length), batch_sampler=batches
    )
    print(f"Read sequences")
    return_contacts = "contacts" in include

    assert all(-(model_esm.num_layers + 1) <= i <= model_esm.num_layers for i in repr_layers)
    repr_layers = [(i + model_esm.num_layers + 1) % (model_esm.num_layers + 1) for i in repr_layers]

    with torch.no_grad():
        for batch_idx, (labels, strs, toks) in enumerate(data_loader):
            print(
                f"Processing {batch_idx + 1} of {len(batches)} batches ({toks.size(0)} sequences)"
            )
            if torch.cuda.is_available():
                toks = toks.to(device="cuda", non_blocking=True)
            out = model_esm(toks, repr_layers=repr_layers, return_contacts=return_contacts)
            logits = out["logits"].to(device="cpu")
            representations = {
                layer: t.to(device="cpu") for layer, t in out["representations"].items()
            }
            if return_contacts:
                contacts = out["contacts"].to(device="cpu")
            for i, label in enumerate(labels):
                result = {"label": label}
                truncate_len = min(truncation_seq_length, len(strs[i]))
                # Call clone on tensors to ensure tensors are not views into a larger representation
                # See https://github.com/pytorch/pytorch/issues/1995
                if "per_tok" in include:
                    result["representations"] = {
                        layer: t[i, 1: truncate_len + 1].clone()
                        for layer, t in representations.items()
                    }
                if "mean" in include:
                    result["mean_representations"] = {
                        layer: t[i, 1: truncate_len + 1].mean(0).clone()
                        for layer, t in representations.items()
                    }
                if "bos" in include:
                    result["bos_representations"] = {
                        layer: t[i, 0].clone() for layer, t in representations.items()
                    }
                if return_contacts:
                    result["contacts"] = contacts[i, : truncate_len, : truncate_len].clone()
            esm_emb = result['representations'][36]
    '''
    inputs = tokenizer([protein], return_tensors="pt", padding=True, truncation=True).to('cuda')
    with torch.no_grad():
        outputs = model_esm(**inputs)
    esm_emb = outputs.last_hidden_state.detach()[0]
    '''
    print("esm embedding generated")
    esm_emb = F.pad(esm_emb.t(), (0, 1024 - len(esm_emb))).t().to('cuda')
    print("esm embedding processed")
    samples = {'name': ['protein_name'],
               'image': torch.unsqueeze(esm_emb, dim=0),
               'text_input': ['none'],
               'prompt': [prompt]}

    del model_esm

    model = Blip2ProteinMistral(config=FAPMConfig(), esm_size='3b')
    model.load_checkpoint("model/checkpoint_mf2.pth")
    model.to('cuda')
    # Generate the output
    prediction = model.generate(samples, length_penalty=0., num_beams=15, num_captions=10, temperature=1.,
                                repetition_penalty=1.0)

    return prediction
    # return "test"


# Define the FAPM interface
description = """Quick demonstration of the FAPM model for protein function prediction. Upload an protein sequence to generate a function description. Modify the Prompt to provide the taxonomy information.

The model used in this app is available at [Hugging Face Model Hub](https://huggingface.co/wenkai/FAPM) and the source code can be found on [GitHub](https://github.com/xiangwenkai/FAPM/tree/main)."""

iface = gr.Interface(
    fn=generate_caption,
    inputs=[gr.Textbox(type="text", label="Upload sequence"), gr.Textbox(type="text", label="Prompt")],
    outputs=gr.Textbox(label="Generated description"),
    description=description
)

# Launch the interface
iface.launch()