models_server.py

import argparse
import logging
import random
import uuid
import numpy as np
from transformers import pipeline
from diffusers import DiffusionPipeline, StableDiffusionControlNetPipeline, ControlNetModel, UniPCMultistepScheduler
from diffusers.utils import load_image
from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
from diffusers.utils import export_to_video
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5ForSpeechToSpeech
from transformers import BlipProcessor, BlipForConditionalGeneration
from transformers import TrOCRProcessor, VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
from datasets import load_dataset
from PIL import Image
import io
from torchvision import transforms
import torch
import torchaudio
from speechbrain.pretrained import WaveformEnhancement
import joblib
from huggingface_hub import hf_hub_url, cached_download
from transformers import AutoImageProcessor, TimesformerForVideoClassification
from transformers import MaskFormerFeatureExtractor, MaskFormerForInstanceSegmentation, AutoFeatureExtractor
from controlnet_aux import OpenposeDetector, MLSDdetector, HEDdetector, CannyDetector, MidasDetector
from controlnet_aux.open_pose.body import Body
from controlnet_aux.mlsd.models.mbv2_mlsd_large import MobileV2_MLSD_Large
from controlnet_aux.hed import Network
from transformers import DPTForDepthEstimation, DPTFeatureExtractor
import warnings
import time
from espnet2.bin.tts_inference import Text2Speech
import soundfile as sf
from asteroid.models import BaseModel
import traceback
import os
import yaml

warnings.filterwarnings("ignore")

parser = argparse.ArgumentParser()
parser.add_argument("--config", type=str, default="config.yaml")
args = parser.parse_args()

if __name__ != "__main__":
    args.config = "config.gradio.yaml"

logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)
handler = logging.StreamHandler()
handler.setLevel(logging.INFO)
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
handler.setFormatter(formatter)
logger.addHandler(handler)

config = yaml.load(open(args.config, "r"), Loader=yaml.FullLoader)

local_deployment = config["local_deployment"]
if config["inference_mode"] == "huggingface":
    local_deployment = "none"

PROXY = None
if config["proxy"]:
    PROXY = {
        "https": config["proxy"],
    }

start = time.time()

# local_models = "models/"
local_models = ""


def load_pipes(local_deployment):
    other_pipes = {}
    standard_pipes = {}
    controlnet_sd_pipes = {}
    if local_deployment in ["full"]:
        other_pipes = {

            # "Salesforce/blip-image-captioning-large": {
            #     "model": BlipForConditionalGeneration.from_pretrained(f"Salesforce/blip-image-captioning-large"),
            #     "processor": BlipProcessor.from_pretrained(f"Salesforce/blip-image-captioning-large"),
            #     "device": "cuda:0"
            # },
            "damo-vilab/text-to-video-ms-1.7b": {
                "model": DiffusionPipeline.from_pretrained(f"{local_models}damo-vilab/text-to-video-ms-1.7b", torch_dtype=torch.float16, variant="fp16"),
                "device": "cuda:0"
            },
            # "facebook/maskformer-swin-large-ade": {
            #     "model": MaskFormerForInstanceSegmentation.from_pretrained(f"facebook/maskformer-swin-large-ade"),
            #     "feature_extractor" : AutoFeatureExtractor.from_pretrained("facebook/maskformer-swin-large-ade"),
            #     "device": "cuda:0"
            # },
            # "microsoft/trocr-base-printed": {
            #     "processor": TrOCRProcessor.from_pretrained(f"microsoft/trocr-base-printed"),
            #     "model": VisionEncoderDecoderModel.from_pretrained(f"microsoft/trocr-base-printed"),
            #     "device": "cuda:0"
            # },
            # "microsoft/trocr-base-handwritten": {
            #     "processor": TrOCRProcessor.from_pretrained(f"microsoft/trocr-base-handwritten"),
            #     "model": VisionEncoderDecoderModel.from_pretrained(f"microsoft/trocr-base-handwritten"),
            #     "device": "cuda:0"
            # },
            "JorisCos/DCCRNet_Libri1Mix_enhsingle_16k": {
                "model": BaseModel.from_pretrained("JorisCos/DCCRNet_Libri1Mix_enhsingle_16k"),
                "device": "cuda:0"
            },

            # "CompVis/stable-diffusion-v1-4": {
            #     "model": DiffusionPipeline.from_pretrained(f"CompVis/stable-diffusion-v1-4"),
            #     "device": "cuda:0"
            # },
            # "stabilityai/stable-diffusion-2-1": {
            #     "model": DiffusionPipeline.from_pretrained(f"stabilityai/stable-diffusion-2-1"),
            #     "device": "cuda:0"
            # },

            # "microsoft/speecht5_tts":{
            #     "processor": SpeechT5Processor.from_pretrained(f"microsoft/speecht5_tts"),
            #     "model": SpeechT5ForTextToSpeech.from_pretrained(f"microsoft/speecht5_tts"),
            #     "vocoder":  SpeechT5HifiGan.from_pretrained(f"microsoft/speecht5_hifigan"),
            #     "embeddings_dataset": load_dataset(f"Matthijs/cmu-arctic-xvectors", split="validation"),
            #     "device": "cuda:0"
            # },
            # "speechbrain/mtl-mimic-voicebank": {
            #     "model": WaveformEnhancement.from_hparams(source="speechbrain/mtl-mimic-voicebank", savedir="models/mtl-mimic-voicebank"),
            #     "device": "cuda:0"
            # },
            "microsoft/speecht5_vc":{
                "processor": SpeechT5Processor.from_pretrained(f"{local_models}microsoft/speecht5_vc"),
                "model": SpeechT5ForSpeechToSpeech.from_pretrained(f"{local_models}microsoft/speecht5_vc"),
                "vocoder": SpeechT5HifiGan.from_pretrained(f"{local_models}microsoft/speecht5_hifigan"),
                "embeddings_dataset": load_dataset(f"{local_models}Matthijs/cmu-arctic-xvectors", split="validation"),
                "device": "cuda:0"
            },
            # "julien-c/wine-quality": {
            #     "model": joblib.load(cached_download(hf_hub_url("julien-c/wine-quality", "sklearn_model.joblib")))
            # },
            # "facebook/timesformer-base-finetuned-k400": {
            #     "processor": AutoImageProcessor.from_pretrained(f"facebook/timesformer-base-finetuned-k400"),
            #     "model": TimesformerForVideoClassification.from_pretrained(f"facebook/timesformer-base-finetuned-k400"),
            #     "device": "cuda:0"
            # },
            "facebook/maskformer-swin-base-coco": {
                "feature_extractor": MaskFormerFeatureExtractor.from_pretrained(f"{local_models}facebook/maskformer-swin-base-coco"),
                "model": MaskFormerForInstanceSegmentation.from_pretrained(f"{local_models}facebook/maskformer-swin-base-coco"),
                "device": "cuda:0"
            },
            "Intel/dpt-hybrid-midas": {
                "model": DPTForDepthEstimation.from_pretrained(f"{local_models}Intel/dpt-hybrid-midas", low_cpu_mem_usage=True),
                "feature_extractor": DPTFeatureExtractor.from_pretrained(f"{local_models}Intel/dpt-hybrid-midas"),
                "device": "cuda:0"
            }
        }

    if local_deployment in ["full", "standard"]:
        standard_pipes = {
            # "nlpconnect/vit-gpt2-image-captioning":{
            #     "model": VisionEncoderDecoderModel.from_pretrained(f"{local_models}nlpconnect/vit-gpt2-image-captioning"),
            #     "feature_extractor": ViTImageProcessor.from_pretrained(f"{local_models}nlpconnect/vit-gpt2-image-captioning"),
            #     "tokenizer": AutoTokenizer.from_pretrained(f"{local_models}nlpconnect/vit-gpt2-image-captioning"),
            #     "device": "cuda:0"
            # },
            "espnet/kan-bayashi_ljspeech_vits": {
                "model": Text2Speech.from_pretrained("espnet/kan-bayashi_ljspeech_vits"),
                "device": "cuda:0"
            },
            # "lambdalabs/sd-image-variations-diffusers": {
            #     "model": DiffusionPipeline.from_pretrained(f"{local_models}lambdalabs/sd-image-variations-diffusers"), #torch_dtype=torch.float16
            #     "device": "cuda:0"
            # },
            "runwayml/stable-diffusion-v1-5": {
                "model": DiffusionPipeline.from_pretrained(f"{local_models}runwayml/stable-diffusion-v1-5"),
                "device": "cuda:0"
            },
            # "superb/wav2vec2-base-superb-ks": {
            #     "model": pipeline(task="audio-classification", model=f"superb/wav2vec2-base-superb-ks"), 
            #     "device": "cuda:0"
            # },
            "openai/whisper-base": {
                "model": pipeline(task="automatic-speech-recognition", model=f"{local_models}openai/whisper-base"), 
                "device": "cuda:0"
            },
            # "microsoft/speecht5_asr": {
            #     "model": pipeline(task="automatic-speech-recognition", model=f"{local_models}microsoft/speecht5_asr"), 
            #     "device": "cuda:0"
            # },
            "Intel/dpt-large": {
                "model": pipeline(task="depth-estimation", model=f"{local_models}Intel/dpt-large"), 
                "device": "cuda:0"
            },
            # "microsoft/beit-base-patch16-224-pt22k-ft22k": {
            #     "model": pipeline(task="image-classification", model=f"microsoft/beit-base-patch16-224-pt22k-ft22k"), 
            #     "device": "cuda:0"
            # },
            "facebook/detr-resnet-50-panoptic": {
                "model": pipeline(task="image-segmentation", model=f"{local_models}facebook/detr-resnet-50-panoptic"), 
                "device": "cuda:0"
            },
            "facebook/detr-resnet-101": {
                "model": pipeline(task="object-detection", model=f"{local_models}facebook/detr-resnet-101"), 
                "device": "cuda:0"
            },
            # "openai/clip-vit-large-patch14": {
            #     "model": pipeline(task="zero-shot-image-classification", model=f"openai/clip-vit-large-patch14"), 
            #     "device": "cuda:0"
            # },
            # "google/owlvit-base-patch32": {
            #     "model": pipeline(task="zero-shot-object-detection", model=f"{local_models}google/owlvit-base-patch32"), 
            #     "device": "cuda:0"
            # },
            # "microsoft/DialoGPT-medium": {
            #     "model": pipeline(task="conversational", model=f"microsoft/DialoGPT-medium"), 
            #     "device": "cuda:0"
            # },
            # "bert-base-uncased": {
            #     "model": pipeline(task="fill-mask", model=f"bert-base-uncased"), 
            #     "device": "cuda:0"
            # },
            # "deepset/roberta-base-squad2": {
            #     "model": pipeline(task = "question-answering", model=f"deepset/roberta-base-squad2"), 
            #     "device": "cuda:0"
            # },
            # "facebook/bart-large-cnn": {
            #     "model": pipeline(task="summarization", model=f"facebook/bart-large-cnn"), 
            #     "device": "cuda:0"
            # },
            # "google/tapas-base-finetuned-wtq": {
            #     "model": pipeline(task="table-question-answering", model=f"google/tapas-base-finetuned-wtq"), 
            #     "device": "cuda:0"
            # },
            # "distilbert-base-uncased-finetuned-sst-2-english": {
            #     "model": pipeline(task="text-classification", model=f"distilbert-base-uncased-finetuned-sst-2-english"), 
            #     "device": "cuda:0"
            # },
            # "gpt2": {
            #     "model": pipeline(task="text-generation", model="gpt2"), 
            #     "device": "cuda:0"
            # },
            # "mrm8488/t5-base-finetuned-question-generation-ap": {
            #     "model": pipeline(task="text2text-generation", model=f"mrm8488/t5-base-finetuned-question-generation-ap"), 
            #     "device": "cuda:0"
            # },
            # "Jean-Baptiste/camembert-ner": {
            #     "model": pipeline(task="token-classification", model=f"Jean-Baptiste/camembert-ner", aggregation_strategy="simple"), 
            #     "device": "cuda:0"
            # },
            # "t5-base": {
            #     "model": pipeline(task="translation", model=f"t5-base"), 
            #     "device": "cuda:0"
            # },
            "impira/layoutlm-document-qa": {
                "model": pipeline(task="document-question-answering", model=f"{local_models}impira/layoutlm-document-qa"), 
                "device": "cuda:0"
            },
            "ydshieh/vit-gpt2-coco-en": {
                "model": pipeline(task="image-to-text", model=f"{local_models}ydshieh/vit-gpt2-coco-en"), 
                "device": "cuda:0"
            },
            "dandelin/vilt-b32-finetuned-vqa": {
                "model": pipeline(task="visual-question-answering", model=f"{local_models}dandelin/vilt-b32-finetuned-vqa"), 
                "device": "cuda:0"
            }
        }

    if local_deployment in ["full", "standard", "minimal"]:
        
        controlnet = ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
        controlnetpipe = StableDiffusionControlNetPipeline.from_pretrained(
            f"{local_models}runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
        )


        hed_network = HEDdetector.from_pretrained('lllyasviel/ControlNet')

        controlnet_sd_pipes = {
            "openpose-control": {
                "model": OpenposeDetector.from_pretrained('lllyasviel/ControlNet')
            },
            "mlsd-control": {
                "model": MLSDdetector.from_pretrained('lllyasviel/ControlNet')
            },
            "hed-control": {
                "model": hed_network
            },
            "scribble-control": {
                "model": hed_network
            },
            "midas-control": {
                "model": MidasDetector.from_pretrained('lllyasviel/ControlNet')
            },
            "canny-control": {
                "model": CannyDetector()
            },
            "lllyasviel/sd-controlnet-canny":{
                "control": controlnet, 
                "model": controlnetpipe,
                "device": "cuda:0"
            },
            "lllyasviel/sd-controlnet-depth":{
                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-depth", torch_dtype=torch.float16),
                "model": controlnetpipe,
                "device": "cuda:0"
            },
            "lllyasviel/sd-controlnet-hed":{
                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-hed", torch_dtype=torch.float16), 
                "model": controlnetpipe,
                "device": "cuda:0"
            },
            "lllyasviel/sd-controlnet-mlsd":{
                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-mlsd", torch_dtype=torch.float16), 
                "model": controlnetpipe,
                "device": "cuda:0"
            },
            "lllyasviel/sd-controlnet-openpose":{
                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-openpose", torch_dtype=torch.float16), 
                "model": controlnetpipe,
                "device": "cuda:0"
            },
            "lllyasviel/sd-controlnet-scribble":{
                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-scribble", torch_dtype=torch.float16), 
                "model": controlnetpipe,
                "device": "cuda:0"
            },
            "lllyasviel/sd-controlnet-seg":{
                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-seg", torch_dtype=torch.float16), 
                "model": controlnetpipe,
                "device": "cuda:0"
            }    
        }
    pipes = {**standard_pipes, **other_pipes, **controlnet_sd_pipes}
    return pipes

pipes = load_pipes(local_deployment)

end = time.time()
during = end - start

print(f"[ ready ] {during}s")

def running():
    return {"running": True}

def status(model_id):
    disabled_models = ["microsoft/trocr-base-printed", "microsoft/trocr-base-handwritten"]
    if model_id in pipes.keys() and model_id not in disabled_models:
        print(f"[ check {model_id} ] success")
        return {"loaded": True}
    else:
        print(f"[ check {model_id} ] failed")
        return {"loaded": False}

def models(model_id, data):
    while "using" in pipes[model_id] and pipes[model_id]["using"]:
        print(f"[ inference {model_id} ] waiting")
        time.sleep(0.1)
    pipes[model_id]["using"] = True
    print(f"[ inference {model_id} ] start")

    start = time.time()

    pipe = pipes[model_id]["model"]
    
    if "device" in pipes[model_id]:
        try:
            pipe.to(pipes[model_id]["device"])
        except:
            pipe.device = torch.device(pipes[model_id]["device"])
            pipe.model.to(pipes[model_id]["device"])
    
    result = None
    try:
        # text to video
        if model_id == "damo-vilab/text-to-video-ms-1.7b":
            pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
            # pipe.enable_model_cpu_offload()
            prompt = data["text"]
            video_frames = pipe(prompt, num_inference_steps=50, num_frames=40).frames
            file_name = str(uuid.uuid4())[:4]
            video_path = export_to_video(video_frames, f"public/videos/{file_name}.mp4")
            
            new_file_name = str(uuid.uuid4())[:4]
            os.system(f"ffmpeg -i {video_path} -vcodec libx264 public/videos/{new_file_name}.mp4")

            if os.path.exists(f"public/videos/{new_file_name}.mp4"): 
                result = {"path": f"/videos/{new_file_name}.mp4"}
            else:
                result = {"path": f"/videos/{file_name}.mp4"}

        # controlnet
        if model_id.startswith("lllyasviel/sd-controlnet-"):
            pipe.controlnet.to('cpu')
            pipe.controlnet = pipes[model_id]["control"].to(pipes[model_id]["device"])
            pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
            control_image = load_image(data["img_url"])
            # generator = torch.manual_seed(66)
            out_image: Image = pipe(data["text"], num_inference_steps=20, image=control_image).images[0]
            file_name = str(uuid.uuid4())[:4]
            out_image.save(f"public/images/{file_name}.png")
            result = {"path": f"/images/{file_name}.png"}

        if model_id.endswith("-control"):
            image = load_image(data["img_url"])
            if "scribble" in model_id:
                control = pipe(image, scribble = True)
            elif "canny" in model_id:
                control = pipe(image, low_threshold=100, high_threshold=200)
            else:
                control = pipe(image)
            file_name = str(uuid.uuid4())[:4]
            control.save(f"public/images/{file_name}.png")
            result = {"path": f"/images/{file_name}.png"}

        # image to image
        if model_id == "lambdalabs/sd-image-variations-diffusers":
            im = load_image(data["img_url"])
            file_name = str(uuid.uuid4())[:4]
            with open(f"public/images/{file_name}.png", "wb") as f:
                f.write(data)
            tform = transforms.Compose([
                transforms.ToTensor(),
                transforms.Resize(
                    (224, 224),
                    interpolation=transforms.InterpolationMode.BICUBIC,
                    antialias=False,
                    ),
                transforms.Normalize(
                [0.48145466, 0.4578275, 0.40821073],
                [0.26862954, 0.26130258, 0.27577711]),
            ])
            inp = tform(im).to(pipes[model_id]["device"]).unsqueeze(0)
            out = pipe(inp, guidance_scale=3)
            out["images"][0].save(f"public/images/{file_name}.jpg")
            result = {"path": f"/images/{file_name}.jpg"}

        # image to text
        if model_id == "Salesforce/blip-image-captioning-large":
            raw_image = load_image(data["img_url"]).convert('RGB')
            text = data["text"]
            inputs = pipes[model_id]["processor"](raw_image, return_tensors="pt").to(pipes[model_id]["device"])
            out = pipe.generate(**inputs)
            caption = pipes[model_id]["processor"].decode(out[0], skip_special_tokens=True)
            result = {"generated text": caption}
        if model_id == "ydshieh/vit-gpt2-coco-en":
            img_url = data["img_url"]
            generated_text = pipe(img_url)[0]['generated_text']
            result = {"generated text": generated_text}
        if model_id == "nlpconnect/vit-gpt2-image-captioning":
            image = load_image(data["img_url"]).convert("RGB")
            pixel_values = pipes[model_id]["feature_extractor"](images=image, return_tensors="pt").pixel_values
            pixel_values = pixel_values.to(pipes[model_id]["device"])
            generated_ids = pipe.generate(pixel_values, **{"max_length": 200, "num_beams": 1})
            generated_text = pipes[model_id]["tokenizer"].batch_decode(generated_ids, skip_special_tokens=True)[0]
            result = {"generated text": generated_text}
        # image to text: OCR
        if model_id == "microsoft/trocr-base-printed" or  model_id == "microsoft/trocr-base-handwritten":
            image = load_image(data["img_url"]).convert("RGB")
            pixel_values = pipes[model_id]["processor"](image, return_tensors="pt").pixel_values
            pixel_values = pixel_values.to(pipes[model_id]["device"])
            generated_ids = pipe.generate(pixel_values)
            generated_text = pipes[model_id]["processor"].batch_decode(generated_ids, skip_special_tokens=True)[0]
            result = {"generated text": generated_text}

        # text to image
        if model_id == "runwayml/stable-diffusion-v1-5":
            file_name = str(uuid.uuid4())[:4]
            text = data["text"]
            out = pipe(prompt=text)
            out["images"][0].save(f"public/images/{file_name}.jpg")
            result = {"path": f"/images/{file_name}.jpg"}

        # object detection
        if model_id == "google/owlvit-base-patch32" or model_id == "facebook/detr-resnet-101":
            img_url = data["img_url"]
            open_types = ["cat", "couch", "person", "car", "dog", "horse", "sheep", "cow", "elephant", "bear", "zebra", "giraffe", "backpack", "umbrella", "handbag", "tie", "suitcase", "frisbee", "skis", "snowboard", "sports ball", "kite", "baseball bat", "baseball glove", "skateboard", "surfboard", "tennis racket", "bottle", "wine glass", "cup", "fork", "knife", "spoon", "bowl", "banana", "apple", "sandwich", "orange", "broccoli", "carrot", "hot dog", "pizza", "donut", "cake", "chair", "couch", "potted plant", "bed", "dining table", "toilet", "tv", "laptop", "mouse", "remote", "keyboard", "cell phone", "microwave", "oven", "toaster", "sink", "refrigerator", "book", "clock", "vase", "scissors", "teddy bear", "hair drier", "toothbrush", "traffic light", "fire hydrant", "stop sign", "parking meter", "bench", "bird"]
            result = pipe(img_url, candidate_labels=open_types)
        
        # VQA
        if model_id == "dandelin/vilt-b32-finetuned-vqa":
            question = data["text"]
            img_url = data["img_url"]
            result = pipe(question=question, image=img_url)
        
        #DQA
        if model_id == "impira/layoutlm-document-qa":
            question = data["text"]
            img_url = data["img_url"]
            result = pipe(img_url, question)

        # depth-estimation
        if model_id == "Intel/dpt-large":
            output = pipe(data["img_url"])
            image = output['depth']
            name = str(uuid.uuid4())[:4]
            image.save(f"public/images/{name}.jpg")
            result = {"path": f"/images/{name}.jpg"}

        if model_id == "Intel/dpt-hybrid-midas" and model_id == "Intel/dpt-large":
            image = load_image(data["img_url"])
            inputs = pipes[model_id]["feature_extractor"](images=image, return_tensors="pt")
            with torch.no_grad():
                outputs = pipe(**inputs)
                predicted_depth = outputs.predicted_depth
            prediction = torch.nn.functional.interpolate(
                predicted_depth.unsqueeze(1),
                size=image.size[::-1],
                mode="bicubic",
                align_corners=False,
            )
            output = prediction.squeeze().cpu().numpy()
            formatted = (output * 255 / np.max(output)).astype("uint8")
            image = Image.fromarray(formatted)
            name = str(uuid.uuid4())[:4]
            image.save(f"public/images/{name}.jpg")
            result = {"path": f"/images/{name}.jpg"}

        # TTS
        if model_id == "espnet/kan-bayashi_ljspeech_vits":
            text = data["text"]
            wav = pipe(text)["wav"]
            name = str(uuid.uuid4())[:4]
            sf.write(f"public/audios/{name}.wav", wav.cpu().numpy(), pipe.fs, "PCM_16")
            result = {"path": f"/audios/{name}.wav"}

        if model_id == "microsoft/speecht5_tts":
            text = data["text"]
            inputs = pipes[model_id]["processor"](text=text, return_tensors="pt")
            embeddings_dataset = pipes[model_id]["embeddings_dataset"]
            speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0).to(pipes[model_id]["device"])
            pipes[model_id]["vocoder"].to(pipes[model_id]["device"])
            speech = pipe.generate_speech(inputs["input_ids"].to(pipes[model_id]["device"]), speaker_embeddings, vocoder=pipes[model_id]["vocoder"])
            name = str(uuid.uuid4())[:4]
            sf.write(f"public/audios/{name}.wav", speech.cpu().numpy(), samplerate=16000)
            result = {"path": f"/audios/{name}.wav"}

        # ASR
        if model_id == "openai/whisper-base" or model_id == "microsoft/speecht5_asr":
            audio_url = data["audio_url"]
            result = { "text": pipe(audio_url)["text"]}

        # audio to audio
        if model_id == "JorisCos/DCCRNet_Libri1Mix_enhsingle_16k":
            audio_url = data["audio_url"]
            wav, sr = torchaudio.load(audio_url)
            with torch.no_grad():
                result_wav = pipe(wav.to(pipes[model_id]["device"]))
            name = str(uuid.uuid4())[:4]
            sf.write(f"public/audios/{name}.wav", result_wav.cpu().squeeze().numpy(), sr)
            result = {"path": f"/audios/{name}.wav"}
        
        if model_id == "microsoft/speecht5_vc":
            audio_url = data["audio_url"]
            wav, sr = torchaudio.load(audio_url)
            inputs = pipes[model_id]["processor"](audio=wav, sampling_rate=sr, return_tensors="pt")
            embeddings_dataset = pipes[model_id]["embeddings_dataset"]
            speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
            pipes[model_id]["vocoder"].to(pipes[model_id]["device"])
            speech = pipe.generate_speech(inputs["input_ids"].to(pipes[model_id]["device"]), speaker_embeddings, vocoder=pipes[model_id]["vocoder"])
            name = str(uuid.uuid4())[:4]
            sf.write(f"public/audios/{name}.wav", speech.cpu().numpy(), samplerate=16000)
            result = {"path": f"/audios/{name}.wav"}
        
        # segmentation
        if model_id == "facebook/detr-resnet-50-panoptic":
            result = []
            segments = pipe(data["img_url"])
            image = load_image(data["img_url"])

            colors = []
            for i in range(len(segments)):
                colors.append((random.randint(100, 255), random.randint(100, 255), random.randint(100, 255), 50))

            for segment in segments:
                mask = segment["mask"]
                mask = mask.convert('L')
                layer = Image.new('RGBA', mask.size, colors[i])
                image.paste(layer, (0, 0), mask)
            name = str(uuid.uuid4())[:4]
            image.save(f"public/images/{name}.jpg")
            result = {"path": f"/images/{name}.jpg"}

        if model_id == "facebook/maskformer-swin-base-coco" or model_id == "facebook/maskformer-swin-large-ade":
            image = load_image(data["img_url"])
            inputs = pipes[model_id]["feature_extractor"](images=image, return_tensors="pt").to(pipes[model_id]["device"])
            outputs = pipe(**inputs)
            result = pipes[model_id]["feature_extractor"].post_process_panoptic_segmentation(outputs, target_sizes=[image.size[::-1]])[0]
            predicted_panoptic_map = result["segmentation"].cpu().numpy()
            predicted_panoptic_map = Image.fromarray(predicted_panoptic_map.astype(np.uint8))
            name = str(uuid.uuid4())[:4]
            predicted_panoptic_map.save(f"public/images/{name}.jpg")
            result = {"path": f"/images/{name}.jpg"}

    except Exception as e:
        print(e)
        traceback.print_exc()
        result = {"error": {"message": "Error when running the model inference."}}

    if "device" in pipes[model_id]:
        try:
            pipe.to("cpu")
            torch.cuda.empty_cache()
        except:
            pipe.device = torch.device("cpu")
            pipe.model.to("cpu")
            torch.cuda.empty_cache()

    pipes[model_id]["using"] = False

    if result is None:
        result = {"error": {"message": "model not found"}}
    
    end = time.time()
    during = end - start
    print(f"[ complete {model_id} ] {during}s")
    print(f"[ result {model_id} ] {result}")

    return result