# Pytorch & Fast API & Huggingface
#### 0. DETR에 대해서
- 논문 링크: https://arxiv.org/pdf/2005.12872.pdf
- 논문 해석(한글): https://www.youtube.com/watch?v=hCWUTvVrG7E
- Baseline:
- DETR R50: https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth
- DETR R101: https://dl.fbaipublicfiles.com/detr/detr-r101-2c7b67e5.pth
- 성능 비교

#### 1. Torch Hub
- Dependency

    pip install torch
    pip install scipy

- Load Pretrained State for transfer learning

    import torch
    state = torch.hub.load_state_dict_from_url('https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth', map_location='cpu', check_hash=True)
    model = custom_class.load_state_dict(state) # 직접 만든 모델 레이어에 state load

- Load Pretrained Model

    import torch
    model = torch.hub.load('facebookresearch/detr:main', 'detr_resnet50', pretrained=True)

#### 2. Hugging Face
##### 1. Code Lab에서 Pretrained 된 모델 로드
- Dependency

    !pip install transformers
    !pip install timm

- Load Pretrained Model

    from transformers import DetrFeatureExtractor, DetrForObjectDetection
    feature_extractor = DetrFeatureExtractor.from_pretrained('facebook/detr-resnet-50')
    model = DetrForObjectDetection.from_pretrained('facebook/detr-resnet-50')

- Load Sample Image

    from PIL import Image
    import requests

    url = 'https://drive.google.com/uc?export=download&id=1satnK-pt3PzOOKypkPd9MDUUXbkjs5ra'
    image = Image.open(requests.get(url, stream=True).raw)

- Feature Extraction

    inputs = feature_extractor(images=image, return_tensors="pt")
    outputs = model(**inputs)

    # model predicts bounding boxes and corresponding COCO classes
    logits = outputs.logits
    bboxes = outputs.pred_boxes

- Softmax

    import torch

    # keep only predictions of queries with 0.9+ confidence (excluding no-object class)
    probas = outputs.logits.softmax(-1)[0, :, :-1]
    keep = probas.max(-1).values > 0.9

    # rescale bounding boxes
    target_sizes = torch.tensor(image.size[::-1]).unsqueeze(0)

    postprocessed_outputs = feature_extractor.post_process(outputs, target_sizes)
    bboxes_scaled = postprocessed_outputs[0]['boxes'][keep]
    bboxes_scaled.shape[0] # Box 갯수

- Count Object

    def count_object(object_category):
      count = 0
      for p in probas[keep]:
        cl = p.argmax()
        if model.config.id2label[cl.item()] == object_category:
          count = count + 1
      print(count)

    count_object('car') # detect된 object 중 car 갯수

- Plot

    import matplotlib.pyplot as plt

    # colors for visualization
    COLORS = [[0.000, 0.447, 0.741], [0.850, 0.325, 0.098], [0.929, 0.694, 0.125],
              [0.494, 0.184, 0.556], [0.466, 0.674, 0.188], [0.301, 0.745, 0.933]]

    def plot_results(pil_img, prob, boxes):
        plt.figure(figsize=(16,10))
        plt.imshow(pil_img)
        ax = plt.gca()
        colors = COLORS * 100
        for p, (xmin, ymin, xmax, ymax), c in zip(prob, boxes.tolist(), colors):
            ax.add_patch(plt.Rectangle((xmin, ymin), xmax - xmin, ymax - ymin,
                                       fill=False, color=c, linewidth=3))
            cl = p.argmax()
            text = f'{model.config.id2label[cl.item()]}: {p[cl]:0.2f}'
            ax.text(xmin, ymin, text, fontsize=15,
                    bbox=dict(facecolor='yellow', alpha=0.5))
        plt.axis('off')
        plt.show()

    plot_results(image, probas[keep], bboxes_scaled)

일산 주차장1
일산 주차장2

##### 2 PyCharm Local 머신에서 Pretrained 된 모델 로드
- Dependency
- Transformer 기반 모델을 사용하기 위한 필수 dependency

    pip install transformers --file requirements.txt
    pip install timm --file requirements.txt

- 원하는 Torch 버전 선택 [PyTorch Site](https://pytorch.org/get-started/locally/)

    # 위에서 생성된 스크립트로 pytorch 설치
    conda install pytorch torchvision torchaudio -c pytorch --file requirements.txt

- Environment 셋팅
- 프로젝트 생성시 Interpreter 설정
- 위에서 생성한 conda python을 Interpreter로 적용

- Code
- Class 선언

    from transformers import DetrFeatureExtractor, DetrForObjectDetection
    from PIL import Image
    import requests
    import torch


    class CarCounterModel:
        def __init__(self):
            self.feature_extractor = DetrFeatureExtractor.from_pretrained('facebook/detr-resnet-50')
            self.model = DetrForObjectDetection.from_pretrained('facebook/detr-resnet-50')
            print('model loaded')

        @staticmethod
        def load_image(url):
            image = Image.open(requests.get(url=url, stream=True).raw)
            print('image size: ', image.size)
            return image

        def feature_extraction(self, image):
            inputs = self.feature_extractor(images=image, return_tensors="pt")
            return self.model(**inputs)

        def softmax(self, image, outputs):
            probas = outputs.logits.softmax(-1)[0, :, :-1]
            keep = probas.max(-1).values > 0.9

            target_sizes = torch.tensor(image.size[::-1]).unsqueeze(0)

            postprocessed_outputs = self.feature_extractor.post_process(outputs, target_sizes)
            bboxes_scaled = postprocessed_outputs[0]['boxes'][keep]
            print('total detected object count: ', bboxes_scaled.shape[0])
            return probas, keep, bboxes_scaled

        def count_object(self, probas_keep, category='car'):
            count = 0
            for p in probas_keep:
                cl = p.argmax()
                if self.model.config.id2label[cl.item()] == category:
                    count = count + 1
            print('total ', category, ' count: ', count)

- 실행

    model = CarCounterModel()
    url = 'https://drive.google.com/uc?export=download&id=1eC2ENVGBSyi04b_UAIryU6X8iTMdgzEh'
    loaded_image = model.load_image(url)
    features = model.feature_extraction(loaded_image)
    probas, keep, bboxes = model.softmax(loaded_image, features)
    model.count_object(probas[keep])

- 결과

    model loaded
    image size:  (4032, 3024)
    total detected object count:  4
    total  car  count:  4

#### 2. Fast API 적용
- Dependency

    pip install "fastapi[all] --file requirements.tst"
    conda install -c conda-forge config --file requirements.txt

- 실행

    uvicorn app:app --reload 

- Code

    from fastapi import FastAPI, File, UploadFile

    from interface.image_url_interface import ImageUrlInterface
    from model import CarCounterModel

    app = FastAPI()
    model = CarCounterModel()


    @app.get("/health")
    async def health_check():
        return {
            'status': 'ok'
        }


    @app.post("/url")
    async def detect_from_image_link(request_body: ImageUrlInterface):
        if len(request_body.url) == 0:
            return 'invalid url'
        loaded_image = model.load_image(request_body.url)
        features = model.feature_extraction(loaded_image)
        probas, keep, bboxes = model.softmax(loaded_image, features)
        count = model.count_object(probas[keep], 'car')
        return {
            'total object count': bboxes.shape[0],
            'car count': count
        }


    @app.post("/file")
    async def detect_from_image_file(file: UploadFile = File()):
        loaded_image = await model.load_image_file(file)
        features = model.feature_extraction(loaded_image)
        probas, keep, bboxes = model.softmax(loaded_image, features)
        count = model.count_object(probas[keep], 'car')
        return {
            'total object count': bboxes.shape[0],
            'car count': count
        }