MLFlow火炬模型

Question

我有一个训练有素的Yolo模型，并且采用model.pt格式，我能够上传模型以创建一个mlflow中的工件。但是，当我查看yaml文件时，它列出了一些依赖项。我确信我装货的方式不对。

频道：

dependencies:

• python=3.6.13

• pip

• pip：**- mlflow
• scikit-learn==0.24.2
• cloudpickle==1.6.0**名称: mlflow-env

任何人，请让我知道如何使用预先训练过的模型将其推送到mlflow以创建工件，然后将依赖项(Docker)容器化以推送到AWS ECR。

Answer 1

底盘( https://www.chassis.ml )是一个开源项目，它可以做你想做的事情。

提供CHassis您的PyTorch文件。它将它包装在一个mlflow模型中，然后是一个grpc服务器，创建一个容器，然后为您将所有东西放到Docker中心。然后，您可以从码头枢纽拉，并推动到ECR自己。

PyTorch的例子有：https://github.com/modzy/chassis/tree/main/chassisml_sdk/examples/pytorch，虽然Yolo不在，但示例中包含了一个快捷笔记本，您可以根据您的需要进行修改。

您将需要访问底盘服务器。您可以按照底盘网站上的说明在本地设置一个(https://chassis.ml/getting-started/deploy-manual/)，或者通过在https://chassis.modzy.com注册使用公开托管的一个

基本代码如下：

#import modules
import chassisml
import pickle
import cv2
import torch
import getpass
import numpy as np
import torchvision.models as models
from torchvision import transforms

#provide docker crednetials
dockerhub_user = getpass.getpass('docker hub username')
dockerhub_pass = getpass.getpass('docker hub password')

#pull model and define pre / post processing of data
model = models.detection.fasterrcnn_mobilenet_v3_large_fpn(pretrained=True)
model.eval()

COCO_INSTANCE_CATEGORY_NAMES = [
    '__background__', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
    'train', 'truck', 'boat', 'traffic light', 'fire hydrant', 'N/A', 'stop sign',
    'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
    'elephant', 'bear', 'zebra', 'giraffe', 'N/A', 'backpack', 'umbrella', 'N/A', 'N/A',
    'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball',
    'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket',
    'bottle', 'N/A', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl',
    'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza',
    'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', 'N/A', 'dining table',
    'N/A', 'N/A', 'toilet', 'N/A', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
    'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'N/A', 'book',
    'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush'
]

transform = transforms.Compose([
    transforms.ToPILImage(),
    transforms.ToTensor(),
])       

device = 'cpu'

def preprocess(input_bytes):
    decoded = cv2.imdecode(np.frombuffer(input_bytes, np.uint8), -1)
    img_t = transform(decoded)
    batch_t = torch.unsqueeze(img_t, 0).to(device)
    return batch_t

def postprocess(num_detections, predictions):
    inference_result = {
        "detections": [
            {
                "xMin": predictions["boxes"][i].detach().cpu().numpy().tolist()[0],
                "xMax": predictions["boxes"][i].detach().cpu().numpy().tolist()[2],
                "yMin": predictions["boxes"][i].detach().cpu().numpy().tolist()[1],
                "yMax": predictions["boxes"][i].detach().cpu().numpy().tolist()[3],
                "class": labels[predictions["labels"][i].detach().cpu().item()],
                "classProbability": predictions["scores"][i].detach().cpu().item(),
            } for i in range(num_detections)
        ]
    }

    structured_output = {
        "data": {
            "result": inference_result,
            "explanation": None,
            "drift": None,
        }
    }    
    return structured_output

def process(input_bytes):
    
    # preprocess
    batch_t = preprocess(input_bytes)
    
    # run inference
    predictions = model(batch_t)[0]
    num_detections = len(predictions["boxes"])
    
    # postprocess
    structured_output = postprocess(num_detections, predictions)
    
    return structured_output

#create chassis client
chassis_client = chassisml.ChassisClient("<chassis_server_url>:<chassis service port>")

#convert pytorch model to mlflow model
# create Chassis model
chassis_model = chassis_client.create_model(process_fn=process)

# test Chassis model (can pass filepath, bufferedreader, bytes, or text here):
sample_filepath = './data/airplane.jpg'
results = chassis_model.test(sample_filepath)
print(results)

# have chassis containerize model
response = chassis_model.publish(
    model_name="PyTorch Faster R-CNN Object Detection",
    model_version="0.0.2",
    registry_user=dockerhub_user,
    registry_pass=dockerhub_pass
)

# wait for packaging to complete.
job_id = response.get('job_id')
final_status = chassis_client.block_until_complete(job_id)

Answer 2

在创建模型和了解依赖关系时，为什么不传递手动建模所需的库列表呢？

    conda_env ={ 
        "channels": ["conda-forge"],
        "dependencies" : [
             "python=<your-python-version>",
             "pip",
             {
                 "pip":[
                    "<your-pip-dependency>==<version>"
                 ],
             },
        ],
        "name": "mlflow-env"
    }

conda_env应该作为参数传递给您的日志模型。

    mlflow.pytorch.log_model(
        artifact_path="",
        conda_env= conda_env
        )