[Azure] Create, deploy, and relearn a model [ML Studio classic]

I don't know the official document.

I've been modeling and deploying several times with ML Studio classic. This time, I read the official document because I wanted to relearn it, but I didn't understand it well and went back and forth. I've managed to do it, so I'll write an article so I don't forget it. Also, since I am writing for those who use it for the first time, I have omitted the details, so please forgive me.

Rough flow

  1. Make a learning experiment
  2. Make a predictive experiment
  3. Deploy predictive experiments
  4. Deploy re-learning
  5. Relearn
  6. Reflect the re-learning result in the prediction experiment
  7. Test if it is reflected

is.

Operating environment

Since Python will be used on the way, write the environment and the imported one.

name
urllib
azure-storage-blob

About the data to be used

Let's start with the data we will use this time. I want to know if I was able to relearn, so I will use data that is too monotonous. Write the training data, test data, and remodeling data below.

train1.csv


id,target
1,1
2,1
3,1
4,1
5,1

remodel0.csv


id,target
1,0
2,0
3,0
4,0
5,0

In the created model, expect a model that always returns 1 no matter how many you enter. So even if you put in the test data, it should all be returned as 1. We expect 0 to be returned by learning that everything is 0.

It is not a good model at all, so I will use it only for testing.

Create a learning experiment

Now let's create a model. After logging in to ML Studio, select ʻEXPERIMENTSand clickNEW` at the bottom of the screen. Web services - Microsoft Azure Machine Learning Studio (classic) - Google Chrome 2020_01_21 17_12_54.png

Then click Blank Experiment. Now you are ready to create a predictive model.

Let's actually make a model. Simply search for the required block from this search window and add it. キャプチャ.PNG

This time it is a binary term (for the time being), so place the box created by searching with two classes. This time, I used Two-Class Boosted Decision Tree. Search for other blocks in the same way and arrange them as shown in the image. キャプチャ.PNG

Next, set each block. Click on the block Setting items will appear on the right side of the page. Parameters can be set in the algorithm block. This time you can leave it as it is.

Next is import Data. Here, specify the data to be used for learning. Write your blob username, Key, and file path. The format of this data is csv, but since the file has a header, let's check File has header row.

Finally, the Train Model. Here, we will specify the column name of the target to be learned. This time I want to learn target, so I will fill in target with Launch column selector. キャプチャ.PNG

When you're done so far, let's connect the blocks with a line. Please note that the Train Model cannot be connected by lines on the left and right sides. If you can arrange it as below, click RUN at the bottom of the screen to execute it. キャプチャ.PNG

When all the boxes are checked, you're done. If you stop in the middle, there must be something wrong. The blob file name may be different (experience story)

If all are checked, the model is complete!

Make a predictive experiment

Now, let's make something that will return an answer when you throw data using the model you created earlier. Click Predictiv web Service from SET UP WEB SERVICE next to RUN. キャプチャ.PNG

Then the box will move and look like the one below. Bring ʻExport Data` from the search window and write the blob user name etc. The test result will be output to the path written here.

キャプチャ.PNG

Let's RUN again. キャプチャ.PNG

If it is checked as before, it is complete. Click DEPLOY WEB SERVICE when you are done. キャプチャ.PNG After a while, the screen will change and API KEY etc. will be displayed. Clicking on REQUEST / RESPONSE will bring up the API, which you can expect to use. You can also do a simple test on this page. Let's experiment by clicking the blue TEST button and putting 3 in id and 1 in target. After a while, the result will be returned at the bottom of the page. The column name, column type, and value are returned in a list.

return


Result: {"Results":{"output1":{"type":"table","value":{"ColumnNames":["id","target","Scored Labels","Scored Probabilities"],"ColumnTypes":["Int32","Int32","Int32","Double"],"Values":[["3","1","1","0.142857149243355"]]}}}}

Since Scored Labels is the result, 1 is returned in this case, which is as expected.

I'll deploy re-learning

Now click View latest to return to the previous screen. キャプチャ.PNG

Then go to the Training expepriment tab and add Web service input and Web service output to RUN. キャプチャ.PNG

Then click Deploy Web Service from SET UP WEB SERVICE below. キャプチャ.PNG

Deployment is now complete. You will need the API Key and API URL later. The API Key is what you see on the screen. The URL will appear when you click BATCH EXECTION. The URL of the API to be used is before ? Api-version…, that is, the URL ending with job. I stumbled here. キャプチャ.PNG

Let's relearn!

Let's relearn. The re-learning itself was done in python. C # is also a sample, but it may not work unless you change it in some places. (That was the case with python3.7.)

First, the program you actually use.

retrain.py


# python 3.7 so change urllib2
import urllib
import urllib.request
import json
import time
from azure.storage.blob import *

def printHttpError(httpError):
    print(f"The request failed with status code: {str(httpError.code)}")
    print(json.loads(httpError.read()))
    return


def processResults(result):
    results = result["Results"]
    for outputName in results:
        result_blob_location = results[outputName]
        sas_token = result_blob_location["SasBlobToken"]
        base_url = result_blob_location["BaseLocation"]
        relative_url = result_blob_location["RelativeLocation"]
        print(f"The results for {outputName} are available at the following Azure Storage location:")
        print(f"BaseLocation: {base_url}")
        print(f"RelativeLocation: {relative_url}")
        print(f"SasBlobToken: {sas_token}")
    return



def uploadFileToBlob(input_file, input_blob_name, storage_container_name, storage_account_name, storage_account_key):
    #It seems that there is no BlobService, so change it
    blob_service = BlockBlobService(account_name=storage_account_name, account_key=storage_account_key)
    print("Uploading the input to blob storage...")
    blob_service.create_blob_from_path(storage_container_name, input_blob_name, input_file)

def invokeBatchExecutionService():

    storage_account_name = "blob username"
    storage_account_key = "blob key"
    storage_container_name = "blob container name"
    connection_string = f"DefaultEndpointsProtocol=https;AccountName={storage_account_name};AccountKey={storage_account_key}"
    api_key = "Re-learning API Key"
    url = "API URL"



    uploadFileToBlob("File path to upload",
                     "File path after upload", 
                     storage_container_name, storage_account_name, storage_account_key)

    payload =  {
        "Inputs": {
            "input1": { 
                "ConnectionString": connection_string, 
                "RelativeLocation": f"/{storage_container_name}/File path of blob to be remodeled" 
            },
        },
        "Outputs": {
            "output1": { 
                "ConnectionString": connection_string, 
                "RelativeLocation": f"/{storage_container_name}/Remodeled blob file path.ilearner" 
            },
        },
        "GlobalParameters": {
        }
    }

    body = str.encode(json.dumps(payload))
    headers = { "Content-Type":"application/json", "Authorization":("Bearer " + api_key)}
    print("Submitting the job...")

    req = urllib.request.Request(url + "?api-version=2.0", body, headers)
    response = urllib.request.urlopen(req)

    result = response.read()
    job_id = result[1:-1] 
    # job_I was angry because id was not str, so I converted it
    job_id=job_id.decode('utf-8')
    print(f"Job ID: {job_id}")

    print("Starting the job...")    
    headers = {"Authorization":("Bearer " + api_key)}
    req = urllib.request.Request(f"{url}/{job_id}/start?api-version=2.0", headers=headers, method="POST")
    response = urllib.request.urlopen(req)

    url2 = url + "/" + job_id + "?api-version=2.0"

    while True:
        print("Checking the job status...")
        req = urllib.request.Request(url2, headers = { "Authorization":("Bearer " + api_key) })

        response = urllib.request.urlopen(req)

        result = json.loads(response.read())
        status = result["StatusCode"]
        if (status == 0 or status == "NotStarted"):
            print(f"Job: {job_id} not yet started...")
        elif (status == 1 or status == "Running"):
            print(f"Job: {job_id} running...")
        elif (status == 2 or status == "Failed"):
            print(f"Job: {job_id} failed!")
            print("Error details: " + result["Details"])
            break
        elif (status == 3 or status == "Cancelled"):
            print(f"Job: {job_id} cancelled!")
            break
        elif (status == 4 or status == "Finished"):
            print(f"Job: {job_id} finished!")
        
            processResults(result)
            break
        time.sleep(1) # wait one second
    return

invokeBatchExecutionService()

It's almost the same as the sample, but some changes have been made. (I wrote it in the comments.) Please rewrite the URL, Key, PATH, etc. according to each environment. Then prepare the data and execute it.

console


>python remodel.py

Uploading the input to blob storage...
Submitting the job...
Job ID: ID
Starting the job...
Checking the job status...
JobID not yet started...
Checking the job status...
JobID running...
Checking the job status...
JobID running...
Checking the job status...
JobID running...
Checking the job status...
JobID running...
Checking the job status...
JobID running...
Checking the job status...
JobID running...
Checking the job status...
JobID running...
Checking the job status...
JobID finished!
The results for output1 are available at the following Azure Storage location:
BaseLocation: URL
RelativeLocation: PATH
SasBlobToken: KEY

You should get the result as above. The following three will be used later.

Reflection of re-learning

Well, I executed the remodeling, but it has not been reflected yet. Open New Web Services Experience from this screen. At this time, if [predictive exp.] Is not added after the model name at the top of the page, click ʻExperiment created on… [Predictive Exp.]` To move it.

キャプチャ.PNG

キャプチャ.PNG

I think it's okay to overwrite the existing one, but most of the time I'll keep it. So create a new endpoint. Press the left button from this screen ... キャプチャ.PNG (If you don't mind that the arrows are messy, you lose.) Click + NEW and save the endpoint with the name you want to use. キャプチャ.PNG

Clicking on the created endpoint name will take you to a screen like the one above, so open the Consume tab. キャプチャ.PNG

When you open it, various KEYs and URLs will appear. This time we will use Primary Key and` Patch. ``

Use these to execute the following code.

update.py


import urllib
import urllib.request
import json 

data =  {
            "Resources": [
                {
                    "Name": "Model name",
                    "Location": 
			            {
                            "BaseLocation": "Result URL",
                            "RelativeLocation": "Result PATH",
                            "SasBlobToken": "Result KEY"
			            }
                }
            ]
        }

body = str.encode(json.dumps(data))

url = "Patch value"
api_key = "Primary Key value" # Replace this with the API key for the web service
headers = {'Content-Type':'application/json', 'Authorization':('Bearer '+ api_key)}

req = urllib.request.Request(url, body, headers)
req.get_method = lambda: 'PATCH'
response = urllib.request.urlopen(req)
result = response.read()
print(result) 

It is like this. If you open API help under Patch on the previous page, there is a sample, so I think that you should write it as it is. Let's run it.

b'' is returned. I see? But this seems to be fine. Finally deploy this. The deployment procedure is

Did you relearn?

Let's test by opening the Test tab at the top of the page. Maybe you need to reload. id = 3, target = 1. キャプチャ.PNG It seems that it is properly set to 0. Was good…

Summary

I managed to deploy the retrained model. Is the accuracy of the model you are using decreasing? If you think so, relearn and continue using the model!

Recommended Posts

[Azure] Create, deploy, and relearn a model [ML Studio classic]
Create and deploy a Django (PTVS) app using Azure Table storage
[Ubuntu] Install Android Studio and create a shortcut
[Azure] Deploy locally created ML model to the cloud
Create a 3D model viewer with PyQt5 and PyQtGraph
Create a deploy script with fabric and cuisine and reuse it
Create a model for your Django schedule
Create and deploy Flask apps in PTVS
Implement a model with state and behavior
Create a REST API using the model learned in Lobe and TensorFlow Serving.