Lab 2 - Train your model

In this lab we are going the train a PyTorch Model that can classify Simpsons using the resources we have created in the previous lab.

1. Connect to your resources

Import dependencies

Start with importing dependencies. If you are using a Notebook in Azure Machine Learning Studio, you have all the latest versions install. If you are running your own Jupyter notebook then you have to install the azureml-sdk (pip install azureml-sdk).

• Paste the code below in the first cell and run this cell.

import os, random
import azureml
import shutil
import urllib.request
from pathlib import Path
import matplotlib.pyplot as plt
from mpl_toolkits.axes_grid1 import AxesGrid
import cv2
import urllib3 
import zipfile

from azureml.core.model import Model, InferenceConfig
from azureml.core import Workspace, Datastore, Experiment, Run, Environment, ScriptRunConfig

from azureml.core.compute import ComputeTarget, AmlCompute, AksCompute, ComputeTarget
from azureml.train.dnn import PyTorch
from azureml.widgets import RunDetails

from azureml.core.webservice import Webservice, AksWebservice, AciWebservice
from azureml.core.dataset import Dataset
from azureml.core.resource_configuration import ResourceConfiguration
from azureml.core.conda_dependencies import CondaDependencies 

# check core SDK version number
print("Azure ML SDK Version: ", azureml.core.VERSION)

Import dependencies

Connect to workspace

• Create a new code cell by clicking on the '+ Code' button

• Paste the code below

  ws = Workspace.from_config()
  print("Connected to workspace: ",ws.name)

• Run the cell

• Performing the interactive authentication using the link and code provide

• Run the cell again

• It should say: "Connected to workspace: "

Connect to Azure Machine Learning Compute Cluster

compute_target = ComputeTarget(workspace=ws, name="gpu-cluster")
print("Connected to compute target:",compute_target.name)

Connect to the default datastore

ds = Datastore.get_default(ws)
print("Connected to datastore:",ds.name)

Connect to resources

Create an experiment

exp = Experiment(workspace=ws, name='Simpsons-PyTorch')
print("Experiment created:",exp.name)

View your created experiment on: https://ml.azure.com

2. Data

Download the dataset from Github

data_url = "https://github.com/hnky/dataset-lego-figures/raw/master/_download/train-and-validate.zip"
data_path = "./data"
download_path = os.path.join(data_path,"train-and-validate.zip")
if not os.path.exists(data_path):
    os.mkdir(data_path,);
urllib.request.urlretrieve(data_url, filename=download_path)

Unzip the dataset

zip_ref = zipfile.ZipFile(download_path, 'r')
zip_ref.extractall(data_path)
zip_ref.close()
print("Data extracted in: {}".format(data_path))

os.remove(download_path)
print("Downloaded file removed: {}".format(download_path))

View your the downloaded dataset: https://ml.azure.com

Choose the refresh button above your folder structure to see data files

Dataset

Preview the dataset

To take a peak at the images in the dataset paste and the run the code below.

from mpl_toolkits.axes_grid1 import AxesGrid
import random
import cv2
import matplotlib.pyplot as plt

path = r"data/train"
random_filenames = []
for tag in os.listdir(path):
    random_filenames.append(path+"/"+tag+"/"+random.choice([
        x for x in os.listdir(os.path.join(path,tag))
        if os.path.isfile(os.path.join(path,tag, x))
    ]))

grid = AxesGrid(plt.figure(1, (20,20)), 111, nrows_ncols=(4, 5), axes_pad=0, label_mode="1")

i = 0
for img_name in random_filenames[0:10]:
    image = cv2.imread(img_name)
    image = cv2.resize(image, (352, 352))
    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

    # Show image in grid
    grid[i].imshow(image)
    i = i+1

Dataset

Upload the data to the datastore

ds.upload(src_dir=data_path, target_path='simpsonslego', overwrite=True, show_progress=True)

Create a dataset from the data in the datastore

datastore_paths = [(ds, 'simpsonslego/**')]
simpsons_ds = Dataset.File.from_files(path=datastore_paths)

Register the dataset

simpsons_ds.register(workspace=ws,
             name='LegoSimpsons',
             description='Simpsons dataset with Lego Figures',
             create_new_version = True)

Connect to the dataset

simpsons_ds = Dataset.get_by_name(ws, name='LegoSimpsons')

Train the model

Download the training script

project_folder = "./trainingscripts"

training_script_url = "https://raw.githubusercontent.com/hnky/DevelopersGuideToAI/master/amls/resources/train.py"

training_script_download_path = os.path.join(project_folder,"train.py")
if not os.path.exists(project_folder):
    os.mkdir(project_folder);
urllib.request.urlretrieve(training_script_url, filename=training_script_download_path)

Refresh your files and validate that 'train.py' is downloaded in the folder 'trainingscripts': https://ml.azure.com

Training Script

curated_env_name = 'AzureML-PyTorch-1.6-GPU'

pytorch_env = Environment.get(workspace=ws, name=curated_env_name)
pytorch_env = pytorch_env.clone(new_name='pytorch-1.6-gpu')

args = [
    '--data-folder', simpsons_ds.as_named_input('simpsons').as_mount(),
    '--num-epochs', 15
]

project_folder = "./trainingscripts"

config = ScriptRunConfig(
    source_directory = project_folder, 
    script = 'train.py', 
    compute_target=compute_target,
    environment = pytorch_env,
    arguments=args,
)

Submit the PyTorch estimator

run = exp.submit(config)

Follow the progress of the run

RunDetails(run).show()

• Click on 'View run details' to view all the details of the run under the experiment.

Training starting

This step can take up to 15 minutes to complete

Training done

Register the model

model = run.register_model(model_name='Simpsons-PyTorch',
                           model_path='outputs',
                           model_framework='PyTorch',
                           model_framework_version='1.3',
                           description="Simpsons PyTorch Classifier",
                           tags={'Conference':'Awesome AI Workshop'},
                           resource_configuration=ResourceConfiguration(cpu=1, memory_in_gb=2))

print("Model '{}' version {} registered ".format(model.name,model.version))

Test results

Validate that your model is visible under 'Models': https://ml.azure.com

Download and test your model

Download the model

model.download(exist_ok=True)

Download test images

test_images_url = "https://github.com/hnky/dataset-lego-figures/raw/master/_download/test-images.zip"
test_images_path = r"./data/test"
test_images_download_path = os.path.join(test_images_path,"test-images.zip")
if not os.path.exists(test_images_path):
    os.mkdir(test_images_path,);
urllib.request.urlretrieve(test_images_url, filename=test_images_download_path)

Unzip test images

zip_ref = zipfile.ZipFile(test_images_download_path, 'r')
zip_ref.extractall(test_images_path)
zip_ref.close()
print("Data extracted in: {}".format(test_images_path))

os.remove(test_images_download_path)
print("Downloaded file removed: {}".format(test_images_download_path))

Run the model over the test images

import os
import torch
import torch.nn as nn
import torchvision
from torchvision import transforms
import json
import urllib
from PIL import Image

# Load the model
loaded_model = torch.load(os.path.join('outputs','model.pth'), map_location=lambda storage, loc: storage)
loaded_model.eval()

# Load the labels
with open(os.path.join('outputs','labels.txt'), 'rt') as lf:
    global labels
    labels = [l.strip() for l in lf.readlines()]

def scoreImage(image_link):
    # Load the image to predict
    input_image = Image.open(image_link)

    # Pre process
    preprocess = transforms.Compose([
        transforms.Resize(225),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    input_tensor = preprocess(input_image)
    input_batch = input_tensor.unsqueeze(0)

    # Predict the image
    if torch.cuda.is_available():
        input_batch = input_batch.to('cuda')
        loaded_model.to('cuda')

    with torch.no_grad():
        output = loaded_model(input_batch)

    index = output.data.cpu().numpy().argmax()
    probability = torch.nn.functional.softmax(output[0], dim=0).data.cpu().numpy().max()

    #Return the result
    return {"label": labels[index], "probability": round(probability*100,2)}



path = r"data/test"
grid = AxesGrid(plt.figure(1, (20,20)), 111, nrows_ncols=(4, 5), axes_pad=0, label_mode="1")

i = 0
for img in os.listdir(path):
    
    #Score the image
    result = scoreImage(os.path.join(path,img))
    
    # Download image
    image = cv2.imread(os.path.join(path,img))
    image = cv2.resize(image, (352, 352))
    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    
    cv2.rectangle(image, (0,260),(352,352),(255,255,255), -1)
    cv2.putText(image, "{} - {}%".format(result['label'],result['probability']),(10, 300), cv2.FONT_HERSHEY_SIMPLEX, 0.65,(0,0,0),2,cv2.LINE_AA)    
    
    # Show image in grid
    grid[i].imshow(image)
    i = i+1

Test results

Continue with lab 3 >