Image Classification using AWS SageMaker

In this project AWS Sagemaker is used to finetune a pretrained model that can perform image classification by using the Sagemaker profiling, debugger, hyperparameter tuning and other ML engineering practices.

The model is ResNet-50 is used, on the dataset of Dog Breed Classification.

Project Set Up and Installation

Enter AWS through the gateway in the course and open SageMaker Studio. Download the starter files. Download/Make the dataset available.

The main files are:

1. train_and_deploy.ipynb: the notebook where we create, train, deploy our estimators and interact with SageMaker
2. hpo.py: the script for hyperparameter tuning
3. train_model.py: the script for training the model with the best hyperparameters and using the debugger and profiles

Make sure to use a kernel type that matches the framework requirements and install smdebug as instructed in the notebook

Dataset

The dataset is the dog breed classification which contains 133 different classes.

Access

Upload the data to an S3 bucket through the AWS Gateway so that SageMaker has access to the data.

Hyperparameter Tuning

What kind of model did you choose for this experiment and why? Give an overview of the types of parameters and their ranges used for the hyperparameter search

I used a residual network ResNet-50. ResNet architecture are known for overcoming vanishing gradients using skip connections. ResNet can achieve good results like VGG for example with being less complex.

The hyperparameters search included the following:

1. number of epochs: this is an important hyperparameter to determine the length of the training and reach optimal results without overfitting.
2. learning rate: the learning rate is important to reach our minima, we need a value that is is not so large that the model oscilates and misses the minimum point nor too small that the weight updates are small
3. batch size: optimally we want to fit our whole dataset and make updates based on it, but this is computationally not possible, therefore, we need to choose a batch size big enough and compatible with our resources

AWS Hyperparameter Tuning Jobs

hyperparameter tuning job

training jobs

best hyperparameters

training logs

Debugging and Profiling

In order to utilize SageMaker Debugger and profiler the following steps are needed:

1. Create rules and configs based on what we need.
2. Pass those configuration to our Estimator
3. Configure a hook in the training script After training is done we can fetch the data from S3.

Results

1. As seen in the profiler, the most triggered rule is the BatchSize which indicates that batchsize is small and I may need to increase to fully utilize my resources or choose a different instance.
2. LowGPUUtilization since I did not use the GPU
3. CPUBottleneck was above the set threshold so I might consider changing my data loaders.

Model Deployment

We can deploy the model using the SDK and specifiying instance count and type.

estimator_path = "s3://sagemaker-us-east-1-066783629629/Breed-classification-2023-01-05-16-01-59-814/output/model.tar.gz"
trained_model = PyTorch(
    model_data=estimator_path,
    entry_point="./infernce2.py",
    framework_version="1.8",
    py_version="py36",
    role=role)
my_estimator = trained_model.attach(training_job_name)

predictor=my_estimator.deploy(initial_instance_count=1, instance_type="ml.m5.xlarge")

Then we can get a data instance from S3 and pass it to our model to get the inference

response = predictor.predict(img_bytes, initial_args={"ContentType": "image/jpeg"})

the deployed end point in SageMaker