2. www.bennett.edu.in
Training Model Free of Cost!!!
• Google Colab (SaaS)
• Microsoft Azure Notebook (SaaS)
• Amazon Sagemaker (SaaS)
• Kaggle Kernel (SaaS)
• Google Cloud Platform (PaaS)
• Amazon Web Services (PaaS)
• Microsoft Azure (PaaS)
• IBM Cloud (PaaS) etc.
3. www.bennett.edu.in
Google Colab
• Cloud based free Jupyter notebook
• Mainly built for research collaboration
• Supports
– Python 2.7 and Python 3.6
– Tensorflow, Keras, PyTorch & OpenCV frameworks
• Configuration (Oct 2018)
– 1 Intel Xeon Processor @2.3 GHz, 13 GB RAM, 33 GB Disk Space
• Two types of hardware accelerators are provided:
– Tesla K80 GPU or Tensor Processing Units (TPUs)
• 12 hour limit on continuous usage
• All data is stored in Google Drive
4. www.bennett.edu.in
Steps To Train Model: Google
Colab
• Sign up using Google Account- colab.research.google.com
• Multiple options
– Upload existing Jupyter Notebook
– Code directly in cloud based notebook
– Run an existing example
– Use a notebook stored on Google Drive
– Use a notebook from GitHub
• Execute the code
• Hardware accelerators are not enabled by default
• Enable GPU or TPU (if required)
5. www.bennett.edu.in
Microsoft Azure Notebook
• Microsoft’s implementation of cloud version of Jupyter
Notebook
• Wider selection of programming languages
– Python, R, F# etc.
• Limitations (Oct 2018)
– 4GB RAM
– 1 GB data limit
– No GPU/TPU
• Pros
– No continuous usage limit
– All code and data is persisted
6. www.bennett.edu.in
Steps To Train Model: Microsoft
Azure Notebook
• Sign in using Microsoft Account – notebooks.azure.com
• Create unique user ID
• Create a new library
– Library Name
– Library ID
• Add Items to library
– New Notebook, folder, blank file
– Upload files from local computer
– Import from a given web location
• Execute the code
7. www.bennett.edu.in
Amazon SageMaker
• Amazon’s implementation of cloud based Jupyter Notebook
• Most flexible among all SaaS platforms (Oct 2018)
– Scale model to Petabytes Scale
– Lower end: 2 cores, 4 GB RAM, No GPU
– Higher end: 64 cores @ 2.3 GHz, 256 GB RAM, No GPU
– Highest end: 64 cores @ 2.3 GHZ (base), 488 GB RAM, 8 NVIDIA
Tesla V100 GPUs, 128 GB GPU memory
• Supported languages- Python & R
• Pros
– Flexibility and Scalability
– Integrates AWS other cloud tools like S3, EC2 etc.
• Cons - very limited free usage; expensive afterwards
8. www.bennett.edu.in
Steps To Train Model: Amazon
SageMaker
• Sign in using AWS Account – aws.amazon.com/sagemaker
– For initial signup credit card info is required
• Create new notebook instance
– Notebook Instance Name
– Instance Type
– IAM role - access control policy (Use default)
– VPC - virtual private cloud (Use default)
– Lifecycle configuration (Use default)
– Encryption key (Use default)
– Volume Size
• Open notebook instance
9. www.bennett.edu.in
Steps To Train Model: Amazon
SageMaker
• Multiple options
– Create new notebook
– Upload existing notebook and resources
– Use existing data stored in AWS S3
• Execute the code
• Lots of other features and pre-trained models
10. www.bennett.edu.in
Kaggle Kernel
• Kaggle - Largest community of data scientist & machine
learners
• Kaggle’s implementation of
– Jupyter Notebook
– Associated tools and machine learning frameworks
• Easily shareable on Kaggle website
• Languages - Python and R
• Hardware Accelerator - NVIDIA Tesla K80 (6 hours)
• Configuration (Oct 2018):
– Without GPU: 4 core CPU, 17 GB RAM, 5 GB Disk Space
– With GPU : 2 core CPU, 14 GB RAM, 1 NVIDIA Tesla K80 GPU
11. www.bennett.edu.in
Steps To Train Model: Kaggle
Kernel
• Sign in using Kaggle/Google/Facebook/Yahoo Account
https://www.kaggle.com/kernels
• Select New Kernel
• Select Notebook
• Multiple options:
– Code directly in web based notebook
– Upload Jupyter notebook
– Upload datasets and resources
• Accessing dataset in notebook – “../input/datasetName”
• Commit – executes your notebook from top to bottom
12. www.bennett.edu.in
Suggestions– When to use
what?
• Google Colab – Highly recommended for
– Education and research
– Free GPUs and TPUs
• Microsoft Azure Notebook – Recommended for
– using R and F# or
– Model needs to be trained for very long duration (>12 hours)
– Data persistence is required
• Amazon SageMaker – Recommended for
– Professional development and deployment
– Very large and computation intensive models
13. www.bennett.edu.in
Suggestions– When to use
what?
• Kaggle Kernel – Recommended for
– Kaggle users to showcase their work on kaggle website
– Free GPU
– Plethora of existing kernels for learning purpose