Machine Learning

Tensor – Data structure in ML

Tensor – a way to store numbers in container

Points about Tensor:

1. 0-D Tensor is scalar
2. 1-D Tensor is vector in ML or 1-D array in programming
   • N-D tensor is collection of (N-1)D array/Tensor
   • Statement – “Vector is 1-D tensor but has its own dimension”
     • Example arr = [1,2,3] – its tensor dimension is 1 & vector dimension is 3 (since x,yz)
3. 2-D Tensor / matrix
   • R X C is a matrix or 2-D tensor
4. 3-D tensor/matrix
   • In addition of R and C, it has Depth (collection of 2-D matrix/2D tensor)
5. Note: In ML 0 to 5D are ones comes in pictturs
6. No. Of dimension Tensor = no. of axis = Rank
7. Size of tensor = no of item = R X C
8. Example of Tensor data
   1. 3D Tensor
      • Time Series data
   2. 4D tensor
      • Image
   3. 5-D tensor
      • Video
        • Domain knowledge – mp4, mpeg, mkv are the video encoder which makes raw format, large size video into small size video

Tools and Lib for ML

1. Anaconda- Contains packages for ML
2. Jupyter Notebook
3. Virtual env
   • Using conda you can create venv
4. Kaggle
   • Pros
     • Also has online version of Jupyternotebook
     • Has kaggle data
5. Google collab
   • Pros-
     • Also has online version of Jupyternotebook
       • It has GPU power Too – used for Deep learning
         • Runtime – > Change runtime – > select GPU
         • There is option to import kaggle data into collab – it needs few configs – like api key setup – which helps in download directly zip to collab which can be extracted with code and used

Image – Anaconda with base(root) having all libraries for data science & virtual env only required lib. so size remains manageable while deploying

E2E flow

Steps involved :

1. Preprocessing , EDA , Feature selection
   • panda lib useful
2. Extract input and output cols
   • panda lib useful
3. scale the values
4. Train & test data split
   • sklearn lib useful
5. model training
   • sklearn lib useful
6. Model evaluation & selection
   • sklearn lib useful
7. deploy
   • can be deployed on any cloud platform (aws/gcp etc)- model can be exported post selection using pickle library