Tuesday, September 29, 2020

Day 66-100 Playing with the Face Recognition

For my Ongoing R&D Internship have to work on a Project involving facial recognition  and past 2 months had been most of research and making a working proof of concept of the project , can't go into much of detail due to stack of paper i signed but the journey into working with ML in the industry is  still new and lack of tried and tested system design and tech stack combination like we have in web development ,now  are plenty of face recongtion tutorial and videos that can explain the topic better than i can , but here are the key takeaways and practicse that will made it easier to navigate my way going deeper and deeper .

  • Having an Understanding of the OS and the Computer Architecture and Organisation and various hardware inside you own Machine and in your phone does help  , to estimate what is the capability of the hardware and estimating the time required to complete the task be it estimating the time taken to complet traning your model  to  time in millisecond that will take for an input to be processed in the model and what all could be the bottle neck in the system when dealing with the edge computing 
    • CPU-Central Procesing Unt
    • DPU-Deep Learning Procesing Unt
    • FPGA-Field Programable Gate Aray
    • GPU-Graphics Procesing Unt
    • IPU-Intelligence/Image Procesing Unt
    • TPU-Tensor Processing Unt
    • NPU-Neural Network Procesing Unt
    • SPU-Stream Procesing Unt
    • VPU-Vision Procesing Unt 

  • " Standing on the Shoulder of the Giants " , looking into the work of other companies who have done similar projcts and the challenges they faced in their enginnering blog , or even in their enginners own blogs does help into giving new insights and possible tried and tested approaches and possible thing that could go wrong over time 

  • When Exploring a completey new topic and field which you have no idea about and have lot of big words to deal with  , it used to be easier to chart it out on white board and draw connected graphs , pseudo code etc which we all miss a lot in work from scenario  so either keep a pen paper to draw out or if you are like me whos diagrams are really big and isnt good on drawing it out look into tools listed below that helps create a clearer mental map when solving thing and also saves time when you need diagrams to present to you team and maybe sometimes one person drawing might trigger someone else thought process in right direction

  • An an ML enginner it will always be dream to have the highest spec , machine do everything locally but , sooner or later get used to setting up and learning to things on the cloud even if deployment isnt in your to do list  thinking and designing system in a way that will be easier to deploy and  practical and effecient and keeps the planet green  (Nvidia or Plants :P ?)
  • Just like we have git for code look into versioning system for data , DVC  
  • Lastly , have look into how same thing is being implemented in countries like China , Japan , Korea , Israel since many of their articles not make it upto to of search result but the results and solution are really good [you know you have been looking really hard into a problem when you end up into a  site or forum for that solution  in mandarin ]
Up next NLP task at work with hard deadline of 23 Nov , so less of looking into images  and more of NLP , Happy Learning

PS : 100 days of code ended back and its been lot many days since then with work and side projects its coding everyday so guess it will #MLjourneyfor now 

Thursday, August 06, 2020

Days of Machine Learning Code Day 28-61



Update : Took a lot of time , understanding to complete the Deep Learning course on Coursera as the topics needed better understanding being CNN , GRU, LSTM , especially the CNN part did send me off to a side quest of learning more about it and some other image recongition works (as background building  part of ongoing office project)  and the last course on Sequence model was tough  and needs more hands on to get a better grip and implementation , the guided projects along the course duration were more of sanboxed and easy to play with but if to jump to some real implemented project with deployment and access on the web  would need to go through tensorflow to start using and deploy  some hellow world equivalent examples in the cloud accessible through web , the main aim is to use these tech to solve a real world problem and not to just get certificates and use ready to eat datasets , So with armed with knowledge imparted by AndrewNG lets use it for making things