Vahid Kazemi

@VahidK

PhD in computer vision & robotics. Engineering manager at . Ex. , , and .

Santa Monica, CA
Vrijeme pridruživanja: travanj 2008.

Tweetovi

Blokirali ste korisnika/cu @VahidK

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @VahidK

  1. 29. stu 2019.

    A practical lesson I learned from doing research in deep learning is to spend considerable amount of time at the beginning of the process on optimizing data loading and common operations making sure 100% of my GPU resources are utilized. It pays off massively in the long run.

    Poništi
  2. 25. stu 2019.

    I made a small package which allows reading tfrecord files in PyTorch with no tf dependency:

    Poništi
  3. 7. stu 2019.

    One way I made myself more productive is to write down everything I want to achieve in the foreseeable future. More recently I started also logging what I actually did and unexpected problems that I had to solve. This has helped me better plan and save time.

    Poništi
  4. 15. lis 2019.

    One major feature I miss from TensorFlow in PyTorch is GPU-based preprocessing operations through tf.image and tf.contrib/tf.contrib.image. Pre-processing in TF can be done very efficiently on GPU. DALI is ok, but doesn't provide the same level of flexibility.

    Poništi
  5. 11. lip 2019.

    I was working on optimizing some Pytorch code today and was amazed how fast Pytorch ran a pretty non-optimal code. So I made some test cases to compare with TensorFlow. Pytorch handily beat TensorFlow running vectorized and non-vectorized code in my test cases. Here's one:

    Poništi
  6. 25. svi 2019.

    Python is so inefficient, Python coders think twice before implementing any new algorithm; they prefer a ready made library (usually written in C++). Paradoxically this has made Python coders much more productive. C++ coders are still writing their own string classes in 2019.

    Poništi
  7. 25. svi 2019.

    We relabeled our test set with more accurate labels and found many of the models we dropped before were actually working better than the models we thought were best. Makes me wonder what percentage of published results are plain wrong because of systematic errors in evaluation.

    Poništi
  8. 22. tra 2019.

    "consider the problem is a self-driving car. In this case there is a very long tail of traffic situations that are very rare and therefore do not show up in your dataset. In this case a purely data-driven method that does not try to model the world is doomed."

    Poništi
  9. 13. tra 2019.

    Since I left Alphabet, I came to realize that computational resources can be limited! Gone are the days of using hundreds of TPUs without anyone raising an eyebrow. Now I spend a lot of time optimizing neural nets to train and run faster. The experience has been very rewarding.

    Poništi
  10. 13. tra 2019.

    TensorFlow tip: Always use _and_batch instead of separate map and batch. I got a 1.7x improvement in reading throughput just by doing that. Another important thing is to make as many of your preprocessing ops as possible in batch and move to GPU.

    Prikaži ovu nit
    Poništi
  11. 7. tra 2019.

    I think if math was taught like this in high schools we would have many more scientists in the world. What an amazing work.

    Poništi
  12. 31. ožu 2019.

    This is a neat tool to learn new languages.

    Poništi
  13. 16. ožu 2019.

    While TensorFlow 2 is more approachable and looks just like Python, it can have surprising behaviors at times if you aren't familiar with its symbolic API. One example with tf.while_loop and variable shape tensors was discussed in the guide.

    Prikaži ovu nit
    Poništi
  14. 16. ožu 2019.

    I just started working on Effective TensorFlow 2.0 today. Ported six of the items: . More to come.

    Prikaži ovu nit
    Poništi
  15. 11. ožu 2019.

    I believe we will build AGI at some point, and by that I mean the kind of ML model that can emulate human brain, but we are no where near that. Pitching AGI to investors now, is like pitching "Uber on Pluto, for aliens". Is it theoretically possible? Yes. Is it feasible? no.

    Poništi
  16. 6. ožu 2019.

    Swift for TensorFlow is exciting. You get a modern and efficient programming language + built-in language support for automatic differentiation + library support for massively parallel computation on GPUs/TPUs. I'm looking forward to say goodbye to Python soon!

    Poništi
  17. 4. ožu 2019.

    My paper on real-time face landmark estimation (used by Snapchat and several other companies) just passed 1000 citations according to Google scholar. Quite a milestone!

    Poništi
  18. 19. velj 2019.

    I wonder how many millions of hours of engineering time would have been saved, if C++ had a built-in standard linear algebra library like Eigen.

    Poništi
  19. 6. velj 2019.

    IMO one of the most poorly designed parts of TensorFlow is tf.estimator API. I have wasted hours and days trying to hack tf.estimator to do what I want, in the hindsight I should have avoided it altogether in the first place. I hope to see a functional high-level API in tf 2.0.

    Poništi
  20. 30. sij 2019.

    Big gains in ML projects are often obtained by understanding the data not randomly trying out things. Visualize predictions/loss on both training/test data. Find/fix problems. Bonus tip: triple check your preprocessing code.

    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·