newsfilter.io
Tutorial, Lecture, Product Demonstration

Theano Tutorial (Pascal Lamblin, MILA)

  • The presentation roadmap includes a hands-on logistic regression example on the MNIST dataset, with potential additional demonstrations on ConfNet and LSTM for character-level text generation if time permits.
  • An IPython notebook containing code snippets from the slides will be provided via a companion GitHub repository for concurrent user execution.
  • TNO library version 0.9 is anticipated soon, introducing easier GPU array interaction from external Python code and support for float16 and integer data types.
  • Basic RNN operations are projected to be completed within the coming days, based on recent development focus.
  • Future development strategies target wrapping additional cuDNN operations for performance gains and implementing enhanced 3D convolution support.
  • Upcoming iterations aim to accelerate graph optimization and advance data parallelism capabilities.
  • The optimizer=none flag can be utilized during compilation to generate backtrace-inclusive error messages, though this necessitates recompiling the function.
  • Distribution of TNO-based models without Python or local compilers remains unsupported due to runtime dependencies on Python for memory management; Docker containerization is recommended as an alternative distribution method.