HISTORY.txt

.. _HISTORY:

=================
Old Release Notes
=================


Theano 0.4.1 (12 August 2011)
=============================

New features:

 * `R_op <http://deeplearning.net/software/theano/tutorial/gradients.html>`_ macro like theano.tensor.grad

   * Not all tests are done yet (TODO)
 * Added alias theano.tensor.bitwise_{and,or,xor,not}. They are the numpy names.
 * Updates returned by Scan (you need to pass them to the theano.function) are now a new Updates class.
   That allow more check and easier work with them. The Updates class is a subclass of dict
 * Scan can now work in a "do while" loop style.

   * We scan until a condition is met.
   * There is a minimum of 1 iteration(can't do "while do" style loop)
 * The "Interactive Debugger" (compute_test_value theano flags)

   * Now should work with all ops (even the one with only C code)
   * In the past some errors were caught and re-raised as unrelated errors (ShapeMismatch replaced with NotImplemented). We don't do that anymore.
 * The new Op.make_thunk function(introduced in 0.4.0) is now used by constant_folding and DebugMode
 * Added A_TENSOR_VARIABLE.astype() as a way to cast. NumPy allows this syntax.
 * New BLAS GER implementation.
 * Insert GEMV more frequently.
 * Added new ifelse(scalar condition, rval_if_true, rval_if_false) Op.

   * This is a subset of the elemwise switch (tensor condition, rval_if_true, rval_if_false).
   * With the new feature in the sandbox, only one of rval_if_true or rval_if_false will be evaluated.

Optimizations:

 * Subtensor has C code
 * {Inc,Set}Subtensor has C code
 * ScalarFromTensor has C code
 * dot(zeros,x) and dot(x,zeros)
 * IncSubtensor(x, zeros, idx) -> x
 * SetSubtensor(x, x[idx], idx) -> x (when x is a constant)
 * subtensor(alloc,...) -> alloc
 * Many new scan optimization 

   * Lower scan execution overhead with a Cython implementation
   * Removed scan double compilation (by using the new Op.make_thunk mechanism)
   * Certain computations from the inner graph are now Pushed out into the outer
     graph. This means they are not re-comptued at every step of scan.
   * Different scan ops get merged now into a single op (if possible), reducing
     the overhead and sharing computations between the two instances

GPU:

 * PyCUDA/CUDAMat/Gnumpy/Theano bridge and `documentation <http://deeplearning.net/software/theano/tutorial/gpu_data_convert.html>`_.

   * New function to easily convert pycuda GPUArray object to and from CudaNdarray object
   * Fixed a bug if you crated a view of a manually created CudaNdarray that are view of GPUArray.
 * Removed a warning when nvcc is not available and the user did not requested it.
 * renamed config option cuda.nvccflags -> nvcc.flags
 * Allow GpuSoftmax and GpuSoftmaxWithBias to work with bigger input.

Bugs fixed:

 * In one case an AdvancedSubtensor1 could be converted to a GpuAdvancedIncSubtensor1 insted of GpuAdvancedSubtensor1.
   It probably didn't happen due to the order of optimizations, but that order is not guaranteed to be the same on all computers.
 * Derivative of set_subtensor was wrong.
 * Derivative of Alloc was wrong.

Crash fixed:

 * On an unusual Python 2.4.4 on Windows
 * When using a C cache copied from another location
 * On Windows 32 bits when setting a complex64 to 0.
 * Compilation crash with CUDA 4
 * When wanting to copy the compilation cache from a computer to another

   * This can be useful for using Theano on a computer without a compiler.
 * GPU:

   * Compilation crash fixed under Ubuntu 11.04
   * Compilation crash fixed with CUDA 4.0

Know bug:

 * CAReduce with nan in inputs don't return the good output (`Ticket <http://trac-hg.assembla.com/theano/ticket/763>`_).

   * This is used in tensor.{max,mean,prod,sum} and in the grad of PermuteRowElements.
   * This is not a new bug, just a bug discovered since the last release that we didn't had time to fix.

Deprecation (will be removed in Theano 0.5, warning generated if you use them):

 * The string mode (accepted only by theano.function()) FAST_RUN_NOGC. Use Mode(linker='c|py_nogc') instead.
 * The string mode (accepted only by theano.function()) STABILIZE. Use Mode(optimizer='stabilize') instead.
 * scan interface change:

   * The use of `return_steps` for specifying how many entries of the output
     scan has been deprecated

     * The same thing can be done by applying a subtensor on the output
       return by scan to select a certain slice
   * The inner function (that scan receives) should return its outputs and
     updates following this order:

        [outputs], [updates], [condition]. One can skip any of the three if not
        used, but the order has to stay unchanged.
 * tensor.grad(cost, wrt) will return an object of the "same type" as wrt 
   (list/tuple/TensorVariable).

   * Currently tensor.grad return a type list when the wrt is a list/tuple of 
     more than 1 element.

Decrecated in 0.4.0(Reminder, warning generated if you use them):

 * Dividing integers with / is deprecated: use // for integer division, or
   cast one of the integers to a float type if you want a float result (you may
   also change this behavior with config.int_division).
 * tag.shape attribute deprecated (#633)
 * CudaNdarray_new_null is deprecated in favour of CudaNdarray_New

Sandbox:

 * MRG random generator now implements the same casting behavior as the regular random generator.

Sandbox New features(not enabled by default):

 * New Linkers (theano flags linker={vm,cvm})

   * The new linker allows lazy evaluation of the new ifelse op, meaning we compute only the true or false branch depending of the condition. This can speed up some types of computation.
   * Uses a new profiling system (that currently tracks less stuff)
   * The cvm is implemented in C, so it lowers Theano's overhead.
   * The vm is implemented in python. So it can help debugging in some cases.
   * In the future, the default will be the cvm.
 * Some new not yet well tested sparse ops: theano.sparse.sandbox.{SpSum, Diag, SquareDiagonal, ColScaleCSC, RowScaleCSC, Remove0, EnsureSortedIndices, ConvolutionIndices}

Documentation:

 * How to compute the `Jacobian, Hessian, Jacobian times a vector, Hessian times a vector <http://deeplearning.net/software/theano/tutorial/gradients.html>`_.
 * Slide for a 3 hours class with exercises that was done at the HPCS2011 Conference in Montreal.

Others:

 * Logger name renamed to be consistent.
 * Logger function simplified and made more consistent.
 * Fixed transformation of error by other not related error with the compute_test_value Theano flag.
 * Compilation cache enhancements.
 * Made compatible with NumPy 1.6 and SciPy 0.9
 * Fix tests when there was new dtype in NumPy that is not supported by Theano.
 * Fixed some tests when SciPy is not available.
 * Don't compile anything when Theano is imported. Compile support code when we compile the first C code.
 * Python 2.4 fix:

   * Fix the file theano/misc/check_blas.py
   * For python 2.4.4 on Windows, replaced float("inf") with numpy.inf.
 * Removes useless inputs to a scan node

   * Beautification mostly, making the graph more visible. Such inputs would appear as a consequence of other optimizations

Core:

 * there is a new mechanism that lets an Op permit that one of its
   inputs to be aliased to another destroyed input.  This will generally
   result in incorrect calculation, so it should be used with care!  The
   right way to use it is when the caller can guarantee that even if
   these two inputs look aliased, they actually will never overlap. This
   mechanism can be used, for example, by a new alternative approach to
   implementing Scan.  If an op has an attribute called
   "destroyhandler_tolerate_aliased" then this is what's going on.
   IncSubtensor is thus far the only Op to use this mechanism.Mechanism

Theano 0.4.0 (2011-06-13)
=========================

Change in output memory storage for Ops:
 If you implemented custom Ops, with either C or Python implementation,
 this will concern you.

 The contract for memory storage of Ops has been changed. In particular,
 it is no longer guaranteed that output memory buffers are either empty,
 or allocated by a previous execution of the same Op.

 Right now, here is the situation:
  * For Python implementation (perform), what is inside output_storage
    may have been allocated from outside the perform() function, for
    instance by another node (e.g., Scan) or the Mode. If that was the
    case, the memory can be assumed to be C-contiguous (for the moment).
  * For C implementations (c_code), nothing has changed yet.

 In a future version, the content of the output storage, both for Python and C
 versions, will either be NULL, or have the following guarantees:
  * It will be a Python object of the appropriate Type (for a Tensor variable,
    a numpy.ndarray, for a GPU variable, a CudaNdarray, for instance)
  * It will have the correct number of dimensions, and correct dtype
 However, its shape and memory layout (strides) will not be guaranteed.

 When that change is made, the config flag DebugMode.check_preallocated_output
 will help you find implementations that are not up-to-date.

Deprecation:
 * tag.shape attribute deprecated (#633)
 * CudaNdarray_new_null is deprecated in favour of CudaNdarray_New
 * Dividing integers with / is deprecated: use // for integer division, or
   cast one of the integers to a float type if you want a float result (you may
   also change this behavior with config.int_division).
 * Removed (already deprecated) sandbox/compile module
 * Removed (already deprecated) incsubtensor and setsubtensor functions,
   inc_subtensor and set_subtensor are to be used instead.

Bugs fixed:
 * In CudaNdarray.__{iadd,idiv}__, when it is not implemented, return the error.
 * THEANO_FLAGS='optimizer=None' now works as expected
 * Fixed memory leak in error handling on GPU-to-host copy
 * Fix relating specifically to Python 2.7 on Mac OS X
 * infer_shape can now handle Python longs
 * Trying to compute x % y with one or more arguments being complex now
   raises an error.
 * The output of random samples computed with uniform(..., dtype=...) is
   guaranteed to be of the specified dtype instead of potentially being of a
   higher-precision dtype.
 * The perform() method of DownsampleFactorMax did not give the right result
   when reusing output storage. This happen only if you use the Theano flags 
   'linker=c|py_nogc' or manually specify the mode to be 'c|py_nogc'.

Crash fixed:
 * Work around a bug in gcc 4.3.0 that make the compilation of 2d convolution
   crash.
 * Some optimizations crashed when the "ShapeOpt" optimization was disabled.

Optimization:
 * Optimize all subtensor followed by subtensor.

GPU:
 * Move to the gpu fused elemwise that have other dtype then float32 in them
   (except float64) if the input and output are float32.
   * This allow to move elemwise comparisons to the GPU if we cast it to
     float32 after that.
 * Implemented CudaNdarray.ndim to have the same interface in ndarray.
 * Fixed slowdown caused by multiple chained views on CudaNdarray objects
 * CudaNdarray_alloc_contiguous changed so as to never try to free
   memory on a view: new "base" property
 * Safer decref behaviour in CudaNdarray in case of failed allocations
 * New GPU implementation of tensor.basic.outer
 * Multinomial random variates now available on GPU

New features:
 * ProfileMode
    * profile the scan overhead
    * simple hook system to add profiler
    * reordered the output to be in the order of more general to more specific
 * DebugMode now checks Ops with different patterns of preallocated memory,
   configured by config.DebugMode.check_preallocated_output.
 * var[vector of index] now work, (grad work recursively, the direct grad
   work inplace, gpu work)
    * limitation: work only of the outer most dimensions.
 * New way to test the graph as we build it. Allow to easily find the source
   of shape mismatch error:
   `http://deeplearning.net/software/theano/tutorial/debug_faq.html#interactive-debugger`__
 * cuda.root inferred if nvcc is on the path, otherwise defaults to
   /usr/local/cuda
 * Better graph printing for graphs involving a scan subgraph
 * Casting behavior can be controlled through config.cast_policy,
   new (experimental) mode.
 * Smarter C module cache, avoiding erroneous usage of the wrong C
   implementation when some options change, and avoiding recompiling the
   same module multiple times in some situations.
 * The "theano-cache clear" command now clears the cache more thoroughly.
 * More extensive linear algebra ops (CPU only) that wrap scipy.linalg
   now available in the sandbox.
 * CUDA devices 4 - 16 should now be available if present.
 * infer_shape support for the View op, better infer_shape support in Scan
 * infer_shape supported in all case of subtensor
 * tensor.grad now gives an error by default when computing the gradient
   wrt a node that is disconnected from the cost (not in the graph, or
   no continuous path from that op to the cost).
 * New tensor.isnan and isinf functions.

Documentation:
 * Better commenting of cuda_ndarray.cu
 * Fixes in the scan documentation: add missing declarations/print statements
 * Better error message on failed __getitem__
 * Updated documentation on profile mode
 * Better documentation of testing on Windows
 * Better documentation of the 'run_individual_tests' script

Unit tests:
 * More strict float comparaison by default
 * Reuse test for subtensor of tensor for gpu tensor(more gpu test)
 * Tests that check for aliased function inputs and assure appropriate copying
   (#374)
 * Better test of copies in CudaNdarray
 * New tests relating to the new base pointer requirements
 * Better scripts to run tests individually or in batches
 * Some tests are now run whenever cuda is available and not just when it has
   been enabled before
 * Tests display less pointless warnings.

Other:
 * Correctly put the broadcast flag to True in the output var of
   a Reshape op when we receive an int 1 in the new shape.
 * pydotprint: high contrast mode is now the default, option to print
   more compact node names.
 * pydotprint: How trunk label that are too long.
 * More compact printing (ignore leading "Composite" in op names)


Theano 0.3.1 (2011-02-21)
=========================

Deprecation:
 * The theano shared variable attribute `value` is deprecated, use `get_value()` or `set_value()`!
    See http://deeplearning.net/software/theano/tutorial/aliasing.html

Bugs fixed:
 * The random number generator in theano/sandbox/rng_mrg.py did not always return the same sequence of number on the CPU and GPU.
    * In some cases, there was a (possibly large) fraction of non-random garbage in the returned sequence.

 * In python mode (not the default mode) when input of elemwise operation was an empty ndarray, we were not returning an empty ndarray.
 * Scan cached the number of steps. This caused no problem because each time you called scan the number of steps would got refreshed.
   The problem was when you called ScanGrad which would use the cached number of steps without refreshing it.
   To be affected by this bug, one would have to compile two graph, one that would contain a Scan and the other the corresponding GradScan, and
   call the first function to cache the number of steps, and then call the second function with a different number of steps.
 * In GpuConv, errors in conv_patch_stack_reduce when the entire kernel doesn't fit into shared memory.
   The error was not found before as the impact was less then the relative tolerance of 1e-3. Now the relative tolerance is 1e-5.

Crash fixed:
 * Add a feature to not have an exception that makes Theano crash when taking the gradient on DimShuffle in some particular case.
 * Compilation crash for GpuElemwise with tensor with high number of dimensions (~6 or more).
 * Disabled C code generator that make gcc crash on complex type.
 * Crash in optimization when an Op has no input.
 * Output shape is now computed correctly for matrix-vector multiplication on GPU.
 * In Scan, when using numbers as inputs, not symbolic variables.
 * In GradScan, when there is only 1 inputs in the Scan.
 * In GpuSum, bug in calculation of n_blocks for the 10 pattern. (Sum on the row of a matrix)
 * Some segfault at exit with GPU code.

Optimization:
 * New SpecifyShape op that allow to pass more shape info in the graph.
 * Speed up gemv by a work around scipy gemv slowness when the matrix is in C order (the default).
 * Remove join of only 1 element.
 * During optimization, consider one more case in get_constant_value.

GPU:
 * cuda_shared.value = X now works inplace!
     * cuda_shared_var.set_value(new_ndarray) will overwrite the old value inplace in the most common case.
 * Allow to create a CudaNdarraySharedVariable from a CudaNdarray.
 * New init_gpu_device theano flags.
 * Fuse GpuElemwise more often (in the case where there are so many inputs that fusing them all would bust the 256 bytes limit of parameter to gpu function).
 * CPU join of only 1 element that was not moved to the GPU.

New features:
 * tensor.reshape now makes dimensions of length 1 broadcastable.
 * tensor.prod now implements the gradient.
 * DebugMode now warns if an Op declared itself as returning a view of the input but did not do so.
    * This behaviour is a problem, because it can block other Ops from being inplace on the same inputs. This could lower the reuse of memory.
 * Sparse.structured_dot now works when both matrices are sparse
 * Sparse type is now supported by the shape op, and the ShapeFeature optimizer works correctly with them.
 * New 3D convolution ops, with CPU and GPU implementations.
 * New colors in pydotprint.

Documentation:
 * Documented lib.amdlibm and (new) init_gpu_device config variables.
 * A new page (was done for 0.3 but an error was hiding it on the web page) on the memory aliasing contract of Theano.
 * Revision to the Windows installation instructions.
 * The cuda documentation is now generated on the web server.
 * Better documentation of .theanorc and its sections.

Unit tests:
 * Stop usage of deprecated functions or syntax in the unit tests.
 * Better testing of GPU convolution nets.
 * Make more tests able to use different random seeds.
 * Tests of sparse now use default mode, not a hard-coded one.
 * Remove some tests of unimplemented features.

Other:
 * The name of compiledir now includes the Python version to make it easier for people with many Python versions
 * Added theano.tensor.std as a shortcut to sqrt(var(input=input, axis=axis)).
 * Whitespace, tabulation and indentation clean-up in the code.
 * Better detection of memory sharing between variables.


Theano 0.3 (2010-11-23)
=======================

This is the first major release of Theano since 0.1. Version 0.2 development started internally but it was never advertised as a release.

There have been so many changes since 0.1 that we have lost track of many of them. Below is a *partial* list of changes since 0.1.

 * GPU code using NVIDIA's CUDA framework is now generated for many Ops.
 * Some interface changes since 0.1:
     * A new "shared variable" system to allow reusing memory space between Theano functions.
         * A new memory contract has been formally written for Theano, for people who want to minimize memory copies.
     * The old module system has been deprecated.
     * By default, inputs to a Theano function will not be silently downcasted (e.g. from float64 to float32).
     * An error is now raised when using the result of logical operation on Theano variable in an 'if' (i.e. an implicit call to __nonzeros__).
     * An error is now raised when we receive a non-aligned ndarray as input to a function (this is not supported).
     * An error is raised when the list of dimensions passed to dimshuffle() contains duplicates or is otherwise not sensible.
     * Call NumPy BLAS bindings for gemv operations in addition to the already supported gemm.
     * If gcc is unavailable at import time, Theano now falls back to a Python-based emulation mode after raising a warning.
     * An error is now raised when tensor.grad is called on a non-scalar Theano variable (in the past we would implicitly do a sum on the tensor to make it a scalar).
     * Added support for "erf" and "erfc" functions.
 * The current default value of the parameter axis of theano.{max,min,argmax,argmin,max_and_argmax} is deprecated. We now use the default NumPy behavior of operating on the entire tensor.
 * Theano is now available from PyPI and installable through "easy_install" or "pip".


Theano 0.1
==========

*Release date: 2009-04-02*

What works
----------

- building symbolic expression.
- arranging symbolic expressions into Modules so that multiple functions
  can work on the same data.
- symbolic gradient descent.
- graph optimization.
- compilation to C for many kinds of expression.
- a debugging mode that checks that your expression results are correct,
  using a variety of sanity checks.

What's missing?
---------------

- An algorithm library. We're missing a library of examples and standard
  component implementations.  Some examples will find their way into
  the Theano repo, but standard algorithms will go into the 'pylearn'
  project (toolbox style). Now that we have a stable foundation, we
  can reach a consensus on style for algorithms.