Theano 0.5.0

Optimizing compiler for evaluating mathematical expressions on CPUs and GPUs.

Theano is a Python library that allows you to define, optimize, and efficiently evaluate mathematical expressions involving multi-dimensional arrays. It is built on top of NumPy_. Theano features:

* **tight integration with NumPy:** a similar interface to NumPy's. numpy.ndarrays are also used internally in Theano-compiled functions.
* **transparent use of a GPU:** perform data-intensive computations up to 140x faster than on a CPU (support for float32 only).
* **efficient symbolic differentiation:** Theano can compute derivatives for functions of one or many inputs.
* **speed and stability optimizations:** avoid nasty bugs when computing expressions such as log(1 + exp(x)) for large values of x.
* **dynamic C code generation:** evaluate expressions faster.
* **extensive unit-testing and self-verification:** includes tools for detecting and diagnosing bugs and/or potential problems.

Theano has been powering large-scale computationally intensive scientific
research since 2007, but it is also approachable enough to be used in the
classroom (IFT6266 at the University of Montreal).

.. _NumPy: http://numpy.scipy.org/


.. _NEWS:

=============
Release Notes
=============

Theano 0.5 (23 February 2012)
=============================

Highlight:
* Moved to github: http://github.com/Theano/Theano/
* Old trac ticket moved to assembla ticket: http://www.assembla.com/spaces/theano/tickets
* Theano vision: http://deeplearning.net/software/theano/introduction.html#theano-vision (Many people)
* Theano with GPU works in some cases on Windows now. Still experimental. (Sebastian Urban)
* Faster dot() call: New/Better direct call to cpu and gpu ger, gemv, gemm
and dot(vector, vector). (James, Frederic, Pascal)
* C implementation of Alloc. (James, Pascal)
* theano.grad() now also work with sparse variable. (Arnaud)
* Macro to implement the Jacobian/Hessian with theano.tensor.{jacobian,hessian} (Razvan)
* See the Interface changes.


Interface Behavior Changes:
* The current default value of the parameter axis of
theano.{max,min,argmax,argmin,max_and_argmax} is now the same as
numpy: None. i.e. operate on all dimensions of the tensor.
(Frederic Bastien, Olivier Delalleau) (was deprecated and generated
a warning since Theano 0.3 released Nov. 23rd, 2010)
* The current output dtype of sum with input dtype [u]int* is now always [u]int64.
You can specify the output dtype with a new dtype parameter to sum.
The output dtype is the one using for the summation.
There is no warning in previous Theano version about this.
The consequence is that the sum is done in a dtype with more precision than before.
So the sum could be slower, but will be more resistent to overflow.
This new behavior is the same as numpy. (Olivier, Pascal)
* When using a GPU, detect faulty nvidia drivers. This was detected
when running Theano tests. Now this is always tested. Faulty
drivers results in in wrong results for reduce operations. (Frederic B.)


Interface Features Removed (most were deprecated):
* The string modes FAST_RUN_NOGC and STABILIZE are not accepted. They
were accepted only by theano.function().
Use Mode(linker='c|py_nogc') or Mode(optimizer='stabilize') instead.
* tensor.grad(cost, wrt) now always returns an object of the "same type" as wrt
(list/tuple/TensorVariable). (Ian Goodfellow, Olivier)
* A few tag.shape and Join.vec_length left have been removed. (Frederic)
* The .value attribute of shared variables is removed, use shared.set_value()
or shared.get_value() instead. (Frederic)
* Theano config option "home" is not used anymore as it was redundant with "base_compiledir".
If you use it, Theano will now raise an error. (Olivier D.)
* scan interface changes: (Razvan Pascanu)
* The use of `return_steps` for specifying how many entries of the output
to return has been removed. Instead, apply a subtensor to the output
returned 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.

Interface bug fix:
* Rop in some case should have returned a list of one Theano variable,
but returned the variable itself. (Razvan)

New deprecation (will be removed in Theano 0.6, warning generated if you use them):
* tensor.shared() renamed to tensor._shared(). You probably want to
call theano.shared() instead! (Olivier D.)


Bug fixes (incorrect results):
* On CPU, if the convolution had received explicit shape information,
they where not checked at runtime. This caused wrong result if the
input shape was not the one expected. (Frederic, reported by Sander
Dieleman)
* Theoretical bug: in some case we could have GPUSum return bad value.
We were not able to reproduce this problem
* patterns affected ({0,1}*nb dim, 0 no reduction on this dim, 1 reduction on this dim):
01, 011, 0111, 010, 10, 001, 0011, 0101 (Frederic)
* div by zero in verify_grad. This hid a bug in the grad of Images2Neibs. (James)
* theano.sandbox.neighbors.Images2Neibs grad was returning a wrong value.
The grad is now disabled and returns an error. (Frederic)
* An expression of the form "1 / (exp(x) +- constant)" was systematically matched to "1 / (exp(x) + 1)"
and turned into a sigmoid regardless of the value of the constant. A warning will be issued if your
code was affected by this bug. (Olivier, reported by Sander Dieleman)
* When indexing into a subtensor of negative stride (for instance, x[a:b:-1][c]),
an optimization replacing it with a direct indexing (x[d]) used an incorrect formula,
leading to incorrect results. (Pascal, reported by Razvan)
* The tile() function is now stricter in what it accepts to allow for better
error-checking/avoiding nonsensical situations. The gradient has been
disabled for the time being as it only implemented (incorrectly) one special
case. The `reps` argument must be a constant (not a tensor variable), and
must have the same length as the number of dimensions in the `x` argument;
this is now checked. (David)


Scan fixes:
* computing grad of a function of grad of scan (reported by Justin Bayer, fix by Razvan)
before : most of the time crash, but could be wrong value with bad number of dimensions (so a visible bug)
now : do the right thing.
* gradient with respect to outputs using multiple taps (reported by Timothy, fix by Razvan)
before : it used to return wrong values
now : do the right thing.
Note: The reported case of this bug was happening in conjunction with the
save optimization of scan that give run time errors. So if you didn't
manually disable the same memory optimization (number in the list4),
you are fine if you didn't manually request multiple taps.
* Rop of gradient of scan (reported by Timothy and Justin Bayer, fix by Razvan)
before : compilation error when computing R-op
now : do the right thing.
* save memory optimization of scan (reported by Timothy and Nicolas BL, fix by Razvan)
before : for certain corner cases used to result in a runtime shape error
now : do the right thing.
* Scan grad when the input of scan has sequences of different lengths. (Razvan, reported by Michael Forbes)
* Scan.infer_shape now works correctly when working with a condition for the number of loops.
In the past, it returned n_steps as the length, which is not always true. (Razvan)
* Scan.infer_shape crash fix. (Razvan)

New features:
* AdvancedIncSubtensor grad defined and tested (Justin Bayer)
* Adding 1D advanced indexing support to inc_subtensor and set_subtensor (James Bergstra)
* tensor.{zeros,ones}_like now support the dtype param as numpy (Frederic)
* Added configuration flag "exception_verbosity" to control the verbosity of exceptions (Ian)
* theano-cache list: list the content of the theano cache (Frederic)
* theano-cache unlock: remove the Theano lock (Olivier)
* tensor.ceil_int_div to compute ceil(a / float(b)) (Frederic)
* MaxAndArgMax.grad now works with any axis (The op supports only 1 axis) (Frederic)
* used by tensor.{max,min,max_and_argmax}
* tensor.{all,any} (Razvan)
* tensor.roll as numpy: (Matthew Rocklin, David Warde-Farley)
* Theano with GPU works in some cases on Windows now. Still experimental. (Sebastian Urban)
* IfElse now allows to have a list/tuple as the result of the if/else branches.
* They must have the same length and corresponding type (Razvan)
* Argmax output dtype is now int64 instead of int32. (Olivier)
* Added the element-wise operation arccos. (Ian)
* Added sparse dot with dense grad output. (Yann Dauphin)
* Optimized to Usmm and UsmmCscDense in some case (Yann)
* Note: theano.dot and theano.sparse.structured_dot() always had a gradient with the same sparsity pattern as the inputs.
The new theano.sparse.dot() has a dense gradient for all inputs.
* GpuAdvancedSubtensor1 supports broadcasted dimensions. (Frederic)
* TensorVariable.zeros_like() and SparseVariable.zeros_like()
* theano.sandbox.cuda.cuda_ndarray.cuda_ndarray.device_properties() (Frederic)
* theano.sandbox.cuda.cuda_ndarray.cuda_ndarray.mem_info() return free and total gpu memory (Frederic)
* Theano flags compiledir_format. Keep the same default as before: compiledir_%(platform)s-%(processor)s-%(python_version)s. (Josh Bleecher Snyder)
* We also support the "theano_version" substitution.
* IntDiv c code (faster and allow this elemwise to be fused with other elemwise) (Pascal)
* Internal filter_variable mechanism in Type. (Pascal, Ian)
* Ifelse works on sparse.
* It makes use of gpu shared variable more transparent with theano.function updates and givens parameter.
* Added a_tensor.transpose(axes) axes is optional (James)
* theano.tensor.transpose(a_tensor, kwargs) We where ignoring kwargs, now it is used as the axes.
* a_CudaNdarray_object[*] = int, now works (Frederic)
* tensor_variable.size (as numpy) computes the product of the shape elements. (Olivier)
* sparse_variable.size (as scipy) computes the number of stored values. (Olivier)
* sparse_variable[N, N] now works (Li Yao, Frederic)
* sparse_variable[M:N, O:P] now works (Li Yao, Frederic, Pascal)
M, N, O, and P can be Python int or scalar tensor variables, None, or
omitted (sparse_variable[:, :M] or sparse_variable[:M, N:] work).
* tensor.tensordot can now be moved to GPU (Sander Dieleman,
Pascal, based on code from Tijmen Tieleman's gnumpy,
http://www.cs.toronto.edu/~tijmen/gnumpy.html)
* Many infer_shape implemented on sparse matrices op. (David W.F.)
* Added theano.sparse.verify_grad_sparse to easily allow testing grad of
sparse op. It support testing the full and structured gradient.
* The keys in our cache now store the hash of constants and not the constant values
themselves. This is significantly more efficient for big constant arrays. (Frederic B.)
* 'theano-cache list' lists key files bigger than 1M (Frederic B.)
* 'theano-cache list' prints an histogram of the number of keys per compiled module (Frederic B.)
* 'theano-cache list' prints the number of compiled modules per op class (Frederic B.)
* The Theano flag "nvcc.fastmath" is now also used for the cuda_ndarray.cu file.
* Add the header_dirs to the hard part of the compilation key. This is
currently used only by cuda, but if we use library that are only headers,
this can be useful. (Frederic B.)
* The Theano flag "nvcc.flags" is now included in the hard part of the key.
This mean that now we recompile all modules for each value of "nvcc.flags".
A change in "nvcc.flags" used to be ignored for module that were already
compiled. (Frederic B.)
* Alloc, GpuAlloc are not always pre-computed (constant_folding optimization)
at compile time if all their inputs are constant.
(Frederic B., Pascal L., reported by Sander Dieleman)
* New Op tensor.sort(), wrapping numpy.sort (Hani Almousli)


New optimizations:
* AdvancedSubtensor1 reuses preallocated memory if available (scan, c|py_nogc linker) (Frederic)
* dot22, dot22scalar work with complex. (Frederic)
* Generate Gemv/Gemm more often. (James)
* Remove scan when all computations can be moved outside the loop. (Razvan)
* scan optimization done earlier. This allows other optimizations to be applied. (Frederic, Guillaume, Razvan)
* exp(x) * sigmoid(-x) is now correctly optimized to the more stable form sigmoid(x). (Olivier)
* Added Subtensor(Rebroadcast(x)) => Rebroadcast(Subtensor(x)) optimization. (Guillaume)
* Made the optimization process faster. (James)
* Allow fusion of elemwise when the scalar op needs support code. (James)
* Better opt that lifts transpose around dot. (James)


Crashes fixed:
* T.mean crash at graph building time. (Ian)
* "Interactive debugger" crash fix. (Ian, Frederic)
* Do not call gemm with strides 0, some blas refuse it. (Pascal Lamblin)
* Optimization crash with gemm and complex. (Frederic)
* GPU crash with elemwise. (Frederic, some reported by Chris Currivan)
* Compilation crash with amdlibm and the GPU. (Frederic)
* IfElse crash. (Frederic)
* Execution crash fix in AdvancedSubtensor1 on 32 bit computers. (Pascal)
* GPU compilation crash on MacOS X. (Olivier)
* Support for OSX Enthought Python Distribution 7.x. (Graham Taylor, Olivier)
* When the subtensor inputs had 0 dimensions and the outputs 0 dimensions. (Frederic)
* Crash when the step to subtensor was not 1 in conjunction with some optimization. (Frederic, reported by Olivier Chapelle)
* Runtime crash related to an optimization with subtensor of alloc (reported by Razvan, fixed by Frederic)
* Fix dot22scalar cast of integer scalars (Justin Bayer, Frederic, Olivier)
* Fix runtime crash in gemm, dot22. FB
* Fix on 32bits computer: make sure all shape are int64.(Olivier)
* Fix to deque on python 2.4 (Olivier)
* Fix crash when not using c code (or using DebugMode) (not used by
default) with numpy 1.6*. Numpy has a bug in the reduction code that
made it crash. (Pascal)
* Crashes of blas functions (Gemv on CPU; Ger, Gemv and Gemm on GPU)
when matrices had non-unit stride in both dimensions (CPU and GPU),
or when matrices had negative strides (GPU only). In those cases,
we are now making copies. (Pascal)
* More cases supported in AdvancedIncSubtensor1. (Olivier D.)
* Fix crash when a broadcasted constant was used as input of an
elemwise Op and needed to be upcasted to match the op's output.
(Reported by John Salvatier, fixed by Pascal L.)
* Fixed a memory leak with shared variable (we kept a pointer to the original value) (Ian G.)


Known bugs:
* CAReduce with nan in inputs don't return the good output (`Ticket <https://www.assembla.com/spaces/theano/tickets/763>`_).
* This is used in tensor.{max,mean,prod,sum} and in the grad of PermuteRowElements.


Sandbox:
* cvm interface more consistent with current linker. (James)
* Now all tests pass with the linker=cvm flags.
* vm linker has a callback parameter. (James)
* review/finish/doc: diag/extract_diag. (Arnaud Bergeron, Frederic, Olivier)
* review/finish/doc: AllocDiag/diag. (Arnaud, Frederic, Guillaume)
* review/finish/doc: MatrixInverse, matrix_inverse. (Razvan)
* review/finish/doc: matrix_dot. (Razvan)
* review/finish/doc: det (determinent) op. (Philippe Hamel)
* review/finish/doc: Cholesky determinent op. (David)
* review/finish/doc: ensure_sorted_indices. (Li Yao)
* review/finish/doc: spectral_radius_boud. (Xavier Glorot)
* review/finish/doc: sparse sum. (Valentin Bisson)
* review/finish/doc: Remove0 (Valentin)
* review/finish/doc: SquareDiagonal (Eric)


Sandbox New features (not enabled by default):
* CURAND_RandomStreams for uniform and normal (not picklable, GPU only) (James)
* New sandbox.linalg.ops.pinv(pseudo-inverse) op (Razvan)


Documentation:
* Many updates. (Many people)
* Updates to install doc on MacOS. (Olivier)
* Updates to install doc on Windows. (David, Olivier)
* Doc on the Rop function (Ian)
* Added how to use scan to loop with a condition as the number of iteration. (Razvan)
* Added how to wrap in Theano an existing python function (in numpy, scipy, ...). (Frederic)
* Refactored GPU installation of Theano. (Olivier)


Others:
* Better error messages in many places. (Many people)
* PEP8 fixes. (Many people)
* Add a warning about numpy bug when using advanced indexing on a
tensor with more than 2**32 elements (the resulting array is not
correctly filled and ends with zeros). (Pascal, reported by David WF)
* Added Scalar.ndim=0 and ScalarSharedVariable.ndim=0 (simplify code) (Razvan)
* New min_informative_str() function to print graph. (Ian)
* Fix catching of exception. (Sometimes we used to catch interrupts) (Frederic, David, Ian, Olivier)
* Better support for utf string. (David)
* Fix pydotprint with a function compiled with a ProfileMode (Frederic)
* Was broken with change to the profiler.
* Warning when people have old cache entries. (Olivier)
* More tests for join on the GPU and CPU. (Frederic)
* Do not request to load the GPU module by default in scan module. (Razvan)
* Fixed some import problems. (Frederic and others)
* Filtering update. (James)
* On Windows, the default compiledir changed to be local to the
computer/user and not transferred with roaming profile. (Sebastian
Urban)
* New theano flag "on_shape_error". Defaults to "warn" (same as previous behavior):
it prints a warning when an error occurs when inferring the shape of some apply node.
The other accepted value is "raise" to raise an error when this happens. (Frederic)
* The buidbot now raises optimization/shape errors instead of just printing a warning. (Frederic)
* better pycuda tests (Frederic)
* check_blas.py now accept the shape and the number of iteration as parameter (Frederic)
* Fix opt warning when the opt ShapeOpt is disabled (enabled by default) (Frederic)
* More internal verification on what each op.infer_shape return. (Frederic, James)
* Argmax dtype to int64 (Olivier)
* Improved docstring and basic tests for the Tile Op (David).

Reviewers (alphabetical order):
* David, Frederic, Ian, James, Olivier, Razvan