config – Theano Configuration

Guide

The config module contains many attributes that modify Theano’s behavior. Many of these attributes are consulted during the import of the theano module and many are assumed to be read-only.

As a rule, the attributes in this module should not be modified by user code.

Theano’s code comes with default values for these attributes, but you can override them from your .theanorc file, and override those values in turn by the THEANO_FLAGS environment variable.

The order of precedence is:

  1. an assignment to theano.config.<property>
  2. an assignment in THEANO_FLAGS
  3. an assignment in the .theanorc file (or the file indicated in THEANORC)

You can print out the current/effective configuration at any time by printing theano.config. For example, to see a list of all active configuration variables, type this from the command-line:

python -c 'import theano; print theano.config' | less

Environment Variables

THEANO_FLAGS

This is a list of comma-delimited key=value pairs that control Theano’s behavior.

For example, in bash, you can override your THEANORC defaults for <myscript>.py by typing this:

THEANO_FLAGS='floatX=float32,device=gpu0,nvcc.fastmath=True'  python <myscript>.py

If a value is defined several times in THEANO_FLAGS, the right-most definition is used. So, for instance, if THEANO_FLAGS='device=cpu,device=gpu0', then gpu0 will be used.

THEANORC

The location[s] of the .theanorc file[s] in ConfigParser format. It defaults to $HOME/.theanorc. On Windows, it defaults to $HOME/.theanorc:$HOME/.theanorc.txt to make Windows users’ life easier.

Here is the .theanorc equivalent to the THEANO_FLAGS in the example above:

[global]
floatX = float32
device = gpu0

[nvcc]
fastmath = True

Configuration attributes that are available directly in config (e.g. config.device, config.mode) should be defined in the [global] section. Attributes from a subsection of config (e.g. config.nvcc.fastmath, config.blas.ldflags) should be defined in their corresponding section (e.g. [nvcc], [blas]).

Multiple configuration files can be specified by separating them with ‘:’ characters (as in $PATH). Multiple configuration files will be merged, with later (right-most) files taking priority over earlier files in the case that multiple files specify values for a common configuration option. For example, to override system-wide settings with personal ones, set THEANORC=/etc/theanorc:~/.theanorc.

Config Attributes

The list below describes some of the more common and important flags that you might want to use. For the complete list (including documentation), import theano and print the config variable, as in:

python -c 'import theano; print theano.config' | less
config.device

String value: either 'cpu', 'gpu', 'gpu0', 'gpu1', 'gpu2', or 'gpu3'

Default device for computations. If gpu*, change the default to try to move computation to it and to put shared variable of float32 on it. Choose the default compute device for theano graphs. Setting this to a gpu* string will make theano to try by default to move computation to it. Also it will make theano put by default shared variable of float32 on it. 'gpu' lets the driver select the GPU to use, while 'gpu?' makes Theano try to use a specific device. If we are not able to use the GPU, either we fall back on the CPU, or an error is raised, depending on the force_device flag.

This flag’s value cannot be modified during the program execution.

Do not use upper case letters, only lower case even if NVIDIA use capital letters.

config.force_device

Bool value: either True or False

Default: False

If True and device=gpu*, we raise an error if we cannot use the specified device. If True and device=cpu, we disable the GPU. If False and device=gpu*, and if the specified device cannot be used, we warn and fall back to the CPU.

This is useful to run Theano’s tests on a computer with a GPU, but without running the GPU tests.

This flag’s value cannot be modified during the program execution.

config.init_gpu_device

String value: either '', 'gpu', 'gpu0', 'gpu1', 'gpu2', or 'gpu3'

Initialize the gpu device to use. When its value is gpu*, the theano flag device must be "cpu". Unlike device, setting this flag to a specific GPU will not try to use this device by default, in particular it will not move computations, nor shared variables, to the specified GPU.

This flag is useful to run GPU-specific tests on a particular GPU, instead of using the default one.

This flag’s value cannot be modified during the program execution.

config.pycuda.init

Bool value: either True or False

Default: False

If True, always initialize PyCUDA when Theano want to initialize the GPU. With PyCUDA version 2011.2.2 or earlier, PyCUDA must initialize the GPU before Theano does it. Setting this flag to True, ensure that, but always import PyCUDA. It can be done manually by importing theano.misc.pycuda_init before Theano initialize the GPU device. Newer version of PyCUDA (currently only in the trunk) don’t have this restriction.

config.print_active_device

Bool value: either True or False

Default: True

Print active device at when the GPU device is initialized.

config.enable_initial_driver_test

Bool value: either True or False

Default: True

Tests the nvidia driver when a GPU device is initialized.

config.floatX

String value: either ‘float64’ or ‘float32’

Default: ‘float64’

This sets the default dtype returned by tensor.matrix(), tensor.vector(), and similar functions. It also sets the default theano bit width for arguments passed as Python floating-point numbers.

config.warn_float64

String value: either ‘ignore’, ‘warn’, ‘raise’ or ‘pdb’

Default: ‘ignore’

When creating a TensorVariable with dtype float64, what should be done? This is useful to help find upcast to float64 in user code.

config.allow_gc

Bool value: either True or False

Default: True

This sets the default for the use of the Theano garbage collector for intermediate results. To use less memory, Theano frees the intermediate results as soon as they are no longer needed. Disabling Theano garbage collection allows Theano to reuse buffers for intermediate results between function calls. This speeds up Theano by no longer spending time reallocating space. This gives significant speed up on functions with many ops that are fast to execute, but this increases Theano’s memory usage.

config.scan.allow_output_prealloc

Bool value, either True or False

Default: True

This enables, or not, an optimization in Scan in which it tries to pre-allocate memory for its outputs. Enabling the optimization can give a significant speed up with Scan at the cost of slightly increased memory usage.

config.scan.allow_gc

Bool value, either True or False

Default: False

Allow/disallow gc inside of Scan.

If config.allow_gc is True, but config.scan.allow_gc is False, then we will gc the inner of scan after all iterations. This is the default.

config.openmp

Bool value: either True or False

Default: True if the environment variable OMP_NUM_THREADS!=1 or
if we detect more than 1 CPU core. Otherwise False.

Enable or not parallel computation on the CPU with OpenMP. It is the default value used when creating an Op that support it. The best is to define it via Theano configuration file or with the environment variable THEANO_FLAGS.

config.openmp_elemwise_minsize

Positive int value, default: 200000.

This specifies the vectors minimum size for which elemwise ops use openmp, if openmp is enabled.

config.cast_policy

String value: either ‘numpy+floatX’ or ‘custom’

Default: ‘custom’

This specifies how data types are implicitly figured out in Theano, e.g. for constants or in the results of arithmetic operations. The ‘custom’ value corresponds to a set of custom rules originally used in Theano (which can be partially customized, see e.g. the in-code help of tensor.NumpyAutocaster), and will be deprecated in the future. The ‘numpy+floatX’ setting attempts to mimic the numpy casting rules, although it prefers to use float32 numbers instead of float64 when config.floatX is set to ‘float32’ and the user uses data that is not explicitly typed as float64 (e.g. regular Python floats). Note that ‘numpy+floatX’ is not currently behaving exactly as planned (it is a work-in-progress), and thus you should consider it as experimental. At the moment it behaves differently from numpy in the following situations:

  • Depending on the value of config.int_division, the resulting type of a division of integer types with the / operator may not match that of numpy.
  • On mixed scalar / array operations, numpy tries to prevent the scalar from upcasting the array’s type unless it is of a fundamentally different type. Theano does not attempt to do the same at this point, so you should be careful that scalars may upcast arrays when they would not when using numpy. This behavior should change in the near future.
config.int_division

String value: either ‘int’, ‘floatX’ or ‘raise’

Default: ‘int’

Specifies what to do when one tries to compute x / y, where both x and y are of integer types (possibly unsigned). ‘int’ means an integer is returned (as in Python 2.X), but this behavior is deprecated. ‘floatX’ returns a number of type given by config.floatX. ‘raise’ is the safest choice (and will become default in a future release of Theano) and raises an error when one tries to do such an operation, enforcing the use of the integer division operator (//) (if a float result is intended, either cast one of the arguments to a float, or use x.__truediv__(y)).

config.mode

String value: ‘Mode’, ‘ProfileMode’(deprecated), ‘DebugMode’, ‘FAST_RUN’, ‘FAST_COMPILE’

Default ‘Mode’

This sets the default compilation mode for theano functions. By default the mode Mode is equivalent to FAST_RUN. See Config attribute linker and optimizer.

config.profile

Bool value: either True or False

Default False

Do the vm/cvm linkers profile the execution time of Theano functions?

See Profiling Theano function for examples.

config.profile_memory

Bool value: either True or False

Default False

Do the vm/cvm linkers profile the memory usage of Theano functions? It only works when profile=True.

config.profile_optimizer

Bool value: either True or False

Default False

Do the vm/cvm linkers profile the optimization phase when compiling a Theano function? It only works when profile=True.

config.profiling.n_apply

Positive int value, default: 20.

The number of Apply nodes to print in the profiler output

config.profiling.n_ops

Positive int value, default: 20.

The number of Ops to print in the profiler output

config.profiling.min_memory_size

Positive int value, default: 1024.

For the memory profile, do not print Apply nodes if the size of their outputs (in bytes) is lower than this.

config.profiling.min_peak_memory

Bool value: either True or False

Default False

Does the memory profile print the min peak memory usage? It only works when profile=True, profile_memory=True

config.profiling.destination

String value: ‘stderr’, ‘stdout’, or a name of a file to be created

Default ‘stderr’

Name of the destination file for the profiling output. The profiling output can be either directed to stderr (default), or stdout or an arbitrary file.

config.profiling.debugprint

Bool value: either True or False

Default False

Do a debugprint of the profiled functions

config.lib.amdlibm

Bool value: either True or False

Default False

This makes the compilation use the amdlibm library, which is faster than the standard libm.

config.lib.cnmem

Float value: >= 0

Do we enable CNMeM or not (a faster CUDA memory allocator). In Theano dev version until 0.7.1 is released.

That library is included in Theano, you do not need to install it.

The value represents the start size (in MB or % of total GPU memory) of the memory pool. If more memory are needed, it will try to get more, but this can cause more memory fragmentation:

  • 0: not enabled.
  • 0 < N <= 1: % of the total GPU memory (clipped to .985 for driver memory)
  • > 0: use that number of MB of memory.

Default 0 (but should change later)

Note

This could cause memory fragmentation. So if you have a memory error while using cnmem, try to allocate more memory at the start or disable it. If you try this, report your result on :ref`theano-dev`.

config.linker

String value: ‘c|py’, ‘py’, ‘c’, ‘c|py_nogc’

Default: ‘c|py’

When the mode is Mode, it sets the default linker used. See Configuration Settings and Compiling Modes for a comparison of the different linkers.

config.optimizer

String value: ‘fast_run’, ‘merge’, ‘fast_compile’, ‘None’

Default: ‘fast_run’

When the mode is Mode, it sets the default optimizer used.

config.on_opt_error

String value: ‘warn’, ‘raise’, ‘pdb’ or ‘ignore’

Default: ‘warn’

When a crash occurs while trying to apply some optimization, either warn the user and skip this optimization (‘warn’), raise the exception (‘raise’), fall into the pdb debugger (‘pdb’) or ignore it (‘ignore’). We suggest to never use ‘ignore’ except in tests.

If you encounter a warning, report it on theano-dev.

config.assert_no_cpu_op

String value: ‘ignore’ or ‘warn’ or ‘raise’ or ‘pdb’

Default: ‘ignore’

If there is a CPU op in the computational graph, depending on its value; this flag can either raise a warning, an exception or stop the compilation with pdb.

config.on_shape_error

String value: ‘warn’ or ‘raise’

Default: ‘warn’

When an exception is raised when inferring the shape of some apply node, either warn the user and use a default value (‘warn’), or raise the exception (‘raise’).

config.warn.ignore_bug_before

String value: ‘None’, ‘all’, ‘0.3’, ‘0.4’, ‘0.4.1’, ‘0.5’, ‘0.6’, ‘0.7’

Default: ‘0.6’

When we fix a Theano bug that generated bad results under some circumstances, we also make Theano raise a warning when it encounters the same circumstances again. This helps to detect if said bug had affected your past experiments, as you only need to run your experiment again with the new version, and you do not have to understand the Theano internal that triggered the bug. A better way to detect this will be implemented. See this ticket.

This flag allows new users not to get warnings about old bugs, that were fixed before their first checkout of Theano. You can set its value to the first version of Theano that you used (probably 0.3 or higher)

None means that all warnings will be displayed. all means all warnings will be ignored.

It is recommended that you put a version, so that you will see future warnings. It is also recommended you put this into your .theanorc, so this setting will always be used.

This flag’s value cannot be modified during the program execution.

config.base_compiledir

Default: On Windows: $LOCALAPPDATA\Theano if $LOCALAPPDATA is defined, otherwise and on other systems: ~/.theano.

This directory stores the platform-dependent compilation directories.

This flag’s value cannot be modified during the program execution.

config.compiledir_format

Default: “compiledir_%(platform)s-%(processor)s-%(python_version)s-%(python_bitwidth)s”

This is a Python format string that specifies the subdirectory of config.base_compiledir in which to store platform-dependent compiled modules. To see a list of all available substitution keys, run python -c "import theano; print theano.config", and look for compiledir_format.

This flag’s value cannot be modified during the program execution.

config.compiledir

Default: config.base_compiledir/config.compiledir_format

This directory stores dynamically-compiled modules for a particular platform.

This flag’s value cannot be modified during the program execution.

config.blas.ldflags

Default: ‘-lblas’

Link arguments to link against a (Fortran) level-3 blas implementation. The default will test if ‘-lblas’ work. If not, we will disable our c code for BLAS.

config.experimental.local_alloc_elemwise_assert

Bool value: either True or False

Default: True

When the local_alloc_optimization is applied, add an assert to highlight shape errors.

Without such asserts this optimization could hide errors in the user code. We add the assert only if we can’t infer that the shapes are equivalent. As such this optimization does not always introduce an assert in the graph. Removing the assert could speed up execution.

config.cuda.root

Default: $CUDA_ROOT or failing that, “/usr/local/cuda”

A directory with bin/, lib/, include/ folders containing cuda utilities.

config.dnn.conv.workmem

Deprecated, use dnn.conv.algo_fwd.

config.dnn.conv.workmem_bwd

Deprecated, use dnn.conv.algo_bwd.

config.dnn.conv.algo_fwd
String value: small, none, large, fft, guess_once,
guess_on_shape_change, time_once, time_on_shape_change.

Default: small

3d convolution only support none, guess_once, guess_on_shape_change, time_once, time_on_shape_change.

config.dnn.conv.algo_bwd
String value: none, deterministic, fft, guess_once,
guess_on_shape_change, time_once, time_on_shape_change.

Default: none

3d convolution only support none, guess_once, guess_on_shape_change, time_once, time_on_shape_change.

config.gcc.cxxflags

Default: “”

Extra parameters to pass to gcc when compiling. Extra include paths, library paths, configuration options, etc.

config.cxx

Default: Full path to g++ if g++ is present. Empty string otherwise.

Indicates which C++ compiler to use. If empty, no C++ code is compiled. Theano automatically detects whether g++ is present and disables C++ compilation when it is not. On darwin systems (Mac OS X), it preferably looks for clang++ and uses that if available.

We print a warning if we detect that no compiler is present. It is recommended to run with C++ compilation as Theano will be much slower otherwise.

This can be any compiler binary (full path or not) but things may break if the interface is not g++-compatible to some degree.

config.nvcc.fastmath

Default: False

If true, this will enable fastmath (–use_fast_math) mode for compiled cuda code which makes div and sqrt faster at the cost of precision. This also disables support for denormal numbers.

config.optimizer_excluding

Default: “”

A list of optimizer tags that we don’t want included in the default Mode. If multiple tags, separate them by ‘:’. Ex: to remove the elemwise inplace optimizer(slow for big graph), use the flags: optimizer_excluding:inplace_opt, where inplace_opt is the name of that optimization.

This flag’s value cannot be modified during the program execution.

config.optimizer_including

Default: “”

A list of optimizer tags that we want included in the default Mode. If multiple tags, separate them by ‘:’.

This flag’s value cannot be modified during the program execution.

config.optimizer_requiring

Default: “”

A list of optimizer tags that we require for optimizer in the default Mode. If multiple tags, separate them by ‘:’.

This flag’s value cannot be modified during the program execution.

config.optimizer_verbose

Bool value: either True or False

Default: False

When True, we print on the stdout the optimization applied.

config.nocleanup

Bool value: either True or False

Default: False

If False, source code files are removed when they are not needed anymore. This means files whose compilation failed are deleted. Set to True to keep those files in order to debug compilation errors.

config.compile

This section contains attributes which influence the compilation of C code for ops. Due to historical reasons many attributes outside of this section also have an influence over compilation, most notably ‘cxx’. This is not expected to change any time soon.

config.compile.timeout

Positive int value, default: compile.wait * 24

Time to wait before an unrefreshed lock is broken and stolen. This is in place to avoid manual cleanup of locks in case a process crashed and left a lock in place.

The refresh time is automatically set to half the timeout value.

config.compile.wait

Positive int value, default: 5

Time to wait between attempts at grabbing the lock if the first attempt is not successful. The actual time will be between compile.wait and compile.wait * 2 to avoid a crowding effect on lock.

config.DebugMode

This section contains various attributes configuring the behaviour of mode DebugMode. See directly this section for the documentation of more configuration options.

config.DebugMode.check_preallocated_output

Default: ''

A list of kinds of preallocated memory to use as output buffers for each Op’s computations, separated by :. Implemented modes are:

  • "initial": initial storage present in storage map (for instance, it can happen in the inner function of Scan),
  • "previous": reuse previously-returned memory,
  • "c_contiguous": newly-allocated C-contiguous memory,
  • "f_contiguous": newly-allocated Fortran-contiguous memory,
  • "strided": non-contiguous memory with various stride patterns,
  • "wrong_size": memory with bigger or smaller dimensions,
  • "ALL": placeholder for all of the above.

In order not to test with preallocated memory, use an empty string, "".

config.DebugMode.check_preallocated_output_ndim

Positive int value, default: 4.

When testing with “strided” preallocated output memory, test all combinations of strides over that number of (inner-most) dimensions. You may want to reduce that number to reduce memory or time usage, but it is advised to keep a minimum of 2.

config.DebugMode.warn_input_not_reused

Bool value, default: True

Generate a warning when the destroy_map or view_map tell that an op work inplace, but the op did not reuse the input for its output.

config.NanGuardMode.nan_is_error

Bool value, default: True

Controls whether NanGuardMode generates an error when it sees a nan.

config.NanGuardMode.inf_is_error

Bool value, default: True

Controls whether NanGuardMode generates an error when it sees an inf.

config.NanGuardMode.nan_is_error

Bool value, default: True

Controls whether NanGuardMode generates an error when it sees a big value (>1e10).

config.numpy

This section contains different attributes for configuring numpy’s behaviour, described by numpy.seterr.

config.numpy.seterr_all

String Value: 'ignore', 'warn', 'raise', 'call', 'print', 'log', 'None'

Default: 'ignore'

Set the default behaviour described by numpy.seterr.

'None' means that numpy’s default behaviour will not be changed (unless one of the other config.numpy.seterr_* overrides it), but this behaviour can change between numpy releases.

This flag sets the default behaviour for all kinds of floating-pont errors, and it can be overriden for specific errors by setting one (or more) of the flags below.

This flag’s value cannot be modified during the program execution.

config.numpy.seterr_divide

String Value: 'None', 'ignore', 'warn', 'raise', 'call', 'print', 'log'

Default: 'None'

Sets numpy’s behavior for division by zero. 'None' means using the default, defined by config.numpy.seterr_all.

This flag’s value cannot be modified during the program execution.

config.numpy.seterr_over

String Value: 'None', 'ignore', 'warn', 'raise', 'call', 'print', 'log'

Default: 'None'

Sets numpy’s behavior for floating-point overflow. 'None' means using the default, defined by config.numpy.seterr_all.

This flag’s value cannot be modified during the program execution.

config.numpy.seterr_under

String Value: 'None', 'ignore', 'warn', 'raise', 'call', 'print', 'log'

Default: 'None'

Sets numpy’s behavior for floating-point underflow. 'None' means using the default, defined by config.numpy.seterr_all.

This flag’s value cannot be modified during the program execution.

config.numpy.seterr_invalid

String Value: 'None', 'ignore', 'warn', 'raise', 'call', 'print', 'log'

Default: 'None'

Sets numpy’s behavior for invalid floating-point operation. 'None' means using the default, defined by config.numpy.seterr_all.

This flag’s value cannot be modified during the program execution.

config.compute_test_value

String Value: 'off', 'ignore', 'warn', 'raise'.

Default: 'off'

Setting this attribute to something other than 'off' activates a debugging mechanism, where Theano executes the graph on-the-fly, as it is being built. This allows the user to spot errors early on (such as dimension mis-match), before optimizations are applied.

Theano will execute the graph using the Constants and/or shared variables provided by the user. Purely symbolic variables (e.g. x = T.dmatrix()) can be augmented with test values, by writing to their 'tag.test_value' attribute (e.g. x.tag.test_value = numpy.random.rand(5,4)).

When not 'off', the value of this option dictates what happens when an Op’s inputs do not provide appropriate test values:

  • 'ignore' will silently skip the debug mechanism for this Op
  • 'warn' will raise a UserWarning and skip the debug mechanism for this Op
  • 'raise' will raise an Exception
config.compute_test_value_opt

As compute_test_value, but it is the value used during Theano optimization phase. Theano user’s do not need to use this. This is to help debug shape error in Theano optimization.

config.print_test_value

Bool value, default: False

If 'True', Theano will override the ‘__str__’ method of its variables to also print the tag.test_value when this is available.

config.reoptimize_unpickled_function

Bool value, default: False (changed in master after Theano 0.7 release)

Theano users can use the standard python pickle tools to save a compiled theano function. When pickling, both graph before and after the optimization are saved, including shared variables. When set to True, the graph is reoptimized when being unpickled. Otherwise, skip the graph optimization and use directly the optimized graph.

config.exception_verbosity

String Value: 'low', 'high'.

Default: 'low'

If 'low', the text of exceptions will generally refer to apply nodes with short names such as 'Elemwise{add_no_inplace}'. If 'high', some exceptions will also refer to apply nodes with long descriptions like:

A. Elemwise{add_no_inplace}
      B. log_likelihood_v_given_h
      C. log_likelihood_h
config.cmodule.warn_no_version

Bool value, default: False

If True, will print a warning when compiling one or more Op with C code that can’t be cached because there is no c_code_cache_version() function associated to at least one of those Ops.

Bool value, default: False

If set to True, breaks certain MacOS installations with the infamous Bus Error.

config.cmodule.remove_gxx_opt

Bool value, default: False

If True, will remove the -O* parameter passed to g++. This is useful to debug in gdb modules compiled by Theano. The parameter -g is passed by default to g++.

config.cmodule.compilation_warning

Bool value, default: False

If True, will print compilation warnings.

config.cmodule.preload_cache

Bool value, default: False

If set to True, will preload the C module cache at import time