The DebugMode evaluation mode includes a number of self-checks and assertions that can help to diagnose several kinds of programmer errors that can lead to incorrect output.
It is much slower to evaluate a function or method with DebugMode than
it would be in
'FAST_RUN' or even
'FAST_COMPILE'. We recommended you use
DebugMode during development, but not when you launch 1000 processes on
DebugMode can be used as follows:
import theano from theano import tensor from theano.compile.debugmode import DebugMode x = tensor.dscalar('x') f = theano.function([x], 10*x, mode='DebugMode') f(5) f(0) f(7)
It can also be used by setting the configuration variable
It can also be used by passing a DebugMode instance as the mode, as in
>>> f = theano.function([x], 10*x, mode=DebugMode(check_c_code=False))
If any problem is detected, DebugMode will raise an exception according to
what went wrong, either at call time (
f(5)) or compile time (
f = theano.function(x, 10*x, mode='DebugMode')). These exceptions
should not be ignored; talk to your local Theano guru or email the
users list if you cannot make the exception go away.
Some kinds of errors can only be detected for certain input value combinations.
In the example above, there is no way to guarantee that a future call to say,
f(-1) won’t cause a problem. DebugMode is not a silver bullet.
If you instantiate DebugMode using the constructor
rather than the keyword
DebugMode you can configure its behaviour via
Evaluation Mode that detects internal theano errors.
This mode catches several kinds of internal error:
- inconsistent outputs when calling the same Op twice with the same inputs, for instance if c_code and perform implementations, are inconsistent, or in case of incorrect handling of output memory (see BadThunkOutput)
- a variable replacing another when their runtime values don’t match. This is a symptom of an incorrect optimization step, or faulty Op implementation (raises BadOptimization)
- stochastic optimization ordering (raises StochasticOrder)
- incomplete destroy_map specification (raises BadDestroyMap)
- an op that returns an illegal value not matching the output Variable Type (raises InvalidValueError)
Each of these exceptions inherits from the more generic DebugModeError.
If there are no internal errors, this mode behaves like FAST_RUN or FAST_COMPILE, but takes a little longer and uses more memory.
If there are internal errors, this mode will raise an DebugModeError exception.
stability_patience = config.DebugMode.patience
When checking for the stability of optimization, recompile the graph this many times. Default 10.
check_c_code = config.DebugMode.check_c
Should we evaluate (and check) the c_code implementations?
check_py_code = config.DebugMode.check_py
Should we evaluate (and check) the perform implementations?
check_isfinite = config.DebugMode.check_finite
Should we check for (and complain about)
require_matching_strides = config.DebugMode.check_strides
Check for (and complain about) Ops whose python and C outputs are ndarrays with different strides. (This can catch bugs, but is generally overly strict.)
0 -> no check, 1 -> warn, 2 -> err.
__init__(self, optimizer='fast_run', stability_patience=None, check_c_code=None, check_py_code=None, check_isfinite=None, require_matching_strides=None, linker=None)¶
Initialize member variables.
If any of these arguments (except optimizer) is not None, it overrides the class default. The linker arguments is not used. It is set there to allow Mode.requiring() and some other functions to work with DebugMode too.
The keyword version of DebugMode (which you get by using
is quite strict, and can raise several different Exception types.
There following are DebugMode exceptions you might encounter:
This is a generic error. All the other exceptions inherit from this one. This error is typically not raised directly. However, you can use
except DebugModeError: ...to catch any of the more specific types of Exception.
This exception means that different calls to the same Op with the same inputs did not compute the same thing like they were supposed to. For instance, it can happen if the python (
perform) and c (
c_code) implementations of the Op are inconsistent (the problem might be a bug in either
c_code(or both)). It can also happen if
c_codedoes not handle correctly output memory that has been preallocated (for instance, if it did not clear the memory before accumulating into it, or if it assumed the memory layout was C-contiguous even if it is not).
This exception indicates that an Optimization replaced one variable (say V1) with another one (say V2) but at runtime, the values for V1 and V2 were different. This is something that optimizations are not supposed to do.
It can be tricky to identify the one-true-cause of an optimization error, but this exception provides a lot of guidance. Most of the time, the exception object will indicate which optimization was at fault. The exception object also contains information such as a snapshot of the before/after graph where the optimization introduced the error.
This happens when an Op’s
c_code()modifies an input that it wasn’t supposed to. If either the
c_codeimplementation of an Op might modify any input, it has to advertise that fact via the
For detailed documentation on the
destroy_mapattribute, see Inplace operations.
This happens when an Op’s perform() or c_code() creates an alias or alias-like dependency between an input and an output... and it didn’t warn the optimization system via the
For detailed documentation on the
view_mapattribute, see Views.
This happens when an optimization does not perform the same graph operations in the same order when run several times in a row. This can happen if any steps are ordered by
id(object)somehow, such as via the default object hash function. A Stochastic optimization invalidates the pattern of work whereby we debug in DebugMode and then run the full-size jobs in FAST_RUN.
This happens when some Op’s
c_codeimplementation computes an output that is invalid with respect to the type of the corresponding output variable. Like if it returned a complex-valued ndarray for a
This can also be triggered when floating-point values such as NaN and Inf are introduced into the computations. It indicates which Op created the first NaN. These floating-point values can be allowed by passing the
check_isfinite=Falseargument to DebugMode.