Fixing last failed test for Release build

[khuck]

The release build has one last failing test:
http://omega.nic.uoregon.edu:8020/#/builders/2/builds/167/steps/16/logs/stdio

I tracked this down to temporaries output.tensor() and target.tensor() going out of scope in cat_cross_operation::cat_cross3d():

phylanx/src/plugins/keras_support/categorical_crossentropy_operation.cpp

Lines 246 to 297 in 64540c3

	primitive_argument_type cat_cross_operation::cat_cross3d(
	arg_type&& target, arg_type&& output, bool from_logits, int axis) const
	{
	assign_tensor<arg_type> output_(output);
	assign_tensor<arg_type> target_(target);
	if(from_logits)
	{
	output_ = softmax3d_axis2(output.tensor());
	}
	else
	{
	if(axis == 0) {
	auto norm = sum3d(output.tensor(),axis);
	for(std::size_t j = 0; j < output.tensor().pages(); ++j) {
	auto slice = blaze::pageslice(output.tensor(),j);
	slice = blaze::map(slice, norm,[](double s_,double n_){
	return s_/n_;
	});
	}
	} else if(axis == 1) {
	auto norm = sum3d(output.tensor(),axis);
	for(std::size_t j = 0; j < output.tensor().rows(); ++j) {
	auto slice = blaze::rowslice(output.tensor(),j);
	slice = blaze::map(slice, norm,[](double s_,double n_){
	return s_/n_;
	});
	}
	} else {
	auto norm = sum3d(output.tensor(),axis);
	for(std::size_t j = 0; j < output.tensor().columns(); ++j) {
	auto slice = blaze::columnslice(output.tensor(),j);
	slice = blaze::map(slice, norm,[](double s_,double n_){
	return s_/n_;
	});
	}
	}
	}
	target_ = blaze::map(
	target.tensor(), output.tensor(), [](double t_, double o_) {
	return -t_ *
	std::log((std::min)(clip_high, (std::max)(clip_low, o_)));
	});
	matrix_type ans = sum3d(target.tensor(), axis);
	if(axis == 1)
	ans = blaze::trans(ans);
	primitive_argument_type part1(std::move(ans)), part2(std::move(output));
	primitive_arguments_type both{part1, part2};
	phylanx::ir::range tup(both);
	return primitive_argument_type{ std::move(tup) };
	}

I'll submit a pull request for that issue (see #1188).

However, after fixing that crash, the python code fails assertions:

AssertionError: [[ 0.96681002  2.18922338  3.41163674  4.6340501 ]
 [ 5.85646346  7.07887682  8.30129018  9.52370354]
 [10.7461169  11.96853026 13.19094362 14.41335698]]
Uncaught exception. Entering post mortem debugging
Running 'cont' or 'step' will restart the program
> /var/lib/buildbot/slaves/phylanx/x86_64-gcc7-release/build/tests/unit/python/execution_tree/categorical_crossentropy.py(79)<module>()
-> assert np.all(np.abs(v1 - v2) < 1.0e-13), v1 - v2
(Pdb) p v1
array([[ 3.,  7., 11., 15.],
       [19., 23., 27., 31.],
       [35., 39., 43., 47.]])
(Pdb) p v2
array([[ 2.03318998,  4.81077662,  7.58836326, 10.3659499 ],
       [13.14353654, 15.92112318, 18.69870982, 21.47629646],
       [24.2538831 , 27.03146974, 29.80905638, 32.58664302]])
(Pdb) list
 74  	    o = 3 * np.ones(10)
 75  	    o[4] = 2
 76  	
 77  	    v1 = cc(t, o, flag, -1)
 78  	    v2 = cat_cross(t, o, flag)
 79  ->	    assert np.all(np.abs(v1 - v2) < 1.0e-13), v1 - v2
 80  	
 81  	    t = np.linspace(1.0e-9, 10e-9, 10)
 82  	    t[2] = 10
 83  	    o = 3 * np.ones(10)
 84  	    o[4] = 2
(Pdb) 

[hkaiser]

Which test is failing, concretely? Could you elaborate, please?

[khuck]

I am trying to hunt that down, which isn't easy because this program requires a mix of gdb and pdb since it's a python script test case.

[khuck]

Which test is failing, concretely? Could you elaborate, please?

it is the tests.unit.python.execution_tree.categorical_crossentropy test, if that's what you mean

[khuck]

@hkaiser Specifically, the temporaries that appear to go out of scope are those bound by the lambda and sum3d() calls here:

phylanx/src/plugins/keras_support/categorical_crossentropy_operation.cpp

Lines 284 to 290 in 64540c3

	target_ = blaze::map(
	target.tensor(), output.tensor(), [](double t_, double o_) {
	return -t_ *
	std::log((std::min)(clip_high, (std::max)(clip_low, o_)));
	});
	matrix_type ans = sum3d(target.tensor(), axis);

So while I've fixed the crash, the assertion fails on line 125 of tests/unit/python/execution_tree/categorical_crossentropy.py in the 3d test when axis is -1 and flag is True.

phylanx/tests/unit/python/execution_tree/categorical_crossentropy.py

Lines 111 to 125 in 64540c3

	for axis in [-1, 0, 1, 2]:
	for flag in [True, False]:
	t = np.linspace(1, 24, 24)
	t = np.reshape(t, (3, 4, 2))
	o = np.linspace(.1 * 1 + .3, .1 * 24 + .3, 24)
	o = np.reshape(o, (3, 4, 2))
	v2 = cat_cross(t, o, flag, axis)
	t = np.linspace(1, 24, 24)
	t = np.reshape(t, (3, 4, 2))
	o = np.linspace(.1 * 1 + .3, .1 * 24 + .3, 24)
	o = np.reshape(o, (3, 4, 2))
	v1 = cc(t, o, flag, axis)
	assert np.all(np.abs(v1 - v2) < 1.0e-13), v1 - v2

I get these matrices:

(Pdb) p v1
array([[ 3.,  7., 11., 15.],
       [19., 23., 27., 31.],
       [35., 39., 43., 47.]])
(Pdb) p v2
array([[ 2.03318998,  4.81077662,  7.58836326, 10.3659499 ],
       [13.14353654, 15.92112318, 18.69870982, 21.47629646],
       [24.2538831 , 27.03146974, 29.80905638, 32.58664302]])
(Pdb) p flag
True
(Pdb) p axis
-1
(Pdb) p o
array([[[0.4, 0.5],
        [0.6, 0.7],
        [0.8, 0.9],
        [1. , 1.1]],

       [[1.2, 1.3],
        [1.4, 1.5],
        [1.6, 1.7],
        [1.8, 1.9]],

       [[2. , 2.1],
        [2.2, 2.3],
        [2.4, 2.5],
        [2.6, 2.7]]])
(Pdb) p t
array([[[ 1.,  2.],
        [ 3.,  4.],
        [ 5.,  6.],
        [ 7.,  8.]],

       [[ 9., 10.],
        [11., 12.],
        [13., 14.],
        [15., 16.]],

       [[17., 18.],
        [19., 20.],
        [21., 22.],
        [23., 24.]]])