def test_debug_multi_output():
input_dim = 2
num_output_classes = 2
f_input = input_variable(input_dim,
np.float32,
needs_gradient=True,
name='features')
p = parameter(shape=(input_dim, ), init=10, name='p')
comb = combine([f_input, p])
ins = InStream(['n', 'n', 'n', 'n', 'n'])
outs = OutStream()
z = times(comb.outputs[0], comb.outputs[1], name='z')
z = debug_model(z, ins, outs)
l_input = input_variable(num_output_classes, np.float32, name='labels')
loss = cross_entropy_with_softmax(z, l_input)
eval_error = classification_error(z, l_input)
_train(z, loss, eval_error, loss.find_by_name('features'),
loss.find_by_name('labels'), num_output_classes, 1)
# outs.written contains something like
# =================================== forward ===================================
# Parameter('p', [], [2]) with uid 'Parameter4'
# Input('features', [#, *], [2]) with uid 'Input3'
# Times: Output('UserDefinedFunction12_Output_0', [#, *], [2]), Output('UserDefinedFunction15_Output_0', [], [2]) -> Output('z', [#, *], [2 x 2]) with uid 'Times21'
# =================================== backward ===================================
# Times: Output('UserDefinedFunction12_Output_0', [#, *], [2]), Output('UserDefinedFunction15_Output_0', [], [2]) -> Output('z', [#, *], [2 x 2]) with uid 'Times21'
# Input('features', [#, *], [2]) with uid 'Input3'
# Parameter('p', [], [2]) with uid 'Parameter4' assert outs.written == out_stuff
assert len(outs.written) == 8
v_p = "Parameter('p', "
v_i = "Input('features'"
v_t = 'Times: '
assert outs.written[0].startswith('=') and 'forward' in outs.written[0]
line_1, line_2, line_3 = outs.written[1:4]
assert outs.written[4].startswith('=') and 'backward' in outs.written[4]
line_5, line_6, line_7 = outs.written[5:8]
assert line_5.startswith(v_t)
assert line_6.startswith(v_p) and line_7.startswith(v_i) or \
line_6.startswith(v_i) and line_7.startswith(v_p)
def test_debug_multi_output():
input_dim = 2
num_output_classes = 2
f_input = input_variable(input_dim, np.float32,
needs_gradient=True, name='features')
p = parameter(shape=(input_dim,), init=10, name='p')
comb = combine([f_input, p])
ins = InStream(['n', 'n', 'n', 'n', 'n'])
outs = OutStream()
z = times(comb.outputs[0], comb.outputs[1], name='z')
z = debug_model(z, ins, outs)
l_input = input_variable(num_output_classes, np.float32, name='labels')
loss = cross_entropy_with_softmax(z, l_input)
eval_error = classification_error(z, l_input)
_train(z, loss, eval_error,
loss.find_by_name('features'),
loss.find_by_name('labels'),
num_output_classes, 1)
# outs.written contains something like
# =================================== forward ===================================
# Parameter('p', [], [2]) with uid 'Parameter4'
# Input('features', [#, *], [2]) with uid 'Input3'
# Times: Output('UserDefinedFunction12_Output_0', [#, *], [2]), Output('UserDefinedFunction15_Output_0', [], [2]) -> Output('z', [#, *], [2 x 2]) with uid 'Times21'
# =================================== backward ===================================
# Times: Output('UserDefinedFunction12_Output_0', [#, *], [2]), Output('UserDefinedFunction15_Output_0', [], [2]) -> Output('z', [#, *], [2 x 2]) with uid 'Times21'
# Input('features', [#, *], [2]) with uid 'Input3'
# Parameter('p', [], [2]) with uid 'Parameter4' assert outs.written == out_stuff
assert len(outs.written) == 8
v_p = "Parameter('p', "
v_i = "Input('features'"
v_t = 'Times: '
assert outs.written[0].startswith('=') and 'forward' in outs.written[0]
line_1, line_2, line_3 = outs.written[1:4]
assert outs.written[4].startswith('=') and 'backward' in outs.written[4]
line_5, line_6, line_7 = outs.written[5:8]
assert line_5.startswith(v_t)
assert line_6.startswith(v_p) and line_7.startswith(v_i) or \
line_6.startswith(v_i) and line_7.startswith(v_p)
