def replace_dense_softmax_by_dense_linear(all_layers, n_features,
nonlin_before_merge, batch_norm_before_merge):
"""Replace dense/conv (n_classes) -> reshape -> softmax
by dense/conv (n_features) -> reshape"""
reshape_layer = [l for l in all_layers if l.__class__.__name__ == 'FinalReshapeLayer']
assert len(reshape_layer) == 1
reshape_layer = reshape_layer[0]
input_to_reshape = reshape_layer.input_layer
# We expect a linear conv2d as "final dense" before the reshape...
assert input_to_reshape.__class__.__name__ == 'Conv2DLayer', (
"expect conv before reshape")
assert input_to_reshape.nonlinearity.func_name == 'linear'
# recreate with different number of filters
assert input_to_reshape.stride == (1,1)
new_input_to_reshape = Conv2DLayer(input_to_reshape.input_layer,
num_filters=n_features,
filter_size=input_to_reshape.filter_size, nonlinearity=nonlin_before_merge,
name='final_dense')
if batch_norm_before_merge:
new_input_to_reshape = batch_norm(new_input_to_reshape,
alpha=0.1,epsilon=0.01)
new_reshape_l = FinalReshapeLayer(new_input_to_reshape)
return lasagne.layers.get_all_layers(new_reshape_l)
def test_inverse_dilation_layer():
from convvisual.receptive_field.convvisual import receptive_field_build_deconv_layers
from braindecode.veganlasagne.layer_util import print_layers
input_var = T.tensor4('inputs').astype(theano.config.floatX)
network1 = lasagne.layers.InputLayer(shape=[None, 1, 15, 1],
input_var=input_var)
network2 = StrideReshapeLayer(network1,
n_stride=2,
invalid_fill_value=np.nan)
network3 = StrideReshapeLayer(network2,
n_stride=2,
invalid_fill_value=np.nan)
network = FinalReshapeLayer(network3, remove_invalids=True, flatten=False)
print_layers(network)
#network = DilationSeparate2DLayer(network,(4,1),((0,1),(0,0)),fill_value=np.nan)
#network = DilationMerge2DLayer(network,(2,1),((0,0),(0,0)))
#network = DilationMerge2DLayer(network,(2,1),((0,1),(0,0)))
deconv_network = receptive_field_build_deconv_layers(network, network2)
print_layers(deconv_network)
input = to_4d_time_array([range(1, 16), range(16, 31)])
preds_cnt = lasagne.layers.get_output(
lasagne.layers.get_all_layers(network))
pred_cnt_func = theano.function([input_var], preds_cnt)
layer_activations = pred_cnt_func(input)
print layer_activations
layer_activations2 = lasagne.layers.get_output(
lasagne.layers.get_all_layers(deconv_network), layer_activations[-1])
print layer_activations[-1].shape
print layer_activations2[3].eval()
assert (np.array_equal(layer_activations2[3].eval(), layer_activations[0]))
def get_layers(self):
l = InputLayer([None, self.in_chans, self.input_time_length, 1])
if self.split_first_layer:
l = DimshuffleLayer(l, pattern=[0, 3, 2, 1])
l = Conv2DLayer(l,
num_filters=self.n_filters_time,
filter_size=[self.filter_time_length, 1],
nonlinearity=identity,
name='time_conv')
l = Conv2DAllColsLayer(l,
num_filters=self.n_filters_spat,
filter_size=[1, -1],
nonlinearity=identity,
name='spat_conv')
else: #keep channel dim in first dim, so it will also be convolved over
l = Conv2DLayer(l,
num_filters=self.num_filters_time,
filter_size=[self.filter_time_length, 1],
nonlinearity=identity,
name='time_conv')
if self.batch_norm:
l = BatchNormLayer(l,
epsilon=1e-4,
alpha=self.batch_norm_alpha,
nonlinearity=self.conv_nonlin)
else:
l = NonlinearityLayer(l, nonlinearity=self.conv_nonlin)
l = Pool2DLayer(l,
pool_size=[self.pool_time_length, 1],
stride=[1, 1],
mode=self.pool_mode)
l = NonlinearityLayer(l, self.pool_nonlin)
l = StrideReshapeLayer(l, n_stride=self.pool_time_stride)
l = DropoutLayer(l, p=self.drop_prob)
l = Conv2DLayer(l,
num_filters=self.n_classes,
filter_size=[self.final_dense_length, 1],
nonlinearity=identity,
name='final_dense')
l = FinalReshapeLayer(l)
l = NonlinearityLayer(l, softmax)
return lasagne.layers.get_all_layers(l)
def get_layers(self):
l = InputLayer([None, self.in_chans, self.input_time_length, 1])
if self.split_first_layer:
l = DimshuffleLayer(l, pattern=[0, 3, 2, 1])
l = DropoutLayer(l, p=self.drop_in_prob)
l = Conv2DLayer(l,
num_filters=self.num_filters_time,
filter_size=[self.filter_time_length, 1],
nonlinearity=identity,
name='time_conv')
if self.double_time_convs:
l = Conv2DLayer(l,
num_filters=self.num_filters_time,
filter_size=[self.filter_time_length, 1],
nonlinearity=identity,
name='time_conv')
l = Conv2DAllColsLayer(l,
num_filters=self.num_filters_spat,
filter_size=[1, -1],
nonlinearity=identity,
name='spat_conv')
else: #keep channel dim in first dim, so it will also be convolved over
l = DropoutLayer(l, p=self.drop_in_prob)
l = Conv2DLayer(l,
num_filters=self.num_filters_time,
filter_size=[self.filter_time_length, 1],
nonlinearity=identity,
name='time_conv')
if self.double_time_convs:
l = Conv2DLayer(l,
num_filters=self.num_filters_time,
filter_size=[self.filter_time_length, 1],
nonlinearity=identity,
name='time_conv')
if self.batch_norm:
l = BatchNormLayer(l,
epsilon=1e-4,
alpha=self.batch_norm_alpha,
nonlinearity=self.first_nonlin)
else:
l = NonlinearityLayer(l, nonlinearity=self.first_nonlin)
l = Pool2DLayer(l,
pool_size=[self.pool_time_length, 1],
stride=[1, 1],
mode=self.first_pool_mode)
l = StrideReshapeLayer(l, n_stride=self.pool_time_stride)
l = NonlinearityLayer(l, self.first_pool_nonlin)
def conv_pool_block(l, num_filters, filter_length, i_block):
l = DropoutLayer(l, p=self.drop_prob)
l = Conv2DLayer(l,
num_filters=num_filters,
filter_size=[filter_length, 1],
nonlinearity=identity,
name='combined_conv_{:d}'.format(i_block))
if self.double_time_convs:
l = Conv2DLayer(l,
num_filters=num_filters,
filter_size=[filter_length, 1],
nonlinearity=identity,
name='combined_conv_{:d}'.format(i_block))
if self.batch_norm:
l = BatchNormLayer(l,
epsilon=1e-4,
alpha=self.batch_norm_alpha,
nonlinearity=self.later_nonlin)
else:
l = NonlinearityLayer(l, nonlinearity=self.later_nonlin)
l = Pool2DLayer(l,
pool_size=[self.pool_time_length, 1],
stride=[1, 1],
mode=self.later_pool_mode)
l = StrideReshapeLayer(l, n_stride=self.pool_time_stride)
l = NonlinearityLayer(l, self.later_pool_nonlin)
return l
l = conv_pool_block(l, self.num_filters_2, self.filter_length_2, 2)
l = conv_pool_block(l, self.num_filters_3, self.filter_length_3, 3)
l = conv_pool_block(l, self.num_filters_4, self.filter_length_4, 4)
# Final part, transformed dense layer
l = DropoutLayer(l, p=self.drop_prob)
l = Conv2DLayer(l,
num_filters=self.n_classes,
filter_size=[self.final_dense_length, 1],
nonlinearity=identity,
name='final_dense')
l = FinalReshapeLayer(l)
l = NonlinearityLayer(l, self.final_nonlin)
return lasagne.layers.get_all_layers(l)
def get_layers(self):
def resnet_residual_block(model,
increase_units_factor=None,
half_time=False):
"""Calling residual_block function with correct attributes
from this object.
Parameters
----------
model :
increase_units_factor :
(Default value = None)
half_time :
(Default value = False)
Returns
-------
Final layer of created residual block.
"""
return residual_block(model,
batch_norm_epsilon=self.batch_norm_epsilon,
batch_norm_alpha=self.batch_norm_alpha,
increase_units_factor=increase_units_factor,
half_time=half_time,
nonlinearity=self.nonlinearity,
projection=self.projection,
survival_prob=self.survival_prob,
add_after_nonlin=self.add_after_nonlin,
reduction_method=self.reduction_method,
reduction_pool_mode=self.reduction_pool_mode)
model = InputLayer([None, self.in_chans, self.input_time_length, 1])
if self.split_first_layer:
# shift channel dim out
model = DimshuffleLayer(model, (0, 3, 2, 1))
# first timeconv
model = Conv2DLayer(model,
num_filters=self.n_first_filters,
filter_size=(self.first_filter_length, 1),
stride=(1, 1),
nonlinearity=identity,
pad='same',
W=lasagne.init.HeNormal(gain='relu'))
# now spatconv
model = batch_norm(Conv2DLayer(
model,
num_filters=self.n_first_filters,
filter_size=(1, self.in_chans),
stride=(1, 1),
nonlinearity=self.nonlinearity,
pad=0,
W=lasagne.init.HeNormal(gain='relu')),
epsilon=self.batch_norm_epsilon,
alpha=self.batch_norm_alpha)
else:
model = batch_norm(Conv2DLayer(
model,
num_filters=self.n_first_filters,
filter_size=(self.first_filter_length, 1),
stride=(1, 1),
nonlinearity=self.nonlinearity,
pad='same',
W=lasagne.init.HeNormal(gain='relu')),
epsilon=self.batch_norm_epsilon,
alpha=self.batch_norm_alpha)
for _ in range(self.n_layers_per_block):
model = resnet_residual_block(model)
model = resnet_residual_block(model,
increase_units_factor=2,
half_time=True)
for _ in range(1, self.n_layers_per_block):
model = resnet_residual_block(model)
model = resnet_residual_block(model,
increase_units_factor=1.5,
half_time=True)
for _ in range(1, self.n_layers_per_block):
model = resnet_residual_block(model)
model = resnet_residual_block(model, half_time=True)
for _ in range(1, self.n_layers_per_block):
model = resnet_residual_block(model)
model = resnet_residual_block(model, half_time=True)
for _ in range(1, self.n_layers_per_block):
model = resnet_residual_block(model)
model = resnet_residual_block(model, half_time=True)
for _ in range(1, self.n_layers_per_block):
model = resnet_residual_block(model)
model = resnet_residual_block(model, half_time=True)
if self.drop_before_pool:
model = DropoutLayer(model, p=0.5)
# Replacement for global mean pooling
if self.final_aggregator == 'pool':
model = Pool2DLayer(model,
pool_size=(self.final_pool_length, 1),
stride=(1, 1),
mode='average_exc_pad')
model = Conv2DLayer(model,
filter_size=(1, 1),
num_filters=4,
W=lasagne.init.HeNormal(),
nonlinearity=identity)
elif self.final_aggregator == 'conv':
model = Conv2DLayer(model,
filter_size=(self.final_pool_length, 1),
num_filters=4,
W=lasagne.init.HeNormal(),
nonlinearity=identity)
else:
raise ValueError("Unknown final aggregator {:s}".format(
self.final_aggregator))
model = FinalReshapeLayer(model)
model = NonlinearityLayer(model, nonlinearity=self.final_nonlin)
model = set_survival_probs_to_linear_decay(model, self.survival_prob)
return lasagne.layers.get_all_layers(model)
def test_raw_net_trial_based_and_continuous():
softmax_rng = RandomState(3094839048)
orig_softmax_weights = softmax_rng.randn(4,20,54,1).astype(theano.config.floatX) * 0.01
lasagne.random.set_rng(RandomState(23903823))
input_var = T.tensor4('inputs')
epo_network = lasagne.layers.InputLayer(shape=[None,22,1200,1],
input_var=input_var)
# we have to switch channel dimension to height axis to not squash/convolve them away
epo_network = lasagne.layers.DimshuffleLayer(epo_network, pattern=(0,3,2,1))
epo_network = lasagne.layers.Conv2DLayer(epo_network, num_filters=20,filter_size=[30, 1], nonlinearity=identity)
epo_network = lasagne.layers.DropoutLayer(epo_network, p=0.5)
epo_network = lasagne.layers.Conv2DLayer(epo_network, num_filters=20,filter_size=[1, 22], nonlinearity=T.sqr)
epo_network = SumPool2dLayer(epo_network, pool_size=(100,1), stride=(20,1), mode='average_exc_pad')
epo_network = lasagne.layers.NonlinearityLayer(epo_network, nonlinearity=safe_log)
epo_network = lasagne.layers.DropoutLayer(epo_network, p=0.5)
epo_network = lasagne.layers.DenseLayer(epo_network, num_units=4,nonlinearity=softmax,
W=orig_softmax_weights.reshape(4,-1).T)
preds = lasagne.layers.get_output(epo_network, deterministic=True)
pred_func = theano.function([input_var], preds)
n_trials = 20
n_samples = 1200 + n_trials - 1
rng = RandomState(343434216)
orig_inputs = rng.randn(1, get_input_shape(epo_network)[1],n_samples,1).astype(
theano.config.floatX)
# reshape to 2000 trials
trialwise_inputs = [orig_inputs[:,:,start_i:start_i+1200] for start_i in range(n_trials)]
trialwise_inputs = np.array(trialwise_inputs)[:,0]
lasagne.random.set_rng(RandomState(23903823))
input_var = T.tensor4('inputs')
cnt_network = lasagne.layers.InputLayer(shape=[None,22,n_samples,1],
input_var=input_var)
# we have to switch channel dimension to height axis to not squash/convolve them away
cnt_network = lasagne.layers.DimshuffleLayer(cnt_network, pattern=(0,3,2,1))
cnt_network = lasagne.layers.Conv2DLayer(cnt_network, num_filters=20,
filter_size=[30, 1], nonlinearity=identity)
cnt_network = lasagne.layers.DropoutLayer(cnt_network, p=0.5)
cnt_network = lasagne.layers.Conv2DLayer(cnt_network, num_filters=20,
filter_size=[1, 22], nonlinearity=T.sqr)
cnt_network = SumPool2dLayer(cnt_network, pool_size=(100,1),
stride=(1,1), mode='average_exc_pad')
cnt_network = lasagne.layers.NonlinearityLayer(cnt_network,
nonlinearity=safe_log)
cnt_network = StrideReshapeLayer(cnt_network, n_stride=20)
cnt_network = lasagne.layers.DropoutLayer(cnt_network, p=0.5)
cnt_network = lasagne.layers.Conv2DLayer(cnt_network, num_filters=4,
filter_size=[54, 1], W=orig_softmax_weights[:,:,::-1,:], stride=(1,1),
nonlinearity=lasagne.nonlinearities.identity)
cnt_network = FinalReshapeLayer(cnt_network)
cnt_network = lasagne.layers.NonlinearityLayer(cnt_network,
nonlinearity=softmax)
preds_cnt = lasagne.layers.get_output(cnt_network, deterministic=True)
pred_cnt_func = theano.function([input_var], preds_cnt)
results = []
batch_size = 5
for i_trial in xrange(0,len(trialwise_inputs),batch_size):
res = pred_func(trialwise_inputs[i_trial:min(len(trialwise_inputs),
i_trial+batch_size)])
results.append(res)
results = np.array(results).squeeze()
res_cnt = pred_cnt_func(orig_inputs)
reshaped_results = np.concatenate(results)
assert np.allclose(reshaped_results, res_cnt[:n_trials])
def test_stride_reshape_layer():
input_var = T.tensor4('inputs').astype(theano.config.floatX)
network = lasagne.layers.InputLayer(shape=[None,1,15,1], input_var=input_var)
network = lasagne.layers.Conv2DLayer(network, num_filters=1,filter_size=[3, 1],
W=lasagne.init.Constant(1), stride=(1,1))
network = StrideReshapeLayer(network, n_stride=2, invalid_fill_value=np.nan)
network = lasagne.layers.Conv2DLayer(network, num_filters=1,filter_size=[2, 1],
W=lasagne.init.Constant(1), stride=(1,1))
network = StrideReshapeLayer(network, n_stride=2, invalid_fill_value=np.nan)
network = lasagne.layers.Conv2DLayer(network, num_filters=4, filter_size=[2, 1],
W=to_4d_time_array([[1,1], [-1,-1], [0.1,0.1], [-0.1,-0.1]]), stride=(1,1),
nonlinearity=lasagne.nonlinearities.identity)
network = FinalReshapeLayer(network, remove_invalids=False)
preds_cnt = lasagne.layers.get_output(lasagne.layers.get_all_layers(network)[1:])
pred_cnt_func = theano.function([input_var], preds_cnt)
layer_activations = pred_cnt_func(to_4d_time_array([range(1,16), range(16,31)]))
assert equal_without_nans(np.array([[[[ 6.], [ 9.], [ 12.], [ 15.], [ 18.], [ 21.],
[ 24.], [ 27.], [ 30.], [ 33.], [ 36.], [ 39.],
[ 42.]]],
[[[ 51.], [ 54.], [ 57.], [ 60.], [ 63.], [ 66.],
[ 69.], [ 72.], [75.], [ 78.], [ 81.], [ 84.],
[ 87.]]]],
dtype=np.float32),
layer_activations[0])
assert equal_without_nans(np.array(
[[[[ 6.], [ 12.], [ 18.], [ 24.], [ 30.], [ 36.], [ 42.]]],
[[[ 51.], [ 57.], [ 63.], [ 69.], [ 75.], [ 81.], [ 87.]]],
[[[ 9.], [ 15.], [ 21.], [ 27.], [ 33.], [ 39.], [ np.nan]]],
[[[ 54.], [ 60.], [ 66.], [ 72.], [ 78.], [ 84.], [ np.nan]]]],
dtype=np.float32),
layer_activations[1])
assert equal_without_nans(np.array([[[[ 18.], [ 30.], [ 42.], [ 54.], [ 66.], [ 78.]]],
[[[ 108.], [ 120.], [ 132.], [ 144.], [ 156.], [ 168.]]],
[[[ 24.], [ 36.], [ 48.], [ 60.], [ 72.], [ np.nan]]],
[[[ 114.], [ 126.], [ 138.], [ 150.], [ 162.], [ np.nan]]]],
dtype=np.float32),
layer_activations[2])
assert equal_without_nans(np.array([[[[ 18.], [ 42.], [ 66.]]],
[[[ 108.], [ 132.], [ 156.]]],
[[[ 24.], [ 48.], [ 72.]]],
[[[ 114.], [ 138.], [ 162.]]],
[[[ 30.], [ 54.], [ 78.]]],
[[[ 120.], [ 144.], [ 168.]]],
[[[ 36.], [ 60.], [ np.nan]]],
[[[ 126.], [ 150.], [ np.nan]]]],
dtype=np.float32),
layer_activations[3])
assert allclose_without_nans(np.array(
[[[[ 60. ], [ 108. ]], [[ -60. ], [-108. ]],
[[ 6.00000048], [ 10.80000019]], [[ -6.00000048], [ -10.80000019]]],
[[[ 240. ], [ 288. ]], [[-240. ], [-288. ]],
[[ 24. ], [ 28.80000114]], [[ -24. ], [ -28.80000114]]],
[[[ 72. ], [ 120. ]], [[ -72. ], [-120. ]],
[[ 7.20000029], [ 12. ]], [[ -7.20000029], [ -12. ]]],
[[[ 252. ], [ 300. ]], [[-252. ], [-300. ]],
[[ 25.20000076], [ 30.00000191]], [[ -25.20000076], [ -30.00000191]]],
[[[ 84. ], [ 132. ]], [[ -84. ], [-132. ]],
[[ 8.40000057], [ 13.19999981]], [[ -8.40000057], [ -13.19999981]]],
[[[ 264. ], [ 312. ]], [[-264. ], [-312. ]],
[[ 26.40000153], [ 31.20000076]], [[ -26.40000153], [ -31.20000076]]],
[[[ 96. ], [ np.nan]], [[ -96. ], [ np.nan]],
[[ 9.60000038], [ np.nan]], [[ -9.60000038], [ np.nan]]],
[[[ 276. ], [ np.nan]], [[-276. ], [ np.nan]],
[[ 27.60000038], [ np.nan]], [[ -27.60000038], [ np.nan]]]],
dtype=np.float32),
layer_activations[4])
assert allclose_without_nans(np.array(
[[ 60. , -60. , 6.0, -6.],
[ 72. , -72. , 7.2, -7.2],
[ 84. , -84. , 8.4, -8.4],
[ 96. , -96. , 9.6, -9.6],
[ 108. , -108. , 10.8, -10.8],
[ 120. , -120. , 12. , -12. ],
[ 132. , -132. , 13.2, -13.2],
[ np.nan, np.nan, np.nan, np.nan],
[ 240. , -240. , 24. , -24. ],
[ 252. , -252. , 25.2, -25.2],
[ 264. , -264. , 26.4, -26.4],
[ 276. , -276. , 27.6, -27.6],
[ 288. , -288. , 28.8, -28.8],
[ 300. , -300. , 30.0, -30.0],
[ 312. , -312. , 31.2, -31.2],
[ np.nan, np.nan, np.nan, np.nan]],
dtype=np.float32),
layer_activations[5])
def test_stride_reshape_layer_with_nonempty_3rd_dim():
input_var = T.tensor4('inputs').astype(theano.config.floatX)
network = lasagne.layers.InputLayer(shape=[None,1,15,2], input_var=input_var)
network = lasagne.layers.Conv2DLayer(network, num_filters=1,filter_size=[3, 1],
W=lasagne.init.Constant(1), stride=(1,1))
network = StrideReshapeLayer(network, n_stride=2, invalid_fill_value=np.nan)
network = lasagne.layers.Conv2DLayer(network, num_filters=1,filter_size=[2, 1],
W=lasagne.init.Constant(1), stride=(1,1))
network = StrideReshapeLayer(network, n_stride=2, invalid_fill_value=np.nan)
network = lasagne.layers.Conv2DLayer(network, num_filters=4, filter_size=[2, 1],
W=to_4d_time_array([[1,1], [-1,-1], [0.1,0.1], [-0.1,-0.1]]), stride=(1,1),
nonlinearity=lasagne.nonlinearities.identity)
network = lasagne.layers.Conv2DLayer(network, num_filters=3, filter_size=[1, 2],
W=np.array([[[[1,1]],[[0.5,0.5]], [[1,1]], [[0.5,0.5]]],
[[[-1,-1]],[[-0.5,-0.5]], [[-1,-1]], [[0,0]]],
[[[0,0]],[[0,0]], [[-1,1]], [[0,0]]]], dtype=np.float32), stride=(1,1),
nonlinearity=lasagne.nonlinearities.identity)
network = FinalReshapeLayer(network, remove_invalids=False)
preds_cnt = lasagne.layers.get_output(lasagne.layers.get_all_layers(network)[1:])
pred_cnt_func = theano.function([input_var], preds_cnt)
inputs= np.array([[range(1,16), range(101,116)],
[range(16,31), range(116,131)]]).astype(np.float32)
inputs = inputs.swapaxes(1,2)[:,np.newaxis,:,:]
layer_activations = pred_cnt_func(inputs)
assert equal_without_nans(np.array([[[[ 6., 306], [ 9., 309], [ 12., 312], [ 15.,315], [ 18.,318], [ 21.,321],
[ 24.,324], [ 27.,327], [ 30.,330], [ 33.,333], [ 36.,336], [ 39.,339],
[ 42.,342]]],
[[[ 51.,351], [ 54.,354], [ 57.,357], [ 60.,360], [ 63.,363], [ 66.,366],
[ 69.,369], [ 72.,372], [75.,375], [ 78.,378], [ 81.,381], [ 84.,384],
[ 87.,387]]]],
dtype=np.float32),
layer_activations[0])
assert equal_without_nans(np.array(
[[[[ 6., 306], [ 12.,312], [ 18.,318], [ 24.,324], [ 30.,330], [ 36.,336], [ 42.,342]]],
[[[ 51., 351], [ 57., 357], [ 63., 363], [ 69., 369], [ 75., 375], [ 81., 381], [ 87., 387]]],
[[[ 9., 309], [ 15., 315], [ 21., 321], [ 27., 327], [ 33., 333], [ 39., 339], [ np.nan, np.nan]]],
[[[ 54., 354], [ 60., 360], [ 66., 366], [ 72., 372], [ 78., 378], [ 84., 384], [ np.nan, np.nan]]]],
dtype=np.float32),
layer_activations[1])
assert equal_without_nans(np.array([[[[ 18.,618.], [ 30., 630], [ 42., 642], [ 54., 654], [ 66., 666], [ 78., 678]]],
[[[ 108., 708], [ 120., 720], [ 132., 732], [ 144., 744], [ 156., 756], [ 168., 768]]],
[[[ 24., 624], [ 36., 636], [ 48., 648], [ 60., 660], [ 72., 672], [ np.nan, np.nan]]],
[[[ 114., 714], [ 126., 726], [ 138., 738], [ 150., 750], [ 162., 762], [ np.nan, np.nan]]]],
dtype=np.float32),
layer_activations[2])
assert equal_without_nans(np.array([[[[ 18., 618], [ 42.,642], [ 66.,666]]],
[[[ 108.,708], [ 132.,732], [ 156.,756]]],
[[[ 24.,624], [ 48.,648], [ 72.,672]]],
[[[ 114.,714], [ 138.,738], [ 162.,762]]],
[[[ 30.,630], [ 54.,654], [ 78.,678]]],
[[[ 120.,720], [ 144.,744], [ 168.,768]]],
[[[ 36.,636], [ 60.,660], [ np.nan, np.nan]]],
[[[ 126., 726], [ 150., 750], [ np.nan, np.nan]]]],
dtype=np.float32),
layer_activations[3])
assert allclose_without_nans(np.array(
[[[[ 60. ,1260], [ 108. , 1308]], [[ -60. , -1260], [-108. ,-1308]],
[[ 6.00000048, 126], [ 10.80000019, 130.80000305]], [[ -6.00000048, -126], [ -10.80000019, -130.80000305]]],
[[[ 240. , 1440. ], [ 288. , 1488. ]], [[ -240. , -1440. ],
[ -288. , -1488. ]], [[ 24. , 144. ], [ 28.79999924, 148.80000305]],
[[ -24. , -144. ], [ -28.79999924, -148.80000305]]],
[[[ 72. , 1272. ], [ 120. , 1320. ]], [[ -72. , -1272. ],
[ -120. , -1320. ]], [[ 7.20000029, 127.20000458], [ 12. , 132. ]],
[[ -7.20000029, -127.20000458], [ -12. , -132. ]]],
[[[ 252. , 1452. ], [ 300. , 1500. ]], [[ -252. , -1452. ],
[ -300. , -1500. ]], [[ 25.20000076, 145.19999695], [ 30. , 150. ]],
[[ -25.20000076, -145.19999695], [ -30. , -150. ]]],
[[[ 84. , 1284. ], [ 132. , 1332. ]], [[ -84. , -1284. ],
[ -132. , -1332. ]], [[ 8.39999962, 128.3999939 ], [ 13.19999981, 133.19999695]],
[[ -8.39999962, -128.3999939 ], [ -13.19999981, -133.19999695]]],
[[[ 264. , 1464. ], [ 312. , 1512. ]], [[ -264. , -1464. ],
[ -312. , -1512. ]], [[ 26.39999962, 146.3999939 ], [ 31.20000076, 151.19999695]],
[[ -26.39999962, -146.3999939 ], [ -31.20000076, -151.19999695]]],
[[[ 96. , 1296. ], [np.nan, np.nan]], [[ -96. , -1296. ],
[np.nan, np.nan]],
[[ 9.60000038, 129.6000061 ], [np.nan, np.nan]], [[ -9.60000038, -129.6000061 ],
[np.nan, np.nan]]],
[[[ 276. , 1476. ], [np.nan, np.nan]],
[[ -276. , -1476. ], [np.nan, np.nan]],
[[ 27.60000038, 147.6000061 ], [np.nan, np.nan]],
[[ -27.60000038, -147.6000061 ], [np.nan, np.nan]]]], dtype=np.float32),
layer_activations[4])
expected_5 = [np.sum(layer_activations[4][:,[0,2]] * 1 + layer_activations[4][:,[1,3]] * 0.5, axis=(1,3)),
np.sum(layer_activations[4][:,[0,2]] * (-1), axis=(1,3)) +
np.sum(layer_activations[4][:,[1]] * (-0.5), axis=(1,3)),
layer_activations[4][:,2,:,0] - layer_activations[4][:,2,:,1]]
expected_5 = np.array(expected_5).swapaxes(0,1)[:,:,:,np.newaxis]
assert allclose_without_nans(expected_5, layer_activations[5])
assert allclose_without_nans(layer_activations[6][:,0],
np.concatenate((np.linspace(726,805.2, 7), [np.nan], np.linspace(924, 1003.2, 7), [np.nan])))
assert allclose_without_nans(layer_activations[6][:,1],
np.concatenate((np.linspace(-792,-878.4, 7), [np.nan], np.linspace(-1008, -1094.4, 7), [np.nan])))
assert allclose_without_nans(layer_activations[6][:,2],
np.concatenate(([-120] * 7, [np.nan], [-120] * 7, [np.nan],)))
for elem in layer_activations[6].flatten():
assert np.isnan(elem) or elem in layer_activations[5]
assert np.sum(np.isnan(layer_activations[5])) == np.sum(np.isnan(layer_activations[6]))