def test_formulate_smt_constraints_convolution_layer(self):
with self.test_session():
# Temporary graphs should be created inside a session. Notice multiple
# graphs are being created in this particular code. So, if each graph
# isn't created inside a separate session, the tensor names will have
# unwanted integer suffices, which then would cause problems while
# accessing tensors by name.
_create_temporary_tf_graph_cnn(self.test_model_path)
image_edge_length = 4
image_channels = 3
# The 1st convolution layer has 48 neurons.
top_k = np.random.randint(low=1, high=48)
image = np.ones((image_edge_length, image_edge_length, image_channels))
tensor_names = {
'input': 'conv2d_input:0',
'first_layer': 'conv2d/BiasAdd:0',
'first_layer_relu': 'conv2d/Relu:0',
'logits': 'dense/BiasAdd:0',
'softmax': 'dense/Softmax:0',
'weights_layer_1': 'conv2d/Conv2D/ReadVariableOp:0',
'biases_layer_1': 'conv2d/bias/Read/ReadVariableOp:0'}
session = utils.restore_model(self.test_model_path)
cnn_predictions = session.run(
tensor_names,
feed_dict={
tensor_names['input']: image.reshape(
(1, image_edge_length, image_edge_length, image_channels))})
z3_mask = [_get_z3_var(index=i) for i in range(image_edge_length ** 2)]
first_layer_activations = masking._reorder(masking._remove_batch_axis(
cnn_predictions['first_layer'])).reshape(-1)
masked_input = masking._encode_input(
image=image, z3_mask=z3_mask, window_size=1)
z3_optimizer = masking._formulate_smt_constraints_convolution_layer(
z3_optimizer=utils.ImageOptimizer(
z3_mask=z3_mask,
window_size=1,
edge_length=image_edge_length),
kernels=masking._reorder(cnn_predictions['weights_layer_1']),
biases=cnn_predictions['biases_layer_1'],
chosen_indices=first_layer_activations.argsort()[-top_k:],
conv_activations=first_layer_activations,
input_activation_maps=masked_input,
output_activation_map_shape=(image_edge_length, image_edge_length),
strides=1,
padding=(0, 1),
gamma=0.5)
mask, result = z3_optimizer.generate_mask()
self.assertEqual(result, 'sat')
self.assertEqual(mask.shape, (image_edge_length, image_edge_length))
session.close()
def test_formulate_smt_constraints_convolution_layer_text(self):
with self.test_session():
# Temporary graphs should be created inside a session. Notice multiple
# graphs are being created in this particular code. So, if each graph
# isn't created inside a separate session, the tensor names will have
# unwanted integer suffices, which then would cause problems while
# accessing tensors by name.
_create_temporary_tf_graph_text_cnn(self.test_model_path)
# The 1st convolution layer has 12 neurons.
image = np.ones(5)
tensor_names = {
'input': 'input_1:0',
'embedding': 'embedding/embedding_lookup/Identity_1:0',
'first_layer': 'conv1d/BiasAdd:0',
'first_layer_relu': 'conv1d/Relu:0',
'logits': 'dense/BiasAdd:0',
'softmax': 'dense/Sigmoid:0',
'weights_layer_1': 'conv1d/conv1d/ExpandDims_1:0',
'biases_layer_1': 'conv1d/BiasAdd/ReadVariableOp:0'
}
session = utils.restore_model(self.test_model_path)
cnn_predictions = session.run(
tensor_names,
feed_dict={tensor_names['input']: image.reshape(1, 5)})
text_embedding = masking._remove_batch_axis(
cnn_predictions['embedding'])
z3_mask = [
z3.Int('mask_%d' % i) for i in range(text_embedding.shape[0])
]
masked_input = []
for mask_bit, embedding_row in zip(z3_mask, text_embedding):
masked_input.append(
[z3.ToReal(mask_bit) * i for i in embedding_row])
first_layer_activations = masking._reorder(
masking._remove_batch_axis(
cnn_predictions['first_layer'])).reshape(-1)
z3_optimizer = masking._formulate_smt_constraints_convolution_layer(
z3_optimizer=utils.TextOptimizer(z3_mask=z3_mask),
kernels=masking._reshape_kernels(
kernels=cnn_predictions['weights_layer_1'],
model_type='text_cnn'),
biases=cnn_predictions['biases_layer_1'],
chosen_indices=first_layer_activations.argsort()[-5:],
conv_activations=first_layer_activations,
input_activation_maps=[masked_input],
output_activation_map_shape=masking._get_activation_map_shape(
activation_maps_shape=cnn_predictions['first_layer'].shape,
model_type='text_cnn'),
strides=1,
padding=(0, 0),
gamma=0.5)
mask, result = z3_optimizer.generate_mask()
self.assertEqual(result, 'sat')
self.assertEqual(mask.shape, (5, ))
session.close()
def test_encode_input(self, image_channels):
# Creates a random image and checks if the encoded image, after multiplying
# the mask bits (set to 1) is the same as the original image.
image_edge_length = 2
image = np.random.rand(image_edge_length, image_edge_length,
image_channels)
z3_var = _get_z3_var(index=0)
# encoded_image has dimensions
# (image_channels, image_edge_length, image_edge_length)
encoded_image = masking._encode_input(
image=image, z3_mask=[z3_var for _ in range(image_edge_length**2)])
solver = z3.Solver()
solver.add(z3_var == 1)
# Swap the axes of the image so that it has the same dimensions as the
# encoded image.
image = masking._reorder(image).reshape(-1)
encoded_image = utils.flatten_nested_lists(encoded_image)
for i in range(image_channels * image_edge_length**2):
solver.add(encoded_image[i] == image[i])
self.assertEqual(str(solver.check()), 'sat')
def test_reorder(self):
shape = tuple(np.random.randint(low=1, high=10, size=4))
self.assertEqual(
masking._reorder(np.ones(shape)).shape,
(shape[3], shape[0], shape[1], shape[2]))