def test_upsample_cam_1d(self):
"""Ensures correct output from _upsample_cam.
In this case the CAM (class-activation map) is 1-D.
"""
this_matrix = gradcam._upsample_cam(
class_activation_matrix=CLASS_ACTIV_MATRIX_COARSE_1D + 0.,
new_dimensions=numpy.array(CLASS_ACTIV_MATRIX_1D.shape, dtype=int))
self.assertTrue(
numpy.allclose(this_matrix, CLASS_ACTIV_MATRIX_1D, atol=TOLERANCE))
def run_gradcam(model_object,
predictor_matrix,
activation_layer_name,
vector_output_layer_name,
output_neuron_indices,
ideal_activation=DEFAULT_IDEAL_ACTIVATION):
"""Runs the Grad-CAM algorithm.
H = number of heights
P = number of predictor variables
:param model_object: Trained model (instance of `keras.models.Model` or
`keras.models.Sequential`).
:param predictor_matrix: H-by-P numpy array of predictor values.
:param activation_layer_name: See doc for `check_metadata`.
:param vector_output_layer_name: Same.
:param output_neuron_indices: Same.
:param ideal_activation: same.
:return: class_activations: length-H numpy array of class activations.
"""
# TODO(thunderhoser): Eventually make this work for dense (scalar) output
# layers as well.
# Check input args.
error_checking.assert_is_numpy_array_without_nan(predictor_matrix)
error_checking.assert_is_numpy_array(predictor_matrix, num_dimensions=2)
predictor_matrix = numpy.expand_dims(predictor_matrix, axis=0)
check_metadata(activation_layer_name=activation_layer_name,
vector_output_layer_name=vector_output_layer_name,
output_neuron_indices=output_neuron_indices,
ideal_activation=ideal_activation)
# Set up loss function.
output_tensor = model_object.get_layer(
name=vector_output_layer_name).output[:, output_neuron_indices[0],
output_neuron_indices[1]]
# TODO(thunderhoser): Is this right?
# loss_tensor = (output_tensor - ideal_activation) ** 2
loss_tensor = output_tensor
# Set up gradient function.
layer_activation_tensor = (model_object.get_layer(
name=activation_layer_name).output)
gradient_tensor = (K.gradients(loss_tensor, [layer_activation_tensor])[0])
gradient_tensor = _normalize_tensor(gradient_tensor)
gradient_function = K.function([model_object.input],
[layer_activation_tensor, gradient_tensor])
# Evaluate gradient function.
layer_activation_matrix, gradient_matrix = gradient_function(
[predictor_matrix])
layer_activation_matrix = layer_activation_matrix[0, ...]
gradient_matrix = gradient_matrix[0, ...]
# Compute class-activation map in activation layer's space.
mean_weight_by_filter = numpy.mean(gradient_matrix, axis=0)
class_activations = numpy.full(layer_activation_matrix.shape[0], 0.)
num_filters = len(mean_weight_by_filter)
for k in range(num_filters):
class_activations += (mean_weight_by_filter[k] *
layer_activation_matrix[:, k])
num_input_heights = predictor_matrix.shape[1]
class_activation_matrix = gradcam_utils._upsample_cam(
class_activation_matrix=class_activations,
new_dimensions=numpy.array([num_input_heights], dtype=int))
return numpy.maximum(class_activation_matrix, 0.)