def matrix_regression_data_prepare(**kwargs):
bufferCollection = buffers.BufferCollection()
bufferCollection.AddBuffer(buffers.X_FLOAT_DATA,
buffers.as_matrix_buffer(kwargs['x']))
if 'y' in kwargs:
bufferCollection.AddBuffer(buffers.YS,
buffers.as_matrix_buffer(kwargs['y']))
return bufferCollection
def construct_axis_aligned_forest(self):
# (0) X[0] > 2.2
# / \
# (1) X[1] > -5 (2) [0.7 0.1 0.2]
# / \
# (3) [0.3 0.3 0.4] (4) [0.3 0.6 0.1]
#
# (0) X[0] > 5.0
# / \
# (1) X[0] > 2.5 (2) [0.8 0.1 0.1]
# / \
# (3) [0.2 0.2 0.6] (4) [0.2 0.7 0.1]
path_1 = buffers.as_matrix_buffer(np.array([[1,2],[3,4],[-1,-1],[-1,-1],[-1,-1]], dtype=np.int32))
int_params_1 = buffers.as_matrix_buffer(np.array([[1,0],[1,1],[1,0],[1,0],[1,0]], dtype=np.int32))
float_params_1 = buffers.as_matrix_buffer(np.array([[2.2],[-5],[0],[0],[0]], dtype=np.float32))
ys_1 = buffers.as_matrix_buffer(np.array([[0,0,0],[0,0,0],[0.7,0.1,0.2],[0.3,0.3,0.4],[0.3,0.6,0.1]], dtype=np.float32))
depth_1 = buffers.as_vector_buffer(np.array([0, 1, 1, 2, 2], dtype=np.int32))
counts_1 = buffers.as_vector_buffer(np.array([5, 5, 5, 5, 5], dtype=np.float32))
tree_1 = forest_data.Tree(path_1, int_params_1, float_params_1, depth_1, counts_1, ys_1)
path_2 = buffers.as_matrix_buffer(np.array([[1,2],[3,4],[-1,-1],[-1,-1],[-1,-1]], dtype=np.int32))
int_params_2 = buffers.as_matrix_buffer(np.array([[1,0],[1,0],[1,0],[1,0],[1,0]], dtype=np.int32))
float_params_2 = buffers.as_matrix_buffer(np.array([[5.0],[2.5],[0],[0],[0]], dtype=np.float32))
ys_2 = buffers.as_matrix_buffer(np.array([[0,0,0],[0,0,0],[0.8,0.1,0.1],[0.2,0.2,0.6],[0.2,0.7,0.1]], dtype=np.float32))
depth_2 = buffers.as_vector_buffer(np.array([0, 1, 1, 2, 2], dtype=np.int32))
counts_2 = buffers.as_vector_buffer(np.array([5, 5, 5, 5, 5], dtype=np.float32))
tree_2 = forest_data.Tree(path_2, int_params_2, float_params_2, depth_2, counts_2, ys_2)
forest = forest_data.Forest([tree_1, tree_2])
return forest
def classify_body_pixels(depth, ground_labels, forest):
assert(depth.ndim == 2)
# setup test data
pixel_indices = to_indices(0, np.where(ground_labels != background))
buffer_collection = buffers.BufferCollection()
buffer_collection.AddInt32MatrixBuffer(buffers.PIXEL_INDICES, buffers.as_matrix_buffer(pixel_indices))
buffer_collection.AddFloat32Tensor3Buffer(buffers.DEPTH_IMAGES, buffers.as_tensor_buffer(depth))
# setup predictor
all_samples_step = pipeline.AllSamplesStep_i32f32i32(buffers.PIXEL_INDICES)
depth_delta_feature = image_features.ScaledDepthDeltaFeature_f32i32(all_samples_step.IndicesBufferId,
buffers.PIXEL_INDICES,
buffers.DEPTH_IMAGES)
combiner = classification.ClassProbabilityCombiner_f32(number_of_body_parts)
forest_predictor = predict.ScaledDepthDeltaClassificationPredictin_f32i32(forest, depth_delta_feature, combiner, all_samples_step)
# predict
yprobs_buffer = buffers.Float32MatrixBuffer()
forest_predictor.PredictYs(buffer_collection, yprobs_buffer)
# convert to image space
yprobs = buffers.as_numpy_array(yprobs_buffer)
(_, ydim) = yprobs.shape
m,n = depth.shape
img_yprobs = np.zeros((m,n), dtype=np.float32)
img_yprobs[ground_labels != background].shape
yprobs.max(axis=1).shape
img_yprobs[ground_labels != background] = yprobs.max(axis=1)
img_yhat = np.zeros((m,n), dtype=np.int32)
img_yhat[ground_labels != background] = np.argmax(yprobs, axis=1)
return img_yhat, img_yprobs
def classify_pixels(depth, forest):
assert(depth.ndim == 2)
# setup test data
pixel_indices = np.array( list(itertools.product( np.zeros(1), range(m), range(n) )), dtype=np.int32 )
buffer_collection = buffers.BufferCollection()
buffer_collection.AddInt32MatrixBuffer(buffers.PIXEL_INDICES, buffers.as_matrix_buffer(pixel_indices))
buffer_collection.AddFloat32Tensor3Buffer(buffers.DEPTH_IMAGES, buffers.as_tensor_buffer(depth))
# setup predictor
all_samples_step = pipeline.AllSamplesStep_i32f32i32(buffers.PIXEL_INDICES)
depth_delta_feature = image_features.ScaledDepthDeltaFeature_f32i32(all_samples_step.IndicesBufferId,
buffers.PIXEL_INDICES,
buffers.DEPTH_IMAGES)
combiner = classification.ClassProbabilityCombiner_f32(number_of_body_parts)
forest_predictor = predict.ScaledDepthDeltaClassificationPredictin_f32i32(forest, depth_delta_feature, combiner, all_samples_step)
# predict
yprobs_buffer = buffers.Float32MatrixBuffer()
forest_predictor.PredictYs(buffer_collection, yprobs_buffer)
# convert to image space
yprobs = buffers.as_numpy_array(yprobs_buffer)
(_, ydim) = yprobs.shape
img_yprobs = yprobs.reshape((m,n,ydim))
img_yhat = np.argmax(img_yprobs, axis=2)
return img_yhat, img_yprobs.max(axis=2)
def test_pickle(self):
# test tensor
for dtype in [np.float32, np.float64, np.int32, np.int64]:
array = np.array([[[3, 21, 1], [22, 1, 5]], [[2, 2, 2], [7, 7, 7]]], dtype=dtype)
b1 = buffers.as_tensor_buffer(array)
pickle.dump(b1, open("tmp.pkl", "wb"))
b2 = pickle.load(open("tmp.pkl", "rb"))
array2 = buffers.as_numpy_array(b2)
self.assertTrue((array == array2).all())
# test matrix
for dtype in [np.float32, np.float64, np.int32, np.int64]:
array = np.array([[3, 21, 1], [22, 1, 5], [2, 2, 2], [7, 7, 7]], dtype=dtype)
b1 = buffers.as_matrix_buffer(array)
pickle.dump(b1, open("tmp.pkl", "wb"))
b2 = pickle.load(open("tmp.pkl", "rb"))
array2 = buffers.as_numpy_array(b2)
self.assertTrue((array == array2).all())
# test vector
for dtype in [np.float32, np.float64, np.int32, np.int64]:
array = np.array([3, 21, 1], dtype=dtype)
b1 = buffers.as_vector_buffer(array)
pickle.dump(b1, open("tmp.pkl", "wb"))
b2 = pickle.load(open("tmp.pkl", "rb"))
array2 = buffers.as_numpy_array(b2)
self.assertTrue((array == array2).all())
def test_pickle(self):
# test tensor
for dtype in [np.float32, np.float64, np.int32, np.int64]:
array = np.array(
[[[3, 21, 1], [22, 1, 5]], [[2, 2, 2], [7, 7, 7]]],
dtype=dtype)
b1 = buffers.as_tensor_buffer(array)
pickle.dump(b1, open('tmp.pkl', 'wb'))
b2 = pickle.load(open('tmp.pkl', 'rb'))
array2 = buffers.as_numpy_array(b2)
self.assertTrue((array == array2).all())
# test matrix
for dtype in [np.float32, np.float64, np.int32, np.int64]:
array = np.array([[3, 21, 1], [22, 1, 5], [2, 2, 2], [7, 7, 7]],
dtype=dtype)
b1 = buffers.as_matrix_buffer(array)
pickle.dump(b1, open('tmp.pkl', 'wb'))
b2 = pickle.load(open('tmp.pkl', 'rb'))
array2 = buffers.as_numpy_array(b2)
self.assertTrue((array == array2).all())
# test vector
for dtype in [np.float32, np.float64, np.int32, np.int64]:
array = np.array([3, 21, 1], dtype=dtype)
b1 = buffers.as_vector_buffer(array)
pickle.dump(b1, open('tmp.pkl', 'wb'))
b2 = pickle.load(open('tmp.pkl', 'rb'))
array2 = buffers.as_numpy_array(b2)
self.assertTrue((array == array2).all())
def matrix_buffer_flatten_helper(self, X, buffer_type):
n = len(X.shape)
matrix_buffer = buffers.as_matrix_buffer(X)
assert isinstance(matrix_buffer, buffer_type)
X_back = buffers.as_numpy_array(matrix_buffer)
self.assertTrue((X == X_back.flatten()).all())
X_back = buffers.as_numpy_array(buffer=matrix_buffer, flatten=True)
self.assertTrue((X == X_back).all())
def construct_axis_aligned_forest(self):
# (0) X[0] > 2.2
# / \
# (1) X[1] > -5 (2) [0.7 0.1 0.2]
# / \
# (3) [0.3 0.3 0.4] (4) [0.3 0.6 0.1]
#
# (0) X[0] > 5.0
# / \
# (1) X[0] > 2.5 (2) [0.8 0.1 0.1]
# / \
# (3) [0.2 0.2 0.6] (4) [0.2 0.7 0.1]
path_1 = buffers.as_matrix_buffer(
np.array([[1, 2], [3, 4], [-1, -1], [-1, -1], [-1, -1]],
dtype=np.int32))
int_params_1 = buffers.as_matrix_buffer(
np.array([[1, 0], [1, 1], [1, 0], [1, 0], [1, 0]], dtype=np.int32))
float_params_1 = buffers.as_matrix_buffer(
np.array([[2.2], [-5], [0], [0], [0]], dtype=np.float32))
ys_1 = buffers.as_matrix_buffer(
np.array([[0, 0, 0], [0, 0, 0], [0.7, 0.1, 0.2], [0.3, 0.3, 0.4],
[0.3, 0.6, 0.1]],
dtype=np.float32))
depth_1 = buffers.as_vector_buffer(
np.array([0, 1, 1, 2, 2], dtype=np.int32))
counts_1 = buffers.as_vector_buffer(
np.array([5, 5, 5, 5, 5], dtype=np.float32))
tree_1 = forest_data.Tree(path_1, int_params_1, float_params_1,
depth_1, counts_1, ys_1)
path_2 = buffers.as_matrix_buffer(
np.array([[1, 2], [3, 4], [-1, -1], [-1, -1], [-1, -1]],
dtype=np.int32))
int_params_2 = buffers.as_matrix_buffer(
np.array([[1, 0], [1, 0], [1, 0], [1, 0], [1, 0]], dtype=np.int32))
float_params_2 = buffers.as_matrix_buffer(
np.array([[5.0], [2.5], [0], [0], [0]], dtype=np.float32))
ys_2 = buffers.as_matrix_buffer(
np.array([[0, 0, 0], [0, 0, 0], [0.8, 0.1, 0.1], [0.2, 0.2, 0.6],
[0.2, 0.7, 0.1]],
dtype=np.float32))
depth_2 = buffers.as_vector_buffer(
np.array([0, 1, 1, 2, 2], dtype=np.int32))
counts_2 = buffers.as_vector_buffer(
np.array([5, 5, 5, 5, 5], dtype=np.float32))
tree_2 = forest_data.Tree(path_2, int_params_2, float_params_2,
depth_2, counts_2, ys_2)
tree_2.GetExtraInfo().AddBuffer(
"first", np.array([3, 21, 1, 22, 1, 5], dtype=np.float32))
forest = forest_data.Forest([tree_1, tree_2])
return forest
def load_and_sample(pose_path, list_of_poses, number_of_pixels_per_image):
depths, labels = load_data(pose_path, list_of_poses)
depths_buffer = buffers.as_tensor_buffer(depths)
del depths
pixel_indices, pixel_labels = sample_pixels_from_images(labels, number_of_pixels_per_image)
del labels
pixel_indices_buffer = buffers.as_matrix_buffer(pixel_indices)
pixel_labels_buffer = buffers.as_vector_buffer(pixel_labels)
del pixel_indices
del pixel_labels
return depths_buffer, pixel_indices_buffer, pixel_labels_buffer
def load_training_data(numpy_filename):
f = open(numpy_filename, 'rb')
depths = np.load(f)
labels = np.load(f)
pixel_indices = np.load(f)
pixel_labels = np.load(f)
depths_buffer = buffers.as_tensor_buffer(depths)
del depths
del labels
pixel_indices_buffer = buffers.as_matrix_buffer(pixel_indices)
del pixel_indices
pixel_labels_buffer = buffers.as_vector_buffer(pixel_labels)
del pixel_labels
return depths_buffer, pixel_indices_buffer, pixel_labels_buffer
def predict_proba(self, x):
buffer_collection = buffers.BufferCollection()
buffer_collection.AddFloat32MatrixBuffer(buffers.X_FLOAT_DATA, buffers.as_matrix_buffer(x))
number_of_classes = self.forest_data.GetTree(0).GetYs().GetN()
all_samples_step = pipeline.AllSamplesStep_f32f32i32(buffers.X_FLOAT_DATA)
combiner = classification.ClassProbabilityCombiner_f32(number_of_classes)
matrix_feature = matrix_features.LinearFloat32MatrixFeature_f32i32(all_samples_step.IndicesBufferId,
buffers.X_FLOAT_DATA)
forest_predicter = predict.LinearMatrixClassificationPredictin_f32i32(forest_data, matrix_feature, combiner, all_samples_step)
result = buffers.Float32MatrixBuffer()
forest_predicter.PredictYs(bufferCollection, result)
return buffers.as_numpy_array(result)
forestStats = predictor.get_forest().GetForestStats()
forestStats.Print()
# For the rest of the passes use all of the data
start_index = 0
end_index = clipped_list_of_sample_counts[-1]
for pass_id in range(1, args.number_of_passes_through_data):
# Randomize the order
perm = buffers.as_vector_buffer(np.array(np.random.permutation(pixel_labels_buffer.GetN()), dtype=np.int32))
pixel_indices_buffer = pixel_indices_buffer.Slice(perm)
pixel_labels_buffer = pixel_labels_buffer.Slice(perm)
# Randomly offset scales
number_of_datapoints = pixel_indices_buffer.GetM()
offset_scales = np.array(np.random.uniform(0.8, 1.2, (number_of_datapoints, 2)), dtype=np.float32)
offset_scales_buffer = buffers.as_matrix_buffer(offset_scales)
predictor = forest_learner.fit(depth_images=depths_buffer,
pixel_indices=pixel_indices_buffer,
offset_scales=offset_scales_buffer,
classes=pixel_labels_buffer)
#pickle forest and data used for training
forest_pickle_filename = "%s/forest-%d-%d.pkl" % (online_run_folder, pass_id, end_index)
pickle.dump(predictor.get_forest(), gzip.open(forest_pickle_filename, 'wb'))
# Print forest stats
forestStats = predictor.get_forest().GetForestStats()
forestStats.Print()
def convert_matrix_buffer_both_directions_helper(self, X, buffer_type):
buf = buffers.as_matrix_buffer(X)
assert isinstance(buf, buffer_type)
X_back = buffers.as_numpy_array(buf)
self.assertTrue((X == X_back).all())
print "Processing %d - %d - %s - %s" % (pass_id, i, pose_filename, str(datetime.now()))
# Load single pose depth and class labels
depths = kinect_utils.load_depth_from_exr("%s%s.exr" % (args.pose_files_input_path, pose_filename))
labels = kinect_utils.load_labels_from_png("%s%s.png" % (args.pose_files_input_path, pose_filename))
pixel_indices, pixel_labels = kinect_utils.sample_pixels_from_image(labels[0,:,:], config.number_of_pixels_per_image)
# Randomly sample pixels and offset scales
(number_of_datapoints, _) = pixel_indices.shape
offset_scales = np.array(np.random.uniform(0.8, 1.2, (number_of_datapoints, 2)), dtype=np.float32)
datapoint_indices = np.array(np.arange(number_of_datapoints), dtype=np.int32)
# Package buffers for learner
bufferCollection = buffers.BufferCollection()
bufferCollection.AddFloat32Tensor3Buffer(buffers.DEPTH_IMAGES, buffers.as_tensor_buffer(depths))
bufferCollection.AddFloat32MatrixBuffer(buffers.OFFSET_SCALES, buffers.as_matrix_buffer(offset_scales))
bufferCollection.AddInt32MatrixBuffer(buffers.PIXEL_INDICES, buffers.as_matrix_buffer(pixel_indices))
bufferCollection.AddInt32VectorBuffer(buffers.CLASS_LABELS, buffers.as_vector_buffer(pixel_labels))
# Update learner
online_learner.Train(bufferCollection, buffers.Int32Vector(datapoint_indices))
#pickle forest and data used for training
if (i+1) % 1 == 0:
# if True:
forest_pickle_filename = "%s/forest-%d-%d.pkl" % (online_run_folder, pass_id, i+1)
pickle.dump(online_learner.GetForest(), gzip.open(forest_pickle_filename, 'wb'))
# Print forest stats
forestStats = online_learner.GetForest().GetForestStats()
forestStats.Print()
labels[0, :, :], config.number_of_pixels_per_image)
# Randomly sample pixels and offset scales
(number_of_datapoints, _) = pixel_indices.shape
offset_scales = np.array(np.random.uniform(
0.8, 1.2, (number_of_datapoints, 2)),
dtype=np.float32)
datapoint_indices = np.array(np.arange(number_of_datapoints),
dtype=np.int32)
# Package buffers for learner
bufferCollection = buffers.BufferCollection()
bufferCollection.AddFloat32Tensor3Buffer(
buffers.DEPTH_IMAGES, buffers.as_tensor_buffer(depths))
bufferCollection.AddFloat32MatrixBuffer(
buffers.OFFSET_SCALES, buffers.as_matrix_buffer(offset_scales))
bufferCollection.AddInt32MatrixBuffer(
buffers.PIXEL_INDICES, buffers.as_matrix_buffer(pixel_indices))
bufferCollection.AddInt32VectorBuffer(
buffers.CLASS_LABELS, buffers.as_vector_buffer(pixel_labels))
# Update learner
online_learner.Train(bufferCollection,
buffers.Int32Vector(datapoint_indices))
#pickle forest and data used for training
if (i + 1) % 1 == 0:
# if True:
forest_pickle_filename = "%s/forest-%d-%d.pkl" % (
online_run_folder, pass_id, i + 1)
pickle.dump(online_learner.GetForest(),
def matrix_regression_data_prepare(**kwargs):
bufferCollection = buffers.BufferCollection()
bufferCollection.AddBuffer(buffers.X_FLOAT_DATA, buffers.as_matrix_buffer(kwargs['x']))
if 'y' in kwargs:
bufferCollection.AddBuffer(buffers.YS, buffers.as_matrix_buffer(kwargs['y']))
return bufferCollection
end_index = clipped_list_of_sample_counts[-1]
for pass_id in range(1, args.number_of_passes_through_data):
# Randomize the order
perm = buffers.as_vector_buffer(
np.array(np.random.permutation(pixel_labels_buffer.GetN()),
dtype=np.int32))
pixel_indices_buffer = pixel_indices_buffer.Slice(perm)
pixel_labels_buffer = pixel_labels_buffer.Slice(perm)
# Randomly offset scales
number_of_datapoints = pixel_indices_buffer.GetM()
offset_scales = np.array(np.random.uniform(0.8, 1.2,
(number_of_datapoints, 2)),
dtype=np.float32)
offset_scales_buffer = buffers.as_matrix_buffer(offset_scales)
predictor = forest_learner.fit(depth_images=depths_buffer,
pixel_indices=pixel_indices_buffer,
offset_scales=offset_scales_buffer,
classes=pixel_labels_buffer)
#pickle forest and data used for training
forest_pickle_filename = "%s/forest-%d-%d.pkl" % (online_run_folder,
pass_id, end_index)
pickle.dump(predictor.get_forest(),
gzip.open(forest_pickle_filename, 'wb'))
# Print forest stats
forestStats = predictor.get_forest().GetForestStats()
forestStats.Print()
