def test_ObserveEmbeddingConvNet3D4C(self):
batch_size = 32
channels = 3
output_dim = 128
input_batch_shape = [batch_size, channels, 16, 16, 16]
output_batch_shape_correct = [batch_size, output_dim]
input_non_batch_shape = [channels, 16, 16, 16]
output_non_batch_shape_correct = [1, output_dim]
input_batch = Variable(util.Tensor(torch.Size(input_batch_shape)))
input_non_batch = Variable(
util.Tensor(torch.Size(input_non_batch_shape)))
nn = ObserveEmbeddingConvNet3D4C(
input_example_non_batch=input_non_batch, output_dim=output_dim)
nn.configure()
output_batch_shape = list(nn.forward(input_batch).size())
output_non_batch_shape = list(
nn.forward(input_non_batch.unsqueeze(0)).size())
util.debug('batch_size', 'channels', 'output_dim', 'input_batch_shape',
'output_batch_shape', 'output_batch_shape_correct',
'input_non_batch_shape', 'output_non_batch_shape',
'output_non_batch_shape_correct')
self.assertEqual(output_batch_shape, output_batch_shape_correct)
self.assertEqual(output_non_batch_shape,
output_non_batch_shape_correct)
def test_dist_empirical(self):
values = Variable(util.Tensor([1, 2, 3]))
log_weights = Variable(util.Tensor([1, 2, 3]))
dist_mean_correct = 2.5752103328704834
dist_stddev_correct = 0.6514633893966675
dist_expectation_sin_correct = 0.3921678960323334
dist_map_sin_mean_correct = 0.3921678960323334
dist_min_correct = 1
dist_max_correct = 3
# dist_sample_shape_correct = []
dist = Empirical(values, log_weights)
dist_empirical = Empirical([dist.sample() for i in range(empirical_samples)])
dist_mean = float(dist.mean)
dist_mean_empirical = float(dist_empirical.mean)
dist_stddev = float(dist.stddev)
dist_stddev_empirical = float(dist_empirical.stddev)
dist_expectation_sin = float(dist.expectation(torch.sin))
dist_map_sin_mean = float(dist.map(torch.sin).mean)
dist_min = float(dist.min)
dist_max = float(dist.max)
dist_sample_shape = list(dist.sample().size())
util.debug('dist_mean', 'dist_mean_empirical', 'dist_mean_correct', 'dist_stddev', 'dist_stddev_empirical', 'dist_stddev_correct', 'dist_expectation_sin', 'dist_expectation_sin_correct', 'dist_map_sin_mean', 'dist_map_sin_mean_correct', 'dist_min', 'dist_min_correct', 'dist_max', 'dist_max_correct')
# self.assertEqual(dist_sample_shape, dist_sample_shape_correct)
self.assertAlmostEqual(dist_mean, dist_mean_correct, places=1)
self.assertAlmostEqual(dist_mean_empirical, dist_mean_correct, places=1)
self.assertAlmostEqual(dist_stddev, dist_stddev_correct, places=1)
self.assertAlmostEqual(dist_stddev_empirical, dist_stddev_correct, places=1)
self.assertAlmostEqual(dist_expectation_sin, dist_expectation_sin_correct, places=1)
self.assertAlmostEqual(dist_map_sin_mean, dist_map_sin_mean_correct, places=1)
self.assertAlmostEqual(dist_min, dist_min_correct, places=1)
self.assertAlmostEqual(dist_max, dist_max_correct, places=1)
def test_dist_empirical_save_load(self):
file_name = os.path.join(tempfile.mkdtemp(), str(uuid.uuid4()))
values = Variable(util.Tensor([1, 2, 3]))
log_weights = Variable(util.Tensor([1, 2, 3]))
dist_mean_correct = 2.5752103328704834
dist_stddev_correct = 0.6514633893966675
dist_expectation_sin_correct = 0.3921678960323334
dist_map_sin_mean_correct = 0.3921678960323334
dist_on_file = Empirical(values, log_weights)
dist_on_file.save(file_name)
dist = Distribution.load(file_name)
os.remove(file_name)
dist_empirical = Empirical([dist.sample() for i in range(empirical_samples)])
dist_mean = float(dist.mean)
dist_mean_empirical = float(dist_empirical.mean)
dist_stddev = float(dist.stddev)
dist_stddev_empirical = float(dist_empirical.stddev)
dist_expectation_sin = float(dist.expectation(torch.sin))
dist_map_sin_mean = float(dist.map(torch.sin).mean)
util.debug('file_name', 'dist_mean', 'dist_mean_empirical', 'dist_mean_correct', 'dist_stddev', 'dist_stddev_empirical', 'dist_stddev_correct', 'dist_expectation_sin', 'dist_expectation_sin_correct', 'dist_map_sin_mean', 'dist_map_sin_mean_correct')
self.assertAlmostEqual(dist_mean, dist_mean_correct, places=1)
self.assertAlmostEqual(dist_mean_empirical, dist_mean_correct, places=1)
self.assertAlmostEqual(dist_stddev, dist_stddev_correct, places=1)
self.assertAlmostEqual(dist_stddev_empirical, dist_stddev_correct, places=1)
self.assertAlmostEqual(dist_expectation_sin, dist_expectation_sin_correct, places=1)
self.assertAlmostEqual(dist_map_sin_mean, dist_map_sin_mean_correct, places=1)
def NDArray_to_Tensor(ndarray):
if ndarray is None:
# util.log_warning('NDArray_to_Tensor: empty NDArray received, returning empty tensor')
return util.Tensor()
else:
b = ndarray._tab.Bytes
o = flatbuffers.number_types.UOffsetTFlags.py_type(
ndarray._tab.Offset(4))
offset = ndarray._tab.Vector(o) if o != 0 else 0
length = ndarray.DataLength()
data_np = np.frombuffer(b,
offset=offset,
dtype=np.dtype('float64'),
count=length)
o = flatbuffers.number_types.UOffsetTFlags.py_type(
ndarray._tab.Offset(6))
offset = ndarray._tab.Vector(o) if o != 0 else 0
length = ndarray.ShapeLength()
shape_np = np.frombuffer(b,
offset=offset,
dtype=np.dtype('int32'),
count=length)
# print('data:', data_np)
# print('shape', shape_np)
data = data_np.reshape(shape_np)
return util.Tensor(data)