From 2ea5843cbf0144b3c9a9bf8341aa941c3eca7618 Mon Sep 17 00:00:00 2001 From: dengkaipeng Date: Thu, 21 Feb 2019 22:24:34 +0800 Subject: [PATCH] add doc and test_layers. test=develop --- paddle/fluid/operators/spectral_norm_op.cc | 26 ++++- python/paddle/fluid/layers/nn.py | 96 +++++++++++-------- .../fluid/tests/unittests/test_layers.py | 13 +++ 3 files changed, 92 insertions(+), 43 deletions(-) diff --git a/paddle/fluid/operators/spectral_norm_op.cc b/paddle/fluid/operators/spectral_norm_op.cc index 56856c45b47..0d43e65c86c 100644 --- a/paddle/fluid/operators/spectral_norm_op.cc +++ b/paddle/fluid/operators/spectral_norm_op.cc @@ -109,10 +109,32 @@ class SpectralNormOpMaker : public framework::OpProtoAndCheckerMaker { .SetDefault(1e-12); AddComment(R"DOC( - This operator samples input X to given output shape by using specified + This layer calculate the spectral normalize value of weight of + fc, conv1d, conv2d, conv3d layers which should be 2-D, 3-D, 4-D, 5-D + tensor. - + Spectral normalization stabilizes the training of critis in GANs + (Generative Adversarial Networks). This layers rescaling weight tensor + wiht spectral normalize value. + For spectral normalization calculations, we rescaling weight + tensor with \sigma, while \sigma{\mathbf{W}} is + + \sigma(\mathbf{W}) = \max_{\mathbf{h}: \mathbf{h} \ne 0} \dfrac{\|\mathbf{W} \mathbf{h}\|_2}{\|\mathbf{h}\|_2} + + We calculate \sigma{\mathbf{W}} through power iterations as + + \mathbf{v} = \mathbf{W}^{T} \mathbf{u} + \mathbf{v} = \frac{\mathbf{v}}{\|\mathbf{v}\|_2} + \mathbf{u} = \mathbf{W}^{T} \mathbf{v} + \mathbf{u} = \frac{\mathbf{u}}{\|\mathbf{u}\|_2} + + And \sigma should be + + \sigma{\mathbf{W}} = \mathbf{u}^{T} \mathbf{W} \mathbf{v} + + For details of spectral normalization, please refer to paper: + `Spectral Normalization `_ . )DOC"); } }; diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 12755c30aea..d243f78325c 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -3348,28 +3348,42 @@ def group_norm(input, @templatedoc() -def spectral_norm(weight, - dim=0, - power_iters=1, - eps=1e-12, - u_attr=None, - v_attr=None, - name=None): +def spectral_norm(weight, dim=0, power_iters=1, eps=1e-12, name=None): """ **Spectral Normalization Layer** + This layer calculate the spectral normalize value of weight parameters of + fc, conv1d, conv2d, conv3d layers which should be 2-D, 3-D, 4-D, 5-D + Parameters. Calculations are showed as followings. + + .. code-block:: text + + Step 1: + Generate vector u in shape of [h], and v in shape of [w]. + While h is the attr:`dim`th dimension of the input weights, + and w is the product result of remain dimensions. + + Step 2: + While attr:`power_iters` is a positive interger, do following + iteration calculations with u and v for attr:`power_iters` + round. + \mathbf{v} = \mathbf{W}^{T} \mathbf{u} + \mathbf{v} = \frac{\mathbf{v}}{\|\mathbf{v}\|_2} + \mathbf{u} = \mathbf{W}^{T} \mathbf{v} + \mathbf{u} = \frac{\mathbf{u}}{\|\mathbf{u}\|_2} + + Step 3: + Calculate \sigma{W} and scale weight values. + \sigma{\mathbf{W}} = \mathbf{u}^{T} \mathbf{W} \mathbf{v} + \mathbf{W} := \frac{\mathbf{W}}{\sigma{\mathbf{W}}} + + Refer to `Spectral Normalization `_ . Args: weight(${weight_type}): ${weight_comment} dim(${dim_type}): ${dim_comment} eps(${eps_type}): ${eps_comment} - u_attr(ParamAttr|None): The parameter attribute for vector u in - spectral calculatings, set None to use default attribute, which - generates random values in normal distribution N(0, 1). Default: None. - v_attr(ParamAttr|None): The parameter attribute for vector v in - spectral calculatings, set None to use default attribute, which - generates random values in normal distribution N(0, 1). Default: None. name (str): The name of this layer. It is optional. Returns: @@ -3382,43 +3396,43 @@ def spectral_norm(weight, >>> x = fluid.layers.spectral_norm(weight=data, dim=1, power_iters=2) """ helper = LayerHelper('spectral_norm', **locals()) - dtype = helper.input_dtype() + dtype = weight.dtype # create intput and parameters inputs = {'Weight': weight} - input_shape = input.shape - if data_layout != 'NCHW': - raise ValueError("unsupported data layout:" + data_layout) - param_shape = [input_shape[1]] - if param_attr: - scale = helper.create_parameter( - attr=helper.param_attr, - shape=param_shape, - dtype=dtype, - default_initializer=Constant(1.0)) - inputs['Scale'] = scale - if bias_attr: - bias = helper.create_parameter( - attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True) - inputs['Bias'] = bias + input_shape = weight.shape + h = input_shape[dim] + w = np.prod(input_shape) // h + + u = helper.create_parameter( + attr=ParamAttr(), + shape=[h], + dtype=dtype, + default_initializer=Normal(0., 1.)) + u.stop_gradient = True + inputs['U'] = u + v = helper.create_parameter( + attr=ParamAttr(), + shape=[w], + dtype=dtype, + default_initializer=Normal(0., 1.)) + inputs['V'] = v + v.stop_gradient = True # create output - mean_out = helper.create_variable(dtype=dtype, stop_gradient=True) - variance_out = helper.create_variable(dtype=dtype, stop_gradient=True) - group_norm_out = helper.create_variable(dtype=dtype) + out = helper.create_variable(dtype=dtype) helper.append_op( - type="group_norm", + type="spectral_norm", inputs=inputs, - outputs={ - "Y": group_norm_out, - "Mean": mean_out, - "Variance": variance_out, - }, - attrs={"epsilon": epsilon, - "groups": groups}) + outputs={"Out": out, }, + attrs={ + "dim": dim, + "power_iters": power_iters, + "eps": eps, + }) - return helper.append_activation(group_norm_out) + return out def conv2d_transpose(input, diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index 30194f8cacf..ff49c1be979 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -1035,6 +1035,19 @@ class TestBook(unittest.TestCase): print(str(program)) + def test_spectral_norm(self): + program = Program() + with program_guard(program): + weight = layers.data( + name='weight', + shape=[2, 3, 32, 32], + dtype="float32", + append_batch_size=False) + out = layers.spectral_norm(weight, dim=1, power_iters=1) + self.assertIsNotNone(out) + + print(str(program)) + def test_shuffle_channel(self): program = Program() with program_guard(program): -- GitLab