tensorlayer3/tensorlayer/cost/tensorflow_cost.py

862 lines
33 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

#! /usr/bin/python
# -*- coding: utf-8 -*-
import numbers
import tensorflow as tf
from tensorflow.python.framework import ops
from tensorflow.python.ops import array_ops, math_ops, nn_ops, standard_ops
from tensorlayer import logging
__all__ = [
'softmax_cross_entropy_with_logits',
'sigmoid_cross_entropy',
'binary_cross_entropy',
'mean_squared_error',
'normalized_mean_square_error',
'absolute_difference_error',
'dice_coe',
'dice_hard_coe',
'iou_coe',
'cross_entropy_seq',
'cross_entropy_seq_with_mask',
'cosine_similarity',
'li_regularizer',
'lo_regularizer',
'maxnorm_regularizer',
'maxnorm_o_regularizer',
'maxnorm_i_regularizer',
]
def softmax_cross_entropy_with_logits(output, target, name=None):
"""Softmax cross-entropy operation, returns the TensorFlow expression of cross-entropy for two distributions,
it implements softmax internally. See ``tf.ops.sparse_softmax_cross_entropy_with_logits``.
Parameters
----------
output : Tensor
A batch of distribution with shape: [batch_size, num of classes].
target : Tensor
A batch of index with shape: [batch_size, ].
name : string
Name of this loss.
Examples
--------
>>> import tensorlayer as tl
>>> ce = tl.cost.softmax_cross_entropy_with_logits(y_logits, y_target_logits, 'my_loss')
References
-----------
- About cross-entropy: `<https://en.wikipedia.org/wiki/Cross_entropy>`__.
- The code is borrowed from: `<https://en.wikipedia.org/wiki/Cross_entropy>`__.
"""
# if name is None:
# raise Exception("Please give a unique name to tl.cost.cross_entropy for TF1.0+")
return tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=target, logits=output), name=name)
def sigmoid_cross_entropy(output, target, name=None):
"""Sigmoid cross-entropy operation, see ``tf.ops.sigmoid_cross_entropy_with_logits``.
Parameters
----------
output : Tensor
A batch of distribution with shape: [batch_size, num of classes].
target : Tensor
A batch of index with shape: [batch_size, ].
name : string
Name of this loss.
"""
return tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=target, logits=output), name=name)
def binary_cross_entropy(output, target, epsilon=1e-8, name='bce_loss'):
"""Binary cross entropy operation.
Parameters
----------
output : Tensor
Tensor with type of `float32` or `float64`.
target : Tensor
The target distribution, format the same with `output`.
epsilon : float
A small value to avoid output to be zero.
name : str
An optional name to attach to this function.
References
-----------
- `ericjang-DRAW <https://github.com/ericjang/draw/blob/master/draw.py#L73>`__
"""
# with ops.op_scope([output, target], name, "bce_loss") as name:
# output = ops.convert_to_tensor(output, name="preds")
# target = ops.convert_to_tensor(targets, name="target")
# with tf.name_scope(name):
return tf.reduce_mean(
tf.reduce_sum(
-(target * tf.math.log(output + epsilon) + (1. - target) * tf.math.log(1. - output + epsilon)), axis=1
), name=name
)
# For brevity, let `x = output`, `z = target`. The binary cross entropy loss is
#
# loss(x, z) = - sum_i (x[i] * log(z[i]) + (1 - x[i]) * log(1 - z[i]))
def mean_squared_error(output, target, is_mean=False, axis=-1, name="mean_squared_error"):
"""Return the TensorFlow expression of mean-square-error (L2) of two batch of data.
Parameters
----------
output : Tensor
2D, 3D or 4D tensor i.e. [batch_size, n_feature], [batch_size, height, width] or [batch_size, height, width, channel].
target : Tensor
The target distribution, format the same with `output`.
is_mean : boolean
Whether compute the mean or sum for each example.
- If True, use ``tf.reduce_mean`` to compute the loss between one target and predict data.
- If False, use ``tf.reduce_sum`` (default).
axis : int or list of int
The dimensions to reduce.
name : str
An optional name to attach to this function.
References
------------
- `Wiki Mean Squared Error <https://en.wikipedia.org/wiki/Mean_squared_error>`__
"""
# with tf.name_scope(name):
# if len(output.shape) == 2: # [batch_size, n_feature]
# axis = 1
# elif len(output.shape) == 3: # [batch_size, w, h]
# axis = [1, 2]
# elif len(output.shape) == 4: # [batch_size, w, h, c]
# axis = [1, 2, 3]
# else:
# raise Exception("Unknow dimension")
if is_mean:
mse = tf.reduce_mean(tf.reduce_mean(tf.math.squared_difference(output, target), axis), name=name)
else:
mse = tf.reduce_mean(tf.reduce_sum(tf.math.squared_difference(output, target), axis), name=name)
return mse
def normalized_mean_square_error(output, target, axis=-1, name="normalized_mean_squared_error_loss"):
"""Return the TensorFlow expression of normalized mean-square-error of two distributions.
Parameters
----------
output : Tensor
2D, 3D or 4D tensor i.e. [batch_size, n_feature], [batch_size, height, width] or [batch_size, height, width, channel].
target : Tensor
The target distribution, format the same with `output`.
axis : int or list of int
The dimensions to reduce.
name : str
An optional name to attach to this function.
"""
with tf.name_scope("normalized_mean_squared_error_loss"):
# if len(output.shape) == 2: # [batch_size, n_feature]
# axis = 1
# elif len(output.shape) == 3: # [batch_size, w, h]
# axis = [1, 2]
# elif len(output.shape) == 4: # [batch_size, w, h, c]
# axis = [1, 2, 3]
nmse_a = tf.sqrt(tf.reduce_sum(tf.math.squared_difference(output, target), axis=axis))
nmse_b = tf.sqrt(tf.reduce_sum(tf.square(target), axis=axis))
nmse = tf.reduce_mean(nmse_a / nmse_b, name=name)
return nmse
def absolute_difference_error(output, target, is_mean=False, axis=-1, name="absolute_difference_error_loss"):
"""Return the TensorFlow expression of absolute difference error (L1) of two batch of data.
Parameters
----------
output : Tensor
2D, 3D or 4D tensor i.e. [batch_size, n_feature], [batch_size, height, width] or [batch_size, height, width, channel].
target : Tensor
The target distribution, format the same with `output`.
is_mean : boolean
Whether compute the mean or sum for each example.
- If True, use ``tf.reduce_mean`` to compute the loss between one target and predict data.
- If False, use ``tf.reduce_sum`` (default).
axis : int or list of int
The dimensions to reduce.
name : str
An optional name to attach to this function.
"""
# # with tf.name_scope("absolute_difference_error_loss"):
# if len(output.shape) == 2: # [batch_size, n_feature]
# axis = 1
# elif len(output.shape) == 3: # [batch_size, w, h]
# axis = [1, 2]
# elif len(output.shape) == 4: # [batch_size, w, h, c]
# axis = [1, 2, 3]
# else:
# raise Exception("Unknow dimension")
if is_mean:
loss = tf.reduce_mean(tf.reduce_mean(tf.abs(output - target), axis), name=name)
else:
loss = tf.reduce_mean(tf.reduce_sum(tf.abs(output - target), axis), name=name)
return loss
def dice_coe(output, target, loss_type='jaccard', axis=(1, 2, 3), smooth=1e-5):
"""Soft dice (Sørensen or Jaccard) coefficient for comparing the similarity
of two batch of data, usually be used for binary image segmentation
i.e. labels are binary. The coefficient between 0 to 1, 1 means totally match.
Parameters
-----------
output : Tensor
A distribution with shape: [batch_size, ....], (any dimensions).
target : Tensor
The target distribution, format the same with `output`.
loss_type : str
``jaccard`` or ``sorensen``, default is ``jaccard``.
axis : tuple of int
All dimensions are reduced, default ``[1,2,3]``.
smooth : float
This small value will be added to the numerator and denominator.
- If both output and target are empty, it makes sure dice is 1.
- If either output or target are empty (all pixels are background), dice = ```smooth/(small_value + smooth)``, then if smooth is very small, dice close to 0 (even the image values lower than the threshold), so in this case, higher smooth can have a higher dice.
Examples
---------
>>> import tensorlayer as tl
>>> outputs = tl.ops.softmax(outputs)
>>> dice_loss = 1 - tl.cost.dice_coe(outputs, y_)
References
-----------
- `Wiki-Dice <https://en.wikipedia.org/wiki/SørensenDice_coefficient>`__
"""
inse = tf.reduce_sum(output * target, axis=axis)
if loss_type == 'jaccard':
l = tf.reduce_sum(output * output, axis=axis)
r = tf.reduce_sum(target * target, axis=axis)
elif loss_type == 'sorensen':
l = tf.reduce_sum(output, axis=axis)
r = tf.reduce_sum(target, axis=axis)
else:
raise Exception("Unknow loss_type")
# old axis=[0,1,2,3]
# dice = 2 * (inse) / (l + r)
# epsilon = 1e-5
# dice = tf.clip_by_value(dice, 0, 1.0-epsilon) # if all empty, dice = 1
# new haodong
dice = (2. * inse + smooth) / (l + r + smooth)
##
dice = tf.reduce_mean(dice, name='dice_coe')
return dice
def dice_hard_coe(output, target, threshold=0.5, axis=(1, 2, 3), smooth=1e-5):
"""Non-differentiable SørensenDice coefficient for comparing the similarity
of two batch of data, usually be used for binary image segmentation i.e. labels are binary.
The coefficient between 0 to 1, 1 if totally match.
Parameters
-----------
output : tensor
A distribution with shape: [batch_size, ....], (any dimensions).
target : tensor
The target distribution, format the same with `output`.
threshold : float
The threshold value to be true.
axis : tuple of integer
All dimensions are reduced, default ``(1,2,3)``.
smooth : float
This small value will be added to the numerator and denominator, see ``dice_coe``.
References
-----------
- `Wiki-Dice <https://en.wikipedia.org/wiki/SørensenDice_coefficient>`__
"""
output = tf.cast(output > threshold, dtype=tf.float32)
target = tf.cast(target > threshold, dtype=tf.float32)
inse = tf.reduce_sum(tf.multiply(output, target), axis=axis)
l = tf.reduce_sum(output, axis=axis)
r = tf.reduce_sum(target, axis=axis)
# old axis=[0,1,2,3]
# hard_dice = 2 * (inse) / (l + r)
# epsilon = 1e-5
# hard_dice = tf.clip_by_value(hard_dice, 0, 1.0-epsilon)
# new haodong
hard_dice = (2. * inse + smooth) / (l + r + smooth)
##
hard_dice = tf.reduce_mean(hard_dice, name='hard_dice')
return hard_dice
def iou_coe(output, target, threshold=0.5, axis=(1, 2, 3), smooth=1e-5):
"""Non-differentiable Intersection over Union (IoU) for comparing the
similarity of two batch of data, usually be used for evaluating binary image segmentation.
The coefficient between 0 to 1, and 1 means totally match.
Parameters
-----------
output : tensor
A batch of distribution with shape: [batch_size, ....], (any dimensions).
target : tensor
The target distribution, format the same with `output`.
threshold : float
The threshold value to be true.
axis : tuple of integer
All dimensions are reduced, default ``(1,2,3)``.
smooth : float
This small value will be added to the numerator and denominator, see ``dice_coe``.
Notes
------
- IoU cannot be used as training loss, people usually use dice coefficient for training, IoU and hard-dice for evaluating.
"""
pre = tf.cast(output > threshold, dtype=tf.float32)
truth = tf.cast(target > threshold, dtype=tf.float32)
inse = tf.reduce_sum(tf.multiply(pre, truth), axis=axis) # AND
union = tf.reduce_sum(tf.cast(tf.add(pre, truth) >= 1, dtype=tf.float32), axis=axis) # OR
# old axis=[0,1,2,3]
# epsilon = 1e-5
# batch_iou = inse / (union + epsilon)
# new haodong
batch_iou = (inse + smooth) / (union + smooth)
iou = tf.reduce_mean(batch_iou, name='iou_coe')
return iou # , pre, truth, inse, union
# ## test soft/hard dice and iou
# import numpy as np
# y = np.zeros((1,10,10,1))
# # y[0,0:5,0:5]=1.0
# o = np.zeros((1,10,10,1))
# # o[:,:,:,:] = 0 # what we want: dice=0 iou=0 OK
# # o[0,0:2,0:2]=0.3 # what we want: dice larger iou=0 OK
# # o[0,0:2,0:2]=0.6 # what we want: dice larger iou small OK
# # o[0,0:3,0:3]=0.6 # what we want: dice larger iou larger OK
# # o[0,0:3,0:3]=1 # what we want: dice larger iou same OK
# # o[0,0:5,0:5]=1 # what we want: dice=1 iou=1 OK
# # o[0,0:5,0:5]=0.3 # what we want: dice smaller iou=0 OK
# # o[0,0:5,0:5]=1e-2 # what we want: dice≈0 iou=0 OK
# # o[0,8:10,8:10]=1.0 # what we want: dice=0 iou=0 OK
# # o[0,8:10,8:10]=1e-10 # what we want: dice=0 iou=0 OK
# # y[:,:,:,:] = o[:,:,:,:] = 0 # what we want: dice=1 iou=1 OK
# ## why in u-net, dice=1 hard-dice=1 iou=1 exist?? print bug?
#
# d = dice_coe(o, y, 'jaccard', smooth=1.)
# hd = dice_hard_coe(o, y, smooth=1e-5)
# i = iou_coe(o, y, smooth=1e-5)
# sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# # sess.run(tf.local_variables_initializer())
# print(sess.run([d,hd,i]))
# # p, t, i, u = sess.run([pre, truth, inse, union])
# # import pprint
# # pprint.pprint(((y>0.5)*(o>0.5)).astype(int).tolist())
# # pprint.pprint(p.tolist())
# # pprint.pprint(t.tolist())
# # pprint.pprint(i)
# # pprint.pprint(u)
# exit()
def sequence_loss_by_example(
logits, targets, weights, average_across_timesteps=True, softmax_loss_function=None, name=None
):
"""Weighted cross-entropy loss for a sequence of logits (per example). see original tensorflow code :
<https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/legacy_seq2seq/python/ops/seq2seq.py#L1057>
Parameters
----------
logits: List
List of 2D Tensors of shape [batch_size x num_decoder_symbols].
targets: List
List of 1D batch-sized int32 Tensors of the same length as logits.
weights: List
List of 1D batch-sized float-Tensors of the same length as logits.
average_across_timesteps: Boolean
If set, divide the returned cost by the total label weight.
softmax_loss_function: None or Function
Function (labels, logits) -> loss-batch to be used instead of the standard softmax (the default if this is None).
**Note that to avoid confusion, it is required for the function to accept named arguments.**
name: None or str
Optional name for this operation, default: "sequence_loss_by_example".
Returns
-------
1D batch-sized float Tensor: The log-perplexity for each sequence.
Raises
------
ValueError: If len(logits) is different from len(targets) or len(weights).
"""
if len(targets) != len(logits) or len(weights) != len(logits):
raise ValueError(
"Lengths of logits, weights, and targets must be the same "
"%d, %d, %d." % (len(logits), len(weights), len(targets))
)
with ops.name_scope(name, "sequence_loss_by_example", logits + targets + weights):
log_perp_list = []
for logit, target, weight in zip(logits, targets, weights):
if softmax_loss_function is None:
# TODO(irving,ebrevdo): This reshape is needed because
# sequence_loss_by_example is called with scalars sometimes, which
# violates our general scalar strictness policy.
target = array_ops.reshape(target, [-1])
crossent = nn_ops.sparse_softmax_cross_entropy_with_logits(labels=target, logits=logit)
else:
crossent = softmax_loss_function(labels=target, logits=logit)
log_perp_list.append(crossent * weight)
log_perps = math_ops.add_n(log_perp_list)
if average_across_timesteps:
total_size = math_ops.add_n(weights)
total_size += 1e-12 # Just to avoid division by 0 for all-0 weights.
log_perps /= total_size
return log_perps
def cross_entropy_seq(logits, target_seqs, batch_size=None):
"""Returns the expression of cross-entropy of two sequences, implement
softmax internally. Normally be used for fixed length RNN outputs, see `PTB example <https://github.com/tensorlayer/tensorlayer/blob/master/example/tutorial_ptb_lstm.py>`__.
Parameters
----------
logits : Tensor
2D tensor with shape of `[batch_size * n_steps, n_classes]`.
target_seqs : Tensor
The target sequence, 2D tensor `[batch_size, n_steps]`, if the number of step is dynamic, please use ``tl.cost.cross_entropy_seq_with_mask`` instead.
batch_size : None or int.
Whether to divide the cost by batch size.
- If integer, the return cost will be divided by `batch_size`.
- If None (default), the return cost will not be divided by anything.
Examples
--------
>>> import tensorlayer as tl
>>> # see `PTB example <https://github.com/tensorlayer/tensorlayer/blob/master/example/tutorial_ptb_lstm.py>`__.for more details
>>> # outputs shape : (batch_size * n_steps, n_classes)
>>> # targets shape : (batch_size, n_steps)
>>> cost = tl.cost.cross_entropy_seq(outputs, targets)
"""
sequence_loss_by_example_fn = sequence_loss_by_example
loss = sequence_loss_by_example_fn(
[logits], [tf.reshape(target_seqs, [-1])], [tf.ones_like(tf.reshape(target_seqs, [-1]), dtype=tf.float32)]
)
# [tf.ones([batch_size * num_steps])])
cost = tf.reduce_sum(loss) # / batch_size
if batch_size is not None:
cost = cost / batch_size
return cost
def cross_entropy_seq_with_mask(logits, target_seqs, input_mask, return_details=False, name=None):
"""Returns the expression of cross-entropy of two sequences, implement
softmax internally. Normally be used for Dynamic RNN with Synced sequence input and output.
Parameters
-----------
logits : Tensor
2D tensor with shape of [batch_size * ?, n_classes], `?` means dynamic IDs for each example.
- Can be get from `DynamicRNNLayer` by setting ``return_seq_2d`` to `True`.
target_seqs : Tensor
int of tensor, like word ID. [batch_size, ?], `?` means dynamic IDs for each example.
input_mask : Tensor
The mask to compute loss, it has the same size with `target_seqs`, normally 0 or 1.
return_details : boolean
Whether to return detailed losses.
- If False (default), only returns the loss.
- If True, returns the loss, losses, weights and targets (see source code).
Examples
--------
>>> import tensorlayer as tl
>>> import tensorflow as tf
>>> import numpy as np
>>> batch_size = 64
>>> vocab_size = 10000
>>> embedding_size = 256
>>> ni = tl.layers.Input([batch_size, None], dtype=tf.int64)
>>> net_lits = []
>>> net_list.append(tl.layers.Embedding(
... vocabulary_size = vocab_size,
... embedding_size = embedding_size,
... name = 'seq_embedding'))
>>> net_list.append(tl.layers.RNN(
... cell =tf.keras.layers.LSTMCell(units=embedding_size, dropout=0.1),
... return_seq_2d = True,
... name = 'dynamicrnn'))
>>> net_list.append(tl.layers.Dense(n_units=vocab_size, name="output"))
>>> model = tl.layers.SequentialLayer(net_list)
>>> input_seqs = np.random.randint(0, 10, size=(batch_size, 10), dtype=np.int64)
>>> target_seqs = np.random.randint(0, 10, size=(batch_size, 10), dtype=np.int64)
>>> input_mask = np.random.randint(0, 2, size=(batch_size, 10), dtype=np.int64)
>>> outputs = model(input_seqs)
>>> loss = tl.cost.cross_entropy_seq_with_mask(outputs, target_seqs, input_mask)
"""
targets = tf.reshape(target_seqs, [-1]) # to one vector
weights = tf.cast(tf.reshape(input_mask, [-1]), dtype=tf.float32) # to one vector like targets
losses = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=targets, name=name) * weights
# losses = tf.reduce_mean(tf.ops.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=targets, name=name)) # for TF1.0 and others
loss = tf.divide(
tf.reduce_sum(losses), # loss from mask. reduce_sum before element-wise mul with mask !!
tf.reduce_sum(weights),
name="seq_loss_with_mask"
)
if return_details:
return loss, losses, weights, targets
else:
return loss
def cosine_similarity(v1, v2):
"""Cosine similarity [-1, 1].
Parameters
----------
v1, v2 : Tensor
Tensor with the same shape [batch_size, n_feature].
References
----------
- `Wiki <https://en.wikipedia.org/wiki/Cosine_similarity>`__.
"""
return tf.reduce_sum(tf.multiply(v1, v2), 1) / \
(tf.sqrt(tf.reduce_sum(tf.multiply(v1, v1), 1)) *
tf.sqrt(tf.reduce_sum(tf.multiply(v2, v2), 1)))
# Regularization Functions
def li_regularizer(scale, scope=None):
"""Li regularization removes the neurons of previous layer. The `i` represents `inputs`.
Returns a function that can be used to apply group li regularization to weights.
The implementation follows `TensorFlow contrib <https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/layers/python/layers/regularizers.py>`__.
Parameters
----------
scale : float
A scalar multiplier `Tensor`. 0.0 disables the regularizer.
scope: str
An optional scope name for this function.
Returns
--------
A function with signature `li(weights, name=None)` that apply Li regularization.
Raises
------
ValueError : if scale is outside of the range [0.0, 1.0] or if scale is not a float.
"""
if isinstance(scale, numbers.Integral):
raise ValueError('scale cannot be an integer: %s' % scale)
if isinstance(scale, numbers.Real):
if scale < 0.:
raise ValueError('Setting a scale less than 0 on a regularizer: %g' % scale)
if scale >= 1.:
raise ValueError('Setting a scale greater than 1 on a regularizer: %g' % scale)
if scale == 0.:
logging.info('Scale of 0 disables regularizer.')
return lambda _, name=None: None
def li(weights):
"""Applies li regularization to weights."""
with tf.name_scope('li_regularizer') as scope:
my_scale = ops.convert_to_tensor(scale, dtype=weights.dtype.base_dtype, name='scale')
# if tf.__version__ <= '0.12':
# standard_ops_fn = standard_ops.mul
# else:
standard_ops_fn = standard_ops.multiply
return standard_ops_fn(
my_scale, standard_ops.reduce_sum(standard_ops.sqrt(standard_ops.reduce_sum(tf.square(weights), 1))),
name=scope
)
return li
def lo_regularizer(scale):
"""Lo regularization removes the neurons of current layer. The `o` represents `outputs`
Returns a function that can be used to apply group lo regularization to weights.
The implementation follows `TensorFlow contrib <https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/layers/python/layers/regularizers.py>`__.
Parameters
----------
scale : float
A scalar multiplier `Tensor`. 0.0 disables the regularizer.
Returns
-------
A function with signature `lo(weights, name=None)` that apply Lo regularization.
Raises
------
ValueError : If scale is outside of the range [0.0, 1.0] or if scale is not a float.
"""
if isinstance(scale, numbers.Integral):
raise ValueError('scale cannot be an integer: %s' % scale)
if isinstance(scale, numbers.Real):
if scale < 0.:
raise ValueError('Setting a scale less than 0 on a regularizer: %g' % scale)
if scale >= 1.:
raise ValueError('Setting a scale greater than 1 on a regularizer: %g' % scale)
if scale == 0.:
logging.info('Scale of 0 disables regularizer.')
return lambda _, name=None: None
def lo(weights, name='lo_regularizer'):
"""Applies group column regularization to weights."""
with tf.name_scope(name) as scope:
my_scale = ops.convert_to_tensor(scale, dtype=weights.dtype.base_dtype, name='scale')
# if tf.__version__ <= '0.12':
# standard_ops_fn = standard_ops.mul
# else:
standard_ops_fn = standard_ops.multiply
return standard_ops_fn(
my_scale, standard_ops.reduce_sum(standard_ops.sqrt(standard_ops.reduce_sum(tf.square(weights), 0))),
name=scope
)
return lo
def maxnorm_regularizer(scale=1.0):
"""Max-norm regularization returns a function that can be used to apply max-norm regularization to weights.
More about max-norm, see `wiki-max norm <https://en.wikipedia.org/wiki/Matrix_norm#Max_norm>`_.
The implementation follows `TensorFlow contrib <https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/layers/python/layers/regularizers.py>`__.
Parameters
----------
scale : float
A scalar multiplier `Tensor`. 0.0 disables the regularizer.
Returns
---------
A function with signature `mn(weights, name=None)` that apply Lo regularization.
Raises
--------
ValueError : If scale is outside of the range [0.0, 1.0] or if scale is not a float.
"""
if isinstance(scale, numbers.Integral):
raise ValueError('scale cannot be an integer: %s' % scale)
if isinstance(scale, numbers.Real):
if scale < 0.:
raise ValueError('Setting a scale less than 0 on a regularizer: %g' % scale)
# if scale >= 1.:
# raise ValueError('Setting a scale greater than 1 on a regularizer: %g' %
# scale)
if scale == 0.:
logging.info('Scale of 0 disables regularizer.')
return lambda _, name=None: None
def mn(weights, name='max_regularizer'):
"""Applies max-norm regularization to weights."""
with tf.name_scope(name) as scope:
my_scale = ops.convert_to_tensor(scale, dtype=weights.dtype.base_dtype, name='scale')
# if tf.__version__ <= '0.12':
# standard_ops_fn = standard_ops.mul
# else:
standard_ops_fn = standard_ops.multiply
return standard_ops_fn(my_scale, standard_ops.reduce_max(standard_ops.abs(weights)), name=scope)
return mn
def maxnorm_o_regularizer(scale):
"""Max-norm output regularization removes the neurons of current layer.
Returns a function that can be used to apply max-norm regularization to each column of weight matrix.
The implementation follows `TensorFlow contrib <https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/layers/python/layers/regularizers.py>`__.
Parameters
----------
scale : float
A scalar multiplier `Tensor`. 0.0 disables the regularizer.
Returns
---------
A function with signature `mn_o(weights, name=None)` that apply Lo regularization.
Raises
---------
ValueError : If scale is outside of the range [0.0, 1.0] or if scale is not a float.
"""
if isinstance(scale, numbers.Integral):
raise ValueError('scale cannot be an integer: %s' % scale)
if isinstance(scale, numbers.Real):
if scale < 0.:
raise ValueError('Setting a scale less than 0 on a regularizer: %g' % scale)
# if scale >= 1.:
# raise ValueError('Setting a scale greater than 1 on a regularizer: %g' %
# scale)
if scale == 0.:
logging.info('Scale of 0 disables regularizer.')
return lambda _, name=None: None
def mn_o(weights, name='maxnorm_o_regularizer'):
"""Applies max-norm regularization to weights."""
with tf.name_scope(name) as scope:
my_scale = ops.convert_to_tensor(scale, dtype=weights.dtype.base_dtype, name='scale')
if tf.__version__ <= '0.12':
standard_ops_fn = standard_ops.mul
else:
standard_ops_fn = standard_ops.multiply
return standard_ops_fn(
my_scale, standard_ops.reduce_sum(standard_ops.reduce_max(standard_ops.abs(weights), 0)), name=scope
)
return mn_o
def maxnorm_i_regularizer(scale):
"""Max-norm input regularization removes the neurons of previous layer.
Returns a function that can be used to apply max-norm regularization to each row of weight matrix.
The implementation follows `TensorFlow contrib <https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/layers/python/layers/regularizers.py>`__.
Parameters
----------
scale : float
A scalar multiplier `Tensor`. 0.0 disables the regularizer.
Returns
---------
A function with signature `mn_i(weights, name=None)` that apply Lo regularization.
Raises
---------
ValueError : If scale is outside of the range [0.0, 1.0] or if scale is not a float.
"""
if isinstance(scale, numbers.Integral):
raise ValueError('scale cannot be an integer: %s' % scale)
if isinstance(scale, numbers.Real):
if scale < 0.:
raise ValueError('Setting a scale less than 0 on a regularizer: %g' % scale)
# if scale >= 1.:
# raise ValueError('Setting a scale greater than 1 on a regularizer: %g' %
# scale)
if scale == 0.:
logging.info('Scale of 0 disables regularizer.')
return lambda _, name=None: None
def mn_i(weights, name='maxnorm_i_regularizer'):
"""Applies max-norm regularization to weights."""
with tf.name_scope(name) as scope:
my_scale = ops.convert_to_tensor(scale, dtype=weights.dtype.base_dtype, name='scale')
if tf.__version__ <= '0.12':
standard_ops_fn = standard_ops.mul
else:
standard_ops_fn = standard_ops.multiply
return standard_ops_fn(
my_scale, standard_ops.reduce_sum(standard_ops.reduce_max(standard_ops.abs(weights), 1)), name=scope
)
return mn_i
def huber_loss(
output, target, is_mean=True, delta=1.0, dynamichuber=False, reverse=False, axis=-1, epsilon=0.00001, name=None
):
"""Huber Loss operation, see ``https://en.wikipedia.org/wiki/Huber_loss`` .
Reverse Huber Loss operation, see ''https://statweb.stanford.edu/~owen/reports/hhu.pdf''.
Dynamic Reverse Huber Loss operation, see ''https://arxiv.org/pdf/1606.00373.pdf''.
Parameters
----------
output : Tensor
A distribution with shape: [batch_size, ....], (any dimensions).
target : Tensor
The target distribution, format the same with `output`.
is_mean : boolean
Whether compute the mean or sum for each example.
- If True, use ``tf.reduce_mean`` to compute the loss between one target and predict data (default).
- If False, use ``tf.reduce_sum``.
delta: float
The point where the huber loss function changes from a quadratic to linear.
dynamichuber: boolean
Whether compute the coefficient c for each batch.
- If True, c is 20% of the maximal per-batch error.
- If False, c is delta.
reverse: boolean
Whether compute the reverse huber loss.
axis : int or list of int
The dimensions to reduce.
epsilon:
Eplison.
name : string
Name of this loss.
"""
if reverse:
if dynamichuber:
huber_c = 0.2 * tf.reduce_max(tf.abs(output - target))
else:
huber_c = delta
if is_mean:
loss = tf.reduce_mean(
tf.where(
tf.less_equal(tf.abs(output - target), huber_c), tf.abs(output - target),
tf.multiply(
tf.pow(output - target, 2.0) + tf.pow(huber_c, 2.0),
tf.math.divide_no_nan(.5, huber_c + epsilon)
)
), name=name
)
else:
loss = tf.reduce_mean(
tf.reduce_sum(
tf.where(
tf.less_equal(tf.abs(output - target), huber_c), tf.abs(output - target),
tf.multiply(
tf.pow(output - target, 2.0) + tf.pow(huber_c, 2.0),
tf.math.divide_no_nan(.5, huber_c + epsilon)
)
), axis
), name=name
)
elif is_mean:
loss = tf.reduce_mean(
tf.where(
tf.less_equal(tf.abs(output - target), delta), 0.5 * tf.pow(output - target, 2),
delta * (tf.abs(output - target) - 0.5 * delta)
), name=name
)
else:
loss = tf.reduce_mean(
tf.reduce_sum(
tf.where(
tf.less_equal(tf.abs(output - target), delta), 0.5 * tf.pow(output - target, 2),
delta * (tf.abs(output - target) - 0.5 * delta)
), axis
), name=name
)
return loss