keras Layer
keras Layer
Simple Introduction
Keras实现了很多层，包括核心层、卷基层、RNN网络层等诸多常用的网络结构。
Core 核心层
Source
class Layer(object):
'''Abstract base layer class. All Keras layers accept certain keyword arguments: trainable: boolean. Set to "False" before model compilation to freeze layer weights (they won't be updated further during training). input_shape: a tuple of integers specifying the expected shape of the input samples. Does not includes the batch size. (e.g. `(100,)` for 100-dimensional inputs). batch_input_shape: a tuple of integers specifying the expected shape of a batch of input samples. Includes the batch size (e.g. `(32, 100)` for a batch of 32 100-dimensional inputs). '''
def __init__(self, **kwargs):
allowed_kwargs = {'input_shape',
'trainable',
'batch_input_shape',
'cache_enabled'}
for kwarg in kwargs:
assert kwarg in allowed_kwargs, 'Keyword argument not understood: ' + kwarg
if 'input_shape' in kwargs:
self.set_input_shape((None,) + tuple(kwargs['input_shape']))
if 'batch_input_shape' in kwargs:
self.set_input_shape(tuple(kwargs['batch_input_shape']))
if 'trainable' in kwargs:
self._trainable = kwargs['trainable']
if not hasattr(self, 'params'):
self.params = []
self._cache_enabled = True
if 'cache_enabled' in kwargs:
self._cache_enabled = kwargs['cache_enabled']
@property
def cache_enabled(self):
return self._cache_enabled
@cache_enabled.setter
def cache_enabled(self, value):
self._cache_enabled = value
def __call__(self, X, mask=None, train=False):
# set temporary input
tmp_input = self.get_input
tmp_mask = None
if hasattr(self, 'get_input_mask'):
tmp_mask = self.get_input_mask
self.get_input_mask = lambda _: mask
self.get_input = lambda _: X
Y = self.get_output(train=train)
# return input to what it was
if hasattr(self, 'get_input_mask'):
self.get_input_mask = tmp_mask
self.get_input = tmp_input
return Y
def set_previous(self, layer, connection_map={}):
'''Connect a layer to its parent in the computational graph. '''
assert self.nb_input == layer.nb_output == 1, 'Cannot connect layers: input count and output count should be 1.'
if hasattr(self, 'input_ndim'):
assert self.input_ndim == len(layer.output_shape), ('Incompatible shapes: layer expected input with ndim=' +
str(self.input_ndim) +
' but previous layer has output_shape ' +
str(layer.output_shape))
if layer.get_output_mask() is not None:
assert self.supports_masked_input(), 'Cannot connect non-masking layer to layer with masked output.'
self.previous = layer
self.build()
def build(self):
'''Instantiation of layer weights. Called after `set_previous`, or after `set_input_shape`, once the layer has a defined input shape. Must be implemented on all layers that have weights. '''
pass
@property
def trainable(self):
if hasattr(self, '_trainable'):
return self._trainable
else:
return True
@trainable.setter
def trainable(self, value):
self._trainable = value
@property
def nb_input(self):
return 1
@property
def nb_output(self):
return 1
@property
def input_shape(self):
# if layer is not connected (e.g. input layer),
# input shape can be set manually via _input_shape attribute.
if hasattr(self, 'previous'):
return self.previous.output_shape
elif hasattr(self, '_input_shape'):
return self._input_shape
else:
raise Exception('Layer is not connected. Did you forget to set "input_shape"?')
def set_input_shape(self, input_shape):
if type(input_shape) not in [tuple, list]:
raise Exception('Invalid input shape - input_shape should be a tuple of int.')
input_shape = tuple(input_shape)
if hasattr(self, 'input_ndim') and self.input_ndim:
if self.input_ndim != len(input_shape):
raise Exception('Invalid input shape - Layer expects input ndim=' +
str(self.input_ndim) +
', was provided with input shape ' + str(input_shape))
self._input_shape = input_shape
self.input = K.placeholder(shape=self._input_shape)
self.build()
@property
def output_shape(self):
# default assumption: tensor shape unchanged.
return self.input_shape
def get_output(self, train=False):
return self.get_input(train)
def get_input(self, train=False):
if hasattr(self, 'previous'):
# to avoid redundant computations,
# layer outputs are cached when possible.
if hasattr(self, 'layer_cache') and self.cache_enabled:
previous_layer_id = '%s_%s' % (id(self.previous), train)
if previous_layer_id in self.layer_cache:
return self.layer_cache[previous_layer_id]
previous_output = self.previous.get_output(train=train)
if hasattr(self, 'layer_cache') and self.cache_enabled:
previous_layer_id = '%s_%s' % (id(self.previous), train)
self.layer_cache[previous_layer_id] = previous_output
return previous_output
elif hasattr(self, 'input'):
return self.input
else:
raise Exception('Layer is not connected' +
'and is not an input layer.')
def supports_masked_input(self):
'''Whether or not this layer respects the output mask of its previous layer in its calculations. If you try to attach a layer that does *not* support masked_input to a layer that gives a non-None output_mask(), an error will be raised. '''
return False
def get_output_mask(self, train=None):
'''For some models (such as RNNs) you want a way of being able to mark some output data-points as "masked", so they are not used in future calculations. In such a model, get_output_mask() should return a mask of one less dimension than get_output() (so if get_output is (nb_samples, nb_timesteps, nb_dimensions), then the mask is (nb_samples, nb_timesteps), with a one for every unmasked datapoint, and a zero for every masked one. If there is *no* masking then it shall return None. For instance if you attach an Activation layer (they support masking) to a layer with an output_mask, then that Activation shall also have an output_mask. If you attach it to a layer with no such mask, then the Activation's get_output_mask shall return None. Some layers have an output_mask even if their input is unmasked, notably Embedding which can turn the entry "0" into a mask. '''
return None
def set_weights(self, weights):
'''Set the weights of the layer. weights: a list of numpy arrays. The number of arrays and their shape must match number of the dimensions of the weights of the layer (i.e. it should match the output of `get_weights`). '''
assert len(self.params) == len(weights), ('Provided weight array does not match layer weights (' +
str(len(self.params)) + ' layer params vs. ' +
str(len(weights)) + ' provided weights)')
for p, w in zip(self.params, weights):
if K.get_value(p).shape != w.shape:
raise Exception('Layer shape %s not compatible with weight shape %s.' % (K.get_value(p).shape, w.shape))
K.set_value(p, w)
def get_weights(self):
'''Return the weights of the layer, as a list of numpy arrays. '''
weights = []
for p in self.params:
weights.append(K.get_value(p))
return weights
def get_config(self):
'''Return the parameters of the layer, as a dictionary. '''
config = {'name': self.__class__.__name__}
if hasattr(self, '_input_shape'):
config['input_shape'] = self._input_shape[1:]
if hasattr(self, '_trainable'):
config['trainable'] = self._trainable
config['cache_enabled'] = self.cache_enabled
return config
def get_params(self):
consts = []
updates = []
if hasattr(self, 'regularizers'):
regularizers = self.regularizers
else:
regularizers = []
if hasattr(self, 'constraints') and len(self.constraints) == len(self.params):
for c in self.constraints:
if c:
consts.append(c)
else:
consts.append(constraints.identity())
elif hasattr(self, 'constraint') and self.constraint:
consts += [self.constraint for _ in range(len(self.params))]
else:
consts += [constraints.identity() for _ in range(len(self.params))]
if hasattr(self, 'updates') and self.updates:
updates += self.updates
return self.params, regularizers, consts, updates
def count_params(self):
'''Return the total number of floats (or ints) composing the weights of the layer. '''
return sum([K.count_params(p) for p in self.params])set_previous
设置previous layer, 使previous layer连接到当前的layer,同时会调用build方法初始化regularizers，weights等参数.
build
被set_previous调用，初始化regularizers，weights等参数.
input_shape
python property. 如果该layer是输入层，返回自身的input shape, 否则返回previous layer的input shape.
set_input_shape
设置input shape(tuple, list), 并调用build方法初始化regularizers，weights等参数.
get_input
返回previous layer的output，如果当前层是输入层，则返回当前的输入.
Activation Layer
主要是计算经过激活函数后输出值，激活函数有softmax， softplus， relu， tanh， sigmoid， hard_sigmoid， linear.
source
class Activation(MaskedLayer):
'''Apply an activation function to an output. # Input shape Arbitrary. Use the keyword argument `input_shape` (tuple of integers, does not include the samples axis) when using this layer as the first layer in a model. # Output shape Same shape as input. # Arguments: activation: name of activation function to use (see: [activations](../activations.md)), or alternatively, a Theano or TensorFlow operation. '''
def __init__(self, activation, **kwargs):
super(Activation, self).__init__(**kwargs)
self.activation = activations.get(activation)
def get_output(self, train=False):
X = self.get_input(train)
return self.activation(X)
def get_config(self):
config = {'name': self.__class__.__name__,
'activation': self.activation.__name__}
base_config = super(Activation, self).get_config()
return dict(list(base_config.items()) + list(config.items()))Lambda Layer
该layer的output是经过lambda计算，如果该layer是input layer,则lambda的input是当前layer 的input，否则是previous layer的input
example
kerasmodel.add_node(Lambda(lambda x:x.sum(2)), name='merge',inputs=['embedding','embedpoint'], merge_mode='mul')source
class Lambda(Layer):
'''Used for evaluating an arbitrary Theano / TensorFlow expression on the output of the previous layer. # Input shape Arbitrary. Use the keyword argument input_shape (tuple of integers, does not include the samples axis) when using this layer as the first layer in a model. # Output shape Specified by `output_shape` argument. # Arguments function: The function to be evaluated. Takes one argument: the output of previous layer output_shape: Expected output shape from function. Could be a tuple or a function of the shape of the input '''
def __init__(self, function, output_shape=None, **kwargs):
super(Lambda, self).__init__(**kwargs)
py3 = sys.version_info[0] == 3
if py3:
self.function = marshal.dumps(function.__code__)
else:
assert hasattr(function, 'func_code'), ('The Lambda layer "function"'
' argument must be a Python function.')
self.function = marshal.dumps(function.func_code)
if output_shape is None:
self._output_shape = None
elif type(output_shape) in {tuple, list}:
self._output_shape = tuple(output_shape)
else:
if py3:
self._output_shape = marshal.dumps(output_shape.__code__)
else:
self._output_shape = marshal.dumps(output_shape.func_code)
super(Lambda, self).__init__()
@property
def output_shape(self):
if self._output_shape is None:
return self.input_shape
elif type(self._output_shape) == tuple:
return (self.input_shape[0], ) + self._output_shape
else:
output_shape_func = marshal.loads(self._output_shape)
output_shape_func = types.FunctionType(output_shape_func, globals())
shape = output_shape_func(self.previous.output_shape)
if type(shape) not in {list, tuple}:
raise Exception('output_shape function must return a tuple')
return tuple(shape)
def get_output(self, train=False):
func = marshal.loads(self.function)
func = types.FunctionType(func, globals())
if hasattr(self, 'previous'):
return func(self.previous.get_output(train))
else:
return func(self.input)Embedding Layer
使用keras实现word2Vector时，需要用到Embedding Layer