Source code for deephyper.nas.space.op.merge

import tensorflow as tf

from deephyper.nas.space.op import Operation
import deephyper as dh


[docs]class Concatenate(Operation): """Concatenate operation. Args: graph: node (Node): stacked_nodes (list(Node)): nodes to concatenate axis (int): axis to concatenate """ def __init__(self, search_space, stacked_nodes=None, axis=-1): self.search_space = search_space self.node = None # current_node of the operation self.stacked_nodes = stacked_nodes self.axis = axis def __str__(self): return "Concatenate"
[docs] def init(self, current_node): self.node = current_node if self.stacked_nodes is not None: for n in self.stacked_nodes: self.search_space.connect(n, self.node)
def __call__(self, values, **kwargs): # case where there is no inputs if len(values) == 0: return [] len_shp = max([len(x.get_shape()) for x in values]) if len_shp > 4: raise RuntimeError( f"This concatenation is for 2D or 3D tensors only but a {len_shp-1}D is passed!" ) # zeros padding if len(values) > 1: if all( map( lambda x: len(x.get_shape()) == len_shp or len(x.get_shape()) == (len_shp - 1), values, ) ): # all tensors should have same number of dimensions 2d or 3d, but we can also accept a mix of 2d en 3d tensors # we have a mix of 2d and 3d tensors so we are expanding 2d tensors to be 3d with last_dim==1 for i, v in enumerate(values): if len(v.get_shape()) < len_shp: values[i] = tf.keras.layers.Reshape( (*tuple(v.get_shape()[1:]), 1) )(v) # for 3d tensors concatenation is applied along last dim (axis=-1), so we are applying a zero padding to make 2nd dimensions (ie. shape()[1]) equals if len_shp == 3: max_len = max(map(lambda x: int(x.get_shape()[1]), values)) paddings = map(lambda x: max_len - int(x.get_shape()[1]), values) for i, (p, v) in enumerate(zip(paddings, values)): lp = p // 2 rp = p - lp values[i] = tf.keras.layers.ZeroPadding1D(padding=(lp, rp))(v) # elif len_shp == 2 nothing to do else: raise RuntimeError( f"All inputs of concatenation operation should have same shape length:\n" f"number_of_inputs=={len(values)}\n" f"shape_of_inputs=={[str(x.get_shape()) for x in values]}" ) # concatenation if len(values) > 1: out = tf.keras.layers.Concatenate(axis=-1)(values) else: out = values[0] return out
[docs]class AddByPadding(Operation): """Add operation. If tensor are of different shapes a padding will be applied before adding them. Args: search_space (KSearchSpace): [description]. Defaults to None. activation ([type], optional): Activation function to apply after adding ('relu', tanh', 'sigmoid'...). Defaults to None. stacked_nodes (list(Node)): nodes to add. axis (int): axis to concatenate. """ def __init__(self, search_space, stacked_nodes=None, activation=None, axis=-1): self.search_space = search_space self.node = None # current_node of the operation self.stacked_nodes = stacked_nodes self.activation = activation self.axis = axis
[docs] def init(self, current_node): self.node = current_node if self.stacked_nodes is not None: for n in self.stacked_nodes: self.search_space.connect(n, self.node)
def __call__(self, values, **kwargs): # case where there is no inputs if len(values) == 0: return [] values = values[:] max_len_shp = max([len(x.get_shape()) for x in values]) # zeros padding if len(values) > 1: for i, v in enumerate(values): if len(v.get_shape()) < max_len_shp: values[i] = tf.keras.layers.Reshape( ( *tuple(v.get_shape()[1:]), *tuple(1 for i in range(max_len_shp - len(v.get_shape()))), ) )(v) def max_dim_i(i): return max(map(lambda x: int(x.get_shape()[i]), values)) max_dims = [None] + list(map(max_dim_i, range(1, max_len_shp))) def paddings_dim_i(i): return list(map(lambda x: max_dims[i] - int(x.get_shape()[i]), values)) paddings_dim = list(map(paddings_dim_i, range(1, max_len_shp))) for i in range(len(values)): paddings = list() for j in range(len(paddings_dim)): p = paddings_dim[j][i] lp = p // 2 rp = p - lp paddings.append([lp, rp]) if sum(map(sum, paddings)) != 0: values[i] = dh.layers.Padding(paddings)(values[i]) # concatenation if len(values) > 1: out = tf.keras.layers.Add()(values) if self.activation is not None: out = tf.keras.layers.Activation(self.activation)(out) else: out = values[0] return out
[docs]class AddByProjecting(Operation): """Add operation. If tensors are of different shapes a projection will be applied before adding them. Args: search_space (KSearchSpace): [description]. Defaults to None. activation ([type], optional): Activation function to apply after adding ('relu', tanh', 'sigmoid'...). Defaults to None. stacked_nodes (list(Node)): nodes to add. axis (int): axis to concatenate. """ def __init__(self, search_space, stacked_nodes=None, activation=None, axis=-1): self.search_space = search_space self.node = None # current_node of the operation self.stacked_nodes = stacked_nodes self.activation = activation self.axis = axis
[docs] def init(self, current_node): self.node = current_node if self.stacked_nodes is not None: for n in self.stacked_nodes: self.search_space.connect(n, self.node)
def __call__(self, values, seed=None, **kwargs): # case where there is no inputs if len(values) == 0: return [] values = values[:] max_len_shp = max([len(x.get_shape()) for x in values]) # projection if len(values) > 1: for i, v in enumerate(values): if len(v.get_shape()) < max_len_shp: values[i] = tf.keras.layers.Reshape( ( *tuple(v.get_shape()[1:]), *tuple(1 for i in range(max_len_shp - len(v.get_shape()))), ) )(v) proj_size = values[0].get_shape()[self.axis] for i in range(len(values)): if values[i].get_shape()[1] != proj_size: values[i] = tf.keras.layers.Dense( units=proj_size, kernel_initializer=tf.keras.initializers.glorot_uniform( seed=seed ), )(values[i]) # concatenation if len(values) > 1: out = tf.keras.layers.Add()(values) if self.activation is not None: out = tf.keras.layers.Activation(self.activation)(out) else: out = values[0] return out