chainer_chemistry.models.RelGAT

class chainer_chemistry.models.RelGAT(out_dim, hidden_dim=16, n_heads=3, negative_slope=0.2, n_edge_types=4, n_layers=4, dropout_ratio=-1.0, activation=<function identity>, n_atom_types=117, softmax_mode='across', concat_hidden=False, concat_heads=False, weight_tying=False)

Relational Graph Attention Networks (GAT)

See: Veličković, Petar, et al. (2017). Graph Attention Networks. arXiv:1701.10903 <https://arxiv.org/abs/1710.10903> Dan Busbridge, et al. (2018). Relational Graph Attention Networks

<https://openreview.net/forum?id=Bklzkh0qFm>

Parameters:

• out_dim (int) – dimension of output feature vector
• hidden_dim (int) – dimension of feature vector associated to each atom
• n_layers (int) – number of layers
• n_atom_types (int) – number of types of atoms
• n_heads (int) – number of multi-head-attentions.
• n_edge_types (int) – number of edge types.
• dropout_ratio (float) – dropout ratio of the normalized attention coefficients
• negative_slope (float) – LeakyRELU angle of the negative slope
• softmax_mode (str) – take the softmax over the logits ‘across’ or ‘within’ relation. If you would like to know the detail discussion, please refer Relational GAT paper.
• concat_hidden (bool) – If set to True, readout is executed in each layer and the result is concatenated
• concat_heads (bool) – Whether to concat or average multi-head attentions
• weight_tying (bool) – enable weight_tying or not

__init__(out_dim, hidden_dim=16, n_heads=3, negative_slope=0.2, n_edge_types=4, n_layers=4, dropout_ratio=-1.0, activation=<function identity>, n_atom_types=117, softmax_mode='across', concat_hidden=False, concat_heads=False, weight_tying=False)

Initialize self. See help(type(self)) for accurate signature.

Methods

__init__(out_dim[, hidden_dim, n_heads, …])	Initialize self.
add_hook(hook[, name])	Registers a link hook.
add_link(name, link)	Registers a child link to this chain.
add_param(name[, shape, dtype, initializer])	Registers a parameter to the link.
add_persistent(name, value)	Registers a persistent value to the link.
addgrads(link)	Accumulates gradient values from given link.
children()	Returns a generator of all child links.
cleargrads()	Clears all gradient arrays.
copy([mode])	Copies the link hierarchy to new one.
copyparams(link[, copy_persistent])	Copies all parameters from given link.
count_params()	Counts the total number of parameters.
delete_hook(name)	Unregisters the link hook.
disable_update()	Disables update rules of all parameters under the link hierarchy.
enable_update()	Enables update rules of all parameters under the link hierarchy.
init_scope()	Creates an initialization scope.
links([skipself])	Returns a generator of all links under the hierarchy.
namedlinks([skipself])	Returns a generator of all (path, link) pairs under the hierarchy.
namedparams([include_uninit])	Returns a generator of all (path, param) pairs under the hierarchy.
params([include_uninit])	Returns a generator of all parameters under the link hierarchy.
register_persistent(name)	Registers an attribute of a given name as a persistent value.
repeat(n_repeat[, mode])	Repeats this link multiple times to make a Sequential.
serialize(serializer)	Serializes the link object.
to_cpu()	Copies parameter variables and persistent values to CPU.
to_gpu([device])	Copies parameter variables and persistent values to GPU.
to_intel64()	Copies parameter variables and persistent values to CPU.
zerograds()	Initializes all gradient arrays by zero.

Attributes

local_link_hooks	Ordered dictionary of registered link hooks.
update_enabled	True if at least one parameter has an update rule enabled.
within_init_scope	True if the current code is inside of an initialization scope.
xp	Array module for this link.