from ._model import model_disjoint
from kgcnn.models.utils import update_model_kwargs
from kgcnn.layers.scale import get as get_scaler
from kgcnn.layers.modules import Input
from kgcnn.models.casting import (template_cast_output, template_cast_list_input,
template_cast_list_input_docs, template_cast_output_docs)
from keras.backend import backend as backend_to_use
from kgcnn.ops.activ import *
# Keep track of model version from commit date in literature.
# To be updated if model is changed in a significant way.
__model_version__ = "2023-09-18"
# Supported backends
__kgcnn_model_backend_supported__ = ["tensorflow", "torch", "jax"]
if backend_to_use() not in __kgcnn_model_backend_supported__:
raise NotImplementedError("Backend '%s' for model 'GATv2' is not supported." % backend_to_use())
# Implementation of GATv2 in `keras` from paper:
# Graph Attention Networks
# by Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, Yoshua Bengio (2018)
# https://arxiv.org/abs/1710.10903
# Improved by
# How Attentive are Graph Attention Networks?
# by Shaked Brody, Uri Alon, Eran Yahav (2021)
# https://arxiv.org/abs/2105.14491
model_default = {
'name': "GATv2",
'inputs': [
{"shape": (None,), "name": "node_attributes", "dtype": "float32"},
{"shape": (None,), "name": "edge_attributes", "dtype": "float32"},
{"shape": (None, 2), "name": "edge_indices", "dtype": "int64"},
{"shape": (), "name": "total_nodes", "dtype": "int64"},
{"shape": (), "name": "total_edges", "dtype": "int64"}
],
"input_tensor_type": "padded",
"input_embedding": None, # deprecated
"cast_disjoint_kwargs": {},
"input_node_embedding": {"input_dim": 95, "output_dim": 64},
"input_edge_embedding": {"input_dim": 5, "output_dim": 64},
'attention_args': {"units": 32, "use_final_activation": False, "use_edge_features": True,
"has_self_loops": True, "activation": "kgcnn>leaky_relu", "use_bias": True},
"pooling_nodes_args": {"pooling_method": "scatter_mean"},
'depth': 3, 'attention_heads_num': 5,
'attention_heads_concat': False, 'verbose': 10,
'output_embedding': 'graph',
"output_to_tensor": None, # deprecated
"output_tensor_type": "padded",
'output_mlp': {"use_bias": [True, True, False], "units": [25, 10, 1],
"activation": ['relu', 'relu', 'sigmoid']},
"output_scaling": None,
}
[docs]@update_model_kwargs(model_default, update_recursive=0, deprecated=["input_embedding", "output_to_tensor"])
def make_model(inputs: list = None,
input_tensor_type: str = None,
cast_disjoint_kwargs: dict = None,
input_embedding: dict = None, # noqa
input_node_embedding: dict = None,
input_edge_embedding: dict = None,
attention_args: dict = None,
pooling_nodes_args: dict = None,
depth: int = None,
attention_heads_num: int = None,
attention_heads_concat: bool = None,
name: str = None,
verbose: int = None, # noqa
output_embedding: str = None,
output_to_tensor: bool = None, # noqa
output_mlp: dict = None,
output_scaling: dict = None,
output_tensor_type: str = None,
):
r"""Make `GATv2 <https://arxiv.org/abs/2105.14491>`__ graph network via functional API.
Default parameters can be found in :obj:`kgcnn.literature.GATv2.model_default`.
**Model inputs**:
Model uses the list template of inputs and standard output template.
The supported inputs are :obj:`[nodes, edges, edge_indices, ...]`
with '...' indicating mask or ID tensors following the template below:
%s
**Model outputs**:
The standard output template:
%s
Args:
inputs (list): List of dictionaries unpacked in :obj:`Input`. Order must match model definition.
cast_disjoint_kwargs (dict): Dictionary of arguments for casting layers if used.
input_tensor_type (str): Input type of graph tensor. Default is "padded".
input_embedding (dict): Deprecated in favour of input_node_embedding etc.
input_node_embedding (dict): Dictionary of arguments for nodes unpacked in :obj:`Embedding` layers.
input_edge_embedding (dict): Dictionary of arguments for edge unpacked in :obj:`Embedding` layers.
attention_args (dict): Dictionary of layer arguments unpacked in :obj:`AttentionHeadGATV2` layer.
pooling_nodes_args (dict): Dictionary of layer arguments unpacked in :obj:`PoolingNodes` layer.
depth (int): Number of graph embedding units or depth of the network.
attention_heads_num (int): Number of attention heads to use.
attention_heads_concat (bool): Whether to concat attention heads, or simply average heads.
name (str): Name of the model.
verbose (int): Level of print output.
output_embedding (str): Main embedding task for graph network. Either "node", "edge" or "graph".
output_to_tensor (bool): Deprecated in favour of `output_tensor_type` .
output_mlp (dict): Dictionary of layer arguments unpacked in the final classification :obj:`MLP` layer block.
Defines number of model outputs and activation.
output_scaling (dict): Dictionary of layer arguments unpacked in scaling layers. Default is None.
output_tensor_type (str): Output type of graph tensors such as nodes or edges. Default is "padded".
Returns:
:obj:`keras.models.Model`
"""
# Make input
model_inputs = [Input(**x) for x in inputs]
dj_model_inputs = template_cast_list_input(
model_inputs,
input_tensor_type=input_tensor_type,
cast_disjoint_kwargs=cast_disjoint_kwargs,
mask_assignment=[0, 1, 1],
index_assignment=[None, None, 0]
)
n, ed, disjoint_indices, batch_id_node, batch_id_edge, node_id, edge_id, count_nodes, count_edges = dj_model_inputs
# Wrapping disjoint model.
out = model_disjoint(
[n, ed, disjoint_indices, batch_id_node, count_nodes],
use_node_embedding=("int" in inputs[0]['dtype']) if input_node_embedding is not None else False,
use_edge_embedding=("int" in inputs[1]['dtype']) if input_edge_embedding is not None else False,
input_node_embedding=input_node_embedding, input_edge_embedding=input_edge_embedding,
attention_args=attention_args, pooling_nodes_args=pooling_nodes_args, depth=depth,
attention_heads_num=attention_heads_num, attention_heads_concat=attention_heads_concat,
output_embedding=output_embedding, output_mlp=output_mlp
)
if output_scaling is not None:
scaler = get_scaler(output_scaling["name"])(**output_scaling)
out = scaler(out)
# Output embedding choice
out = template_cast_output(
[out, batch_id_node, batch_id_edge, node_id, edge_id, count_nodes, count_edges],
output_embedding=output_embedding, output_tensor_type=output_tensor_type,
input_tensor_type=input_tensor_type, cast_disjoint_kwargs=cast_disjoint_kwargs,
)
# Define model output
model = ks.models.Model(inputs=model_inputs, outputs=out, name=name)
model.__kgcnn_model_version__ = __model_version__
if output_scaling is not None:
def set_scale(*args, **kwargs):
scaler.set_scale(*args, **kwargs)
setattr(model, "set_scale", set_scale)
return model
make_model.__doc__ = make_model.__doc__ % (template_cast_list_input_docs, template_cast_output_docs)