def recv(dst, uniq_dst, bucketing_index, msg, reduce_function, node_ids):
"""Recv message from given msg to dst nodes.
"""
if reduce_function == "sum":
if isinstance(msg, dict):
raise TypeError("The message for build-in function"
" should be Tensor not dict.")
try:
out_dims = msg.shape[-1]
init_output = fluid.layers.fill_constant_batch_size_like(
node_ids, shape=[1, out_dims], value=0, dtype="float32")
init_output.stop_gradient = False
output = paddle_helper.scatter_add(init_output, dst, msg)
return output
except TypeError as e:
warnings.warn(
"scatter_add is not supported with paddle version <= 1.5")
def sum_func(message):
return fluid.layers.sequence_pool(message, "sum")
reduce_function = sum_func
# convert msg into lodtensor
bucketed_msg = op.nested_lod_reset(msg, bucketing_index)
# Check dim for bucketed_msg equal to out_dims
output = reduce_function(bucketed_msg)
out_dims = output.shape[-1]
init_output = fluid.layers.fill_constant_batch_size_like(
node_ids, shape=[1, out_dims], value=0, dtype="float32")
init_output.stop_gradient = False
output = fluid.layers.scatter(init_output, uniq_dst, output)
return output
def recv(dst, uniq_dst, bucketing_index, msg, reduce_function, num_nodes,
num_edges):
"""Recv message from given msg to dst nodes.
"""
if reduce_function == "sum":
if isinstance(msg, dict):
raise TypeError("The message for build-in function"
" should be Tensor not dict.")
try:
out_dim = msg.shape[-1]
init_output = L.fill_constant(shape=[num_nodes, out_dim],
value=0,
dtype=msg.dtype)
init_output.stop_gradient = False
empty_msg_flag = L.cast(num_edges > 0, dtype=msg.dtype)
msg = msg * empty_msg_flag
output = paddle_helper.scatter_add(init_output, dst, msg)
return output
except TypeError as e:
warnings.warn(
"scatter_add is not supported with paddle version <= 1.5")
def sum_func(message):
return L.sequence_pool(message, "sum")
reduce_function = sum_func
bucketed_msg = op.nested_lod_reset(msg, bucketing_index)
output = reduce_function(bucketed_msg)
output_dim = output.shape[-1]
empty_msg_flag = L.cast(num_edges > 0, dtype=output.dtype)
output = output * empty_msg_flag
init_output = L.fill_constant(shape=[num_nodes, output_dim],
value=0,
dtype=output.dtype)
init_output.stop_gradient = True
final_output = L.scatter(init_output, uniq_dst, output)
return final_output
