有可能在火把中运行散点剂吗？

Question

编辑:显然DGL已经在做了：https://github.com/dmlc/dgl/pull/3641

我有几种类型的嵌入，每一种都需要自己的线性投影。我可以使用类型为for的循环来解决这个问题：

emb_out = dict()
for ntype in ntypes:
    emb_out[ntype] = self.lin_layer[ntype](emb[ntype])

但理想情况下，我想做一些分散操作来并行运行它。类似于：

pytorch_scatter(lin_layers, embeddings, layer_map, reduce='matmul')，图层映射指示应该通过哪个层进行嵌入。如果我有两种类型的线性层，并且batch_size = 5，那么layer_map应该是1，0-1，1，0。

是否有可能像在散落中那样以一种高效的方式将for循环矢量化？请检查下面的最低限度的例子。

import torch
import random
import numpy as np 

seed = 42
torch.manual_seed(seed)
random.seed(seed)

def matmul_single_embtype(lin_layers, embeddings, layer_map):
   #run single linear layer over all embeddings, irrespective of type
   output_embeddings = torch.matmul(lin_layers[0], embeddings.T).T
   return output_embeddings

def matmul_for_loop(lin_layers, embeddings, layer_map):
   #let each embedding type have its own projection, looping over emb types
   output_embeddings = dict()
   for emb_type in np.unique(layer_map):
       output_embeddings[emb_type] = torch.matmul(lin_layers[emb_type], embeddings[layer_map == emb_type].T).T
   return output_embeddings

def matmul_scatter(lin_layers, embeddings, layer_map):
   #parallelize the for loop by creating a diagonal matrix of lin layers
   #this is very innefficient, because creates a copy of the layer for each embedding, instead of broadcasting
   mapped_lin_layers = [lin_layers[i] for i in layer_map]
   mapped_lin_layers = torch.block_diag(*mapped_lin_layers) #batch_size*inp_size x batch_size*output_size
   embeddings_stacked = embeddings.view(-1,1) #stack all embeddings to multiply the linear block
   output_embeddings = torch.matmul(mapped_lin_layers, embeddings_stacked).view(embeddings.shape)
   return output_embeddings

"""
GENERATE DATA
lin_layers:
   List of matrices of size n_layer x inp_size x output_size
embeddings:
   Matrix of size batch_size x inp_size
layer_map:
   Vector os size batch_size stating which embedding should go thorugh each layer
"""

emb_size = 32
batch_size = 500
emb_types = 20
layer_map = [random.choice(list(range(emb_types))) for i in range(batch_size)]

lin_layers = [torch.arange(emb_size*emb_size, dtype=torch.float32).view(emb_size,emb_size) for i in range(emb_types)]
embeddings = torch.arange(batch_size*emb_size, dtype=torch.float32).view(batch_size,emb_size)
grouped_emb = {i: embeddings[layer_map==i] for i in np.unique(layer_map)} #separate embeddings by embedding type

#Run experiments
%timeit matmul_scatter(lin_layers, embeddings, layer_map)
%timeit matmul_for_loop(lin_layers, embeddings, layer_map)
%timeit matmul_single_embtype(lin_layers, embeddings, layer_map)

>>>>>133 ms ± 2.47 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
>>>>>1.64 ms ± 14 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
>>>>>31.4 µs ± 805 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)

相关堆栈溢出问题：离散矩阵运算的矢量化方法

火炬手相关问题：https://github.com/pytorch/pytorch/issues/31942

Answer 1

刚刚发现DGL已经在处理这个特性了：https://github.com/dmlc/dgl/pull/3641