list to torch

方法一：

直接转换

torch.tensor(pick, dtype=torch.long)

方法二：

numpy转换

pick = numpy.array(pick)
torch.from_numpy(pick).type(torch.LongTensor)

两种方法测试下来，在list长度过大的情况下，间接转换效率高；但在list长度少的情况下，使用直接转换效率高。

测试代码：

import random
import sys
import time
from datetime import datetime

import numpy
import torch

# 有gpu的情况下使用gpu
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')



total_time = 0
"""
测试模式，0表示直接转换，1表示使用numpy转换
"""
test_mode = 1
loop_time = 100
list_size = 64000
if test_mode == 0:

    for _ in range(loop_time):
        pick = [random.randint(0, 10000) for _ in range(list_size)]
        begin_time = time.perf_counter()
        torch.tensor(pick, dtype=torch.long).to(device)
        end_time = time.perf_counter()
        total_time += end_time - begin_time
else:
    for _ in range(loop_time):
        pick = [random.randint(0, 10000) for _ in range(list_size)]
        pick = numpy.array(pick)
        begin_time = time.perf_counter()
        torch.from_numpy(pick).type(torch.LongTensor).to(device)
        end_time = time.perf_counter()
        total_time += end_time - begin_time
print("%.5fms" % (total_time * 1000.0))