Questions about LSTR's performance.

I just tested the performance，and the result seems not that fast，just about 100fps. Is there something wrong with the testing method？Here’s the testing code below.

speed_simple

import torch
import time
import numpy as np
from nnet.py_factory import NetworkFactory
from config import system_configs
import os
import json

cfg_file = os.path.join(system_configs.config_dir, "LSTR.json")
print("cfg_file: {}".format(cfg_file))
with open(cfg_file, "r") as f:
    configs = json.load(f)

configs["system"]["snapshot_name"] = "LSTR"
system_configs.update_config(configs["system"])

torch.cuda.set_device(0)
torch.backends.cudnn.benchmark = True
nnet = NetworkFactory()
nnet.load_params(500000)
nnet.cuda()
nnet.eval_mode()

input_size = [360, 640]
images = torch.zeros((1, 3, input_size[0], input_size[1]), dtype=torch.float32).cuda() + 1
masks = torch.ones((1, 1, input_size[0], input_size[1]), dtype=torch.float32).cuda() + 2

for i in range(10):
    outputs, weights = nnet.test([images, masks])

t_all = []
for i in range(200):
    torch.cuda.synchronize(0)
    t1 = time.time()
    outputs, weights = nnet.test([images, masks])
    torch.cuda.synchronize(0)
    t2 = time.time()
    t_all.append(t2 - t1)

print('average time:', np.mean(t_all) / 1)
print('average fps:', 1 / np.mean(t_all))

print('fastest time:', min(t_all) / 1)
print('fastest fps:', 1 / min(t_all))

print('slowest time:', max(t_all) / 1)
print('slowest fps:', 1 / max(t_all))

[nigyiii]

我佛了

[liuruijin17]

Hi, thanks for you understanding. I will update the arxiv to clarify batch=1’s FPS and batch=16’s FPS on the Table.