-
Notifications
You must be signed in to change notification settings - Fork 7
Description
Hi,
Your publication is interesting and I would like to reproduce result especially about FPS.
This is my code to measure the FPS
from core.models.cabinet import CABiNet
import torch
import time
import numpy as np
from ptflops import get_model_complexity_info
net = CABiNet(n_classes=19, backbone_weights="core/models/pretrained_backbones/mobilenetv3-small-55df8e1f.pth")
net.cuda()
net.eval()
time_avg = []
for i in range(0,120):
# imgs = imgs.cuda()
imgs = torch.randn(1, 3, 1024, 1024, requires_grad=False).cuda()
print("imgs size: ",imgs.shape)
start_time = time.time()
out = net(imgs)[0]
torch.cuda.synchronize()
end_time = time.time()
print("out size: ",out.shape)
print("time: ", end_time-start_time)
time_avg.append(end_time-start_time)
print("avg time: ", np.mean(time_avg[10:-10]))
I was wondering that it is about 8fps with 2048x1024 resolution.
However, with my device, I cannot reproduce the result of FPS with 1024x2048 resolution and can measure the FPS of 8 fps 1024x1024 image resolution.
Could you help the measurement codes?