权重11m,速度35ms
内存小,1060可以20个416的图片
p100 batch可以32 与pelee可以匹配。
这个作者说比原版好,caffe的:
Model VOC2007 mAP (%) VOC2012 mAP (%)
PVANet+ (VOC2007) 84.9 N/A
PVANet+ (VOC2012) 89.8 84.2
PVANet+ (VOC2012 + compressed) 87.8 83.7
https://github.com/JohnHush/pvanet_modified
import time
import torch
import torchvision
import torch.nn as nn
import torch.nn.functional as F
from utils import *
# XXX: Not tested
class CReLU(nn.Module):
def __init__(self, act=F.relu):
super(CReLU, self).__init__()
self.act = act
def