search

shicai / MobileNet-Caffe

Caffe Implementation of Google's MobileNets (v1 and v2)
BSD 3-Clause "New" or "Revised" License
1.26k stars 708 forks source link

problem in conv4_7 layer #50

Closed linchaozhang closed 6 years ago

linchaozhang commented 6 years ago

for mobilenet _v2, stride=1 for conv4_7 in paper

shicai commented 6 years ago

there may be some typos in Table 2 in the paper, so it is not clear when to use stride=2. if you take a look at colomn 1, spatial size is reduced from 28 to 14, so there should be stride=2 used here.

linchaozhang commented 6 years ago

@shicai got it, there may be some typos in Table 2, you take input_size as a standard, use stride = 1 in conv4_3, thank you！

wujiyang commented 6 years ago

@linchaozhang @shicai there is a little difference in Table 2 between the arXiv paper v1 and the latest v3. In paper v3, it should be stride=1 for conv4_7 and stride=2 for conv4_3.

leochli commented 6 years ago

Hi, @wujiyang Thanks for pointing out the differences! @shicai The paper says "The first layer of each sequence has a stride s and all others use stride 1." Yet it seems that you are doing down sampling at the end of each phase. Could you explain for that?

shicai commented 6 years ago

they share the same architecture, but are named in different ways. @leochli