Modification of the definition of networks#18
Conversation
| if self.blocks >= 5: | ||
| net = layers.repeat(net, 3, layers.conv2d, 512, [3, 3], scope='conv5') | ||
| intermediate_levels.append(net) | ||
| net = layers.max_pool2d(net, [2, 2], scope='pool5') |
There was a problem hiding this comment.
The output of the last max_pool2d is not returned, shouldn't it be ?
There was a problem hiding this comment.
mmhh... good catch. In a way, it depends what we want to do with the model definition, it can make sense to keep the feature maps with the best resolution but the lower receptive field. But I think if we want to be consistent with what we had, outputs should have the pooled versions, thoughts?
|
|
||
| @classmethod | ||
| def parse(cls, prediction_type): | ||
| def parse(cls, prediction_type) -> 'PredictionType': |
There was a problem hiding this comment.
Unrelated to this diff, but I wonder if we use Enum we don't get a parse method for free?
|
Uh why close it?
Le lun. 5 nov. 2018 à 09:40, solivr <notifications@github.com> a écrit :
… Closed #18 <#18>.
—
You are receiving this because you were assigned.
Reply to this email directly, view it on GitHub
<#18 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/AGzWkTKqx6tNRmO80Gy6eUx2yGarkEHJks5ur_l9gaJpZM4XZ-ok>
.
|
|
oops sorry I closed |
doc : Correcting typo in masks creation script (annotating)
Fixing instruction
weights in IOU evaluation
|
Is it recommended to use this branch instead of master? It seems to have a lot of fixes and extensions... |
|
I just wanted to note I was unable to run demo.py on the master branch or the latest release tag. It would fail with a ValueError. That being said, demo.py ran just fine on this encoder_revamp branch. It's somewhat confusing to a newcomer like me that this isn't noted anywhere. |
|
A lot of the changes suggested here are actually part of the |
|
@wrznr The official branch is the |
|
To pursue on @solivr comment, we do not have anybody continuing the development at the moment ( in fact I have not been part of the lab anymore for some time already). A lot of the commits are actually merged from master though so it might appear that there is much more modifications than there are actually. |
|
We did a bit of merging today, things left to correct.
|
This branch seems ready to be merged with |
6 participants
So now the network definition is based on two bases classes, an
Encoderand aDecoder.This simplifies quite a bit of things, and allows more flexibility. Also, it will allows anyone to plug-in its own network definition without having to clone dh_segment.