There is no 'upsample' in concat layer in PVAnet_train?

[sleepfin]

I notice some difference between PVAnet_train and PVAnet_test, which makes the Tensor does not match when loading a training ckpt to a test model.

PVAnet_train.py -> line 96 (self.feed('downsample', 'conv4_4').concat(axis=3, name='concat')) PAVnet_test.py -> line 92 (self.feed('downsample', 'conv4_4', 'upsample').concat(axis=3, name='concat'))

Train: Tensor("concat:0", shape=(?, ?, ?, 384), dtype=float32) Test: Tensor("concat:0", shape=(?, ?, ?, 768), dtype=float32)

According to the paper, I think the missing 'upsample' in PVAnet_train is necessary.

AND

PAVnet_test.py -> line 85 .scale(c_in=384, name='conv5_4/last_bn_scale') which is also missing in PVAnet_train.py (Between line 89 and 90)

AND

In PVAnet_train.py line 130-133

# (self.feed('convf_rpn', 'convf_2')
# .concat(axis=3, name='convf'))

(self.feed('convf_2', 'roi-data')

In PVAnet_test.py line 118-121

(self.feed('convf_rpn', 'convf_2')
.concat(axis=3, name='convf'))

(self.feed('convf', 'rois')

But Y??