Faster R-CNN training in TensorFlow < 1.4 ?

[csnemes2]

Can I retrain Faster R-CNN with TF 1.3?
Evaluation worked perfectly with TF1.3, and I can confirm your published results.

BUT, when I train I run into an assert at https://github.com/ppwwyyxx/tensorpack/blob/master/tensorpack/models/batch_norm.py

if use_local_stat:
blabla
else:
if ctx.is_training:
assert get_tf_version_number() >= 1.4,
"Fine tuning a BatchNorm model with fixed statistics is only "
"supported after tensorflow/tensorflow#12580 "

Without pain, Is there any chance to set use_local_stat to True?
According to my quick check

• Faster RCNN relies on BNReLUs (see https://github.com/ppwwyyxx/tensorpack/blob/master/tensorpack/models/nonlin.py)
• BNReLU defaults use_local_stat to None
• I can modify this but still hesitating, would it work? Any suggestion?

[ppwwyyxx]

You cannot. That's why it requires TF>=1.4

[csnemes2]

Thanks for the quick and clear answer, I am already downloading the 1.4 :)
Any chance to update the readme?
I have just checked it and no hints, just something tf>=1.2 which is true but not very tight :)

[machanic]

tf 1.5 has already released , which fully support dynamic graph( like in PyTorch), which will ease the burden of debug breakpoint tf. Will tensorpack supply a switch argument to open dynamic graph let us to easy debug?

[ppwwyyxx]

Any chance to update the readme?

The readme is correct. Tensorpack depends on TF>=1.2; FasterRCNN depends on TF>=1.4. Both are mentioned in the corresponding readme.

@sharpstill I'll look at that in the future at #463.