fizyr / keras-retinanet

Keras implementation of RetinaNet object detection.
Apache License 2.0
4.38k stars 1.96k forks source link

Update effnet.py #1448

Open karthikayan4u opened 4 years ago

karthikayan4u commented 4 years ago

I guess as keras has been updated with tensorflow, efficientnet.keras is no longer available and hence I am updating this line.

ha-nso-li commented 4 years ago

I think effnet.py used efficientnet pip package, neither core of keras or tensorflow. Is the implementation of tf.keras compatible?

karthikayan4u commented 4 years ago

I think effnet.py used efficientnet pip package, neither core of keras or tensorflow. Is the implementation of tf.keras compatible?

Actually there is a problem. We can't use pretrained weights as tf comes with weights itself and I don't know how to integrate the imagenet weights. Rest of all is fine.

hgaiser commented 4 years ago

What issue are you having? I'm guessing it is related to https://github.com/qubvel/efficientnet/issues/127 ? This should be fixed with https://github.com/qubvel/efficientnet/pull/128 . When that gets released, efficientnet should work with the latest tensorflow again.

karthikayan4u commented 4 years ago

Yeah the same was the issue along with another..I couldn't use the pretrained iamgenet weights too!

gosha20777 commented 4 years ago

I use the qubvel/efficientnet package and personally know the author of this package. This package works nicely and is well compatible with the latest TF versions. I believe there is no need to replace this package. Plus it breaks compatibility.

gosha20777 commented 4 years ago

In addition, I think it would be nice to replace the original resnet with the resnet from this package. This can solve many problems and will allow you to abandon the keras code and completely move to TF.