You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
TensorFlow version (2.3): Although this request feels very TF 2.4.
Are you willing to contribute it: I could be involved in testing.
Describe the feature and the current behavior/state.
tf.keras.applications is amazing. With all the model architectures and the pre-trained weights there is a lot
of possibilities for transfer learning and feature extraction. However, there are not any object detection models.
I would love to see some object detection models baked into tf.keras.applications as well as pre-trained weights (trained on object detection tasks like coco for example) that would make it possible for training new object detection models, via transfer learning, on custom data. Very much like how the current API works for classification models for example i.e. (import resnet, remove top/head, add layer, freeze, train).
Since we already have Resnet backbones and Efficient backbones (as of TF 2.3), would it be possible to add object detection architectures on top of these such as Retinanet or EfficientDet. As well as have the pre-trained weights trained on coco for example for such models. Then having some sort of easy way to adjust these for training on custom object detection data. Much like with the current models we can remove the top/head and add on top for custom classifiers.
It's not straight forward to get started with object detection using tf.keras. I am also aware of the newly designed TF Models Detection API but it still feels bloated and very disconnected from TF main library. It would be lovely to see something built into tf.keras.applications.
Will this change the current api? How?
Just by adding the new feature request.
Who will benefit with this feature?
All users of tf.keras.applications. Anyone working in industry with current TF 2 in their tech stack looking to add object detection functionality.
Any Other info.
I think this would be amazing!
There is the project https://github.com/google/automl/tree/master/efficientdet which is Google and it seems like they are trying to use tf.keras there some. There is also https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2.md but it has so much else going on in it. How can we take these projects and pick out the pieces and build for example EfficientDet into tf.keras.applications. At the moment things are convoluted and lots of outside projects. Having this within tf.keras.applications main code would be nice and easier to build off in existing tech stacks that already use Tensorflow.
The text was updated successfully, but these errors were encountered:
Hello @ymodak, I would like to work on this issue. I am new to this but this can help me learn alot.
Can you tell where I have to add this in your code base?
@aavishkarmishra , love the enthusiasm but I think this is a good project for internal Google developers and core tf.keras team developers. It’s a fairly big ask but one that I think would have tons of value.
System information
Describe the feature and the current behavior/state.
tf.keras.applications
is amazing. With all the model architectures and the pre-trained weights there is a lotof possibilities for transfer learning and feature extraction. However, there are not any object detection models.
I would love to see some object detection models baked into
tf.keras.applications
as well as pre-trained weights (trained on object detection tasks like coco for example) that would make it possible for training new object detection models, via transfer learning, on custom data. Very much like how the current API works for classification models for example i.e. (import resnet, remove top/head, add layer, freeze, train).Since we already have Resnet backbones and Efficient backbones (as of TF 2.3), would it be possible to add object detection architectures on top of these such as Retinanet or EfficientDet. As well as have the pre-trained weights trained on coco for example for such models. Then having some sort of easy way to adjust these for training on custom object detection data. Much like with the current models we can remove the top/head and add on top for custom classifiers.
It's not straight forward to get started with object detection using tf.keras. I am also aware of the newly designed TF Models Detection API but it still feels bloated and very disconnected from TF main library. It would be lovely to see something built into tf.keras.applications.
Will this change the current api? How?
Just by adding the new feature request.
Who will benefit with this feature?
All users of tf.keras.applications. Anyone working in industry with current TF 2 in their tech stack looking to add object detection functionality.
Any Other info.
I think this would be amazing!
There is the project https://github.com/google/automl/tree/master/efficientdet which is Google and it seems like they are trying to use tf.keras there some. There is also https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2.md but it has so much else going on in it. How can we take these projects and pick out the pieces and build for example EfficientDet into tf.keras.applications. At the moment things are convoluted and lots of outside projects. Having this within tf.keras.applications main code would be nice and easier to build off in existing tech stacks that already use Tensorflow.
The text was updated successfully, but these errors were encountered: