Skip to content

GitLab

Switch branch/tag
  1. 27 Sep, 2020 1 commit
    • Shixin's avatar
      [MobileNet] Add Mobilenet Backbone Implementation (#9303) · e61588cd
      Shixin authored 
      

* factor make_divisible function and move round_filters to nn_layers

* modify SqueezeExcitation to add two additional parameter: divisible_by and gating_activation

* modify the InvertedBottleneckBlock to include 1. use_depthwise, 2. use_residual, 3. regularize_depthwise additional boolean flag; Add control for depthwise activation and regularizer; remove expand_ratio from SqueezeExcitation

* add Conv2DBNBlock definition

* add mobilenet v2, v3 implementation

* add mobilenet v1

* put mobilenet_base into class body

* fix a type hint error

* the invertedbottlenetblock is different for mobilenet and efficientnet. Made necessary changes to cope both.

* add target_backbone while call invertedbottleneckblock

* add relu6 and hard_sigmoid

* add test for mobilenet

* add mobilenet to factory

* fix some typo; link the reference to the architectures

* remove future import
Co-authored-by: default avatarShixin Luo <luoshixin@google.com>
      e61588cd
  3. 08 Sep, 2020 1 commit
  5. 04 Sep, 2020 1 commit
  7. 03 Sep, 2020 1 commit
  9. 02 Sep, 2020 2 commits
  11. 27 Aug, 2020 1 commit
  13. 26 Aug, 2020 4 commits
  15. 13 Aug, 2020 2 commits
  17. 12 Aug, 2020 2 commits
  19. 05 Aug, 2020 6 commits
  21. 04 Aug, 2020 2 commits
  23. 03 Aug, 2020 2 commits
  25. 31 Jul, 2020 3 commits
  27. 30 Jul, 2020 2 commits
    • Tomer Kaftan's avatar
      Pre-emptively disable the KerasTensors refactoring for the detection models in... · 071bb243
      Tomer Kaftan authored 
      Pre-emptively disable the KerasTensors refactoring for the detection models in tensorflow models/official/vision/detection, because they rely on several unsupported things that will stop working entirely when the refactoring goes live.

Specifically:
* The custom layers implement `__call__` instead of `call` and rely on manually enter the keras backend graph
* The vision models try to use `tf.while_loop` as Keras op layers during functional API construction, which is unsupported.

Updating the models to avoid this would subtly change the variable names and break the pre-existing tf1-style name-based checkpoints, so for now we will just disable the KerasTensors refactoring for these models.

PiperOrigin-RevId: 323937426
      071bb243
    • Tomer Kaftan's avatar
      Pre-emptively disable the KerasTensors refactoring for the detection models in... · ccc60760
      Tomer Kaftan authored 
      Pre-emptively disable the KerasTensors refactoring for the detection models in tensorflow models/official/vision/detection, because they rely on several unsupported things that will stop working entirely when the refactoring goes live.

Specifically:
* The custom layers implement `__call__` instead of `call` and rely on manually enter the keras backend graph
* The vision models try to use `tf.while_loop` as Keras op layers during functional API construction, which is unsupported.

Updating the models to avoid this would subtly change the variable names and break the pre-existing tf1-style name-based checkpoints, so for now we will just disable the KerasTensors refactoring for these models.

PiperOrigin-RevId: 323937426
      ccc60760
  29. 24 Jul, 2020 1 commit
    • Srihari Humbarwadi's avatar
      Fixed typos · b3f63b00
      Srihari Humbarwadi authored 
       - Fixed `intermediate_scale` in  `Anchor`
 - "ratio" and "divisible" in doc string
      b3f63b00
  31. 20 Jul, 2020 2 commits
    • Tomer Kaftan's avatar
      Make hack in official/vision/detection models that enters the backend keras... · 363a36cd
      Tomer Kaftan authored 
      Make hack in official/vision/detection models that enters the backend keras graph stop happening once we enable the Functional API KerasTensors refactoring:

As a workaround hack for the tf op layer conversion being fragile, the detection models have to explicitly enter the Keras backend graph.

When we enable the KerasTensors refactoring of the Functional API internals, the op layer conversion will be much more reliable and this hack will not be necessary. In addition, the hack actually causes the models to break when we enable the refactoring (because it causes tensors to leak out of a graph).

So, this CL changes the existing hack to stop applying once we've enabled the KerasTensors refactoring.

PiperOrigin-RevId: 322229802
      363a36cd
    • Tomer Kaftan's avatar
      Make hack in official/vision/detection models that enters the backend keras... · ae8845ce
      Tomer Kaftan authored 
      Make hack in official/vision/detection models that enters the backend keras graph stop happening once we enable the Functional API KerasTensors refactoring:

As a workaround hack for the tf op layer conversion being fragile, the detection models have to explicitly enter the Keras backend graph.

When we enable the KerasTensors refactoring of the Functional API internals, the op layer conversion will be much more reliable and this hack will not be necessary. In addition, the hack actually causes the models to break when we enable the refactoring (because it causes tensors to leak out of a graph).

So, this CL changes the existing hack to stop applying once we've enabled the KerasTensors refactoring.

PiperOrigin-RevId: 322229802
      ae8845ce
  33. 17 Jul, 2020 2 commits
  35. 15 Jul, 2020 2 commits
  37. 13 Jul, 2020 2 commits
  39. 01 Jul, 2020 1 commit