Deepcell

🐛 Bug Fixes

<details>
<summary>Add matplotlib to setup.py msschwartz21 (610)</summary>

What
* Add matplotlib requirement to setup.py

Why
* Pip installations used the requirements listed in setup.py so currently matplotlib is not installed when pip installing deepcell
</details>

<details>
<summary>Update mesmer post-processing args ngreenwald (609)</summary>

What
Updated the post-processing parameters for the Mesmer model. Also updates the notebook to describe how post-processing can be modified.

Why
The newly retrained model has different parameters that give the best results. In addition, I've gotten questions from a few different people about how to tweak the model output, having it in the notebook will make it easy for people to see the effects.

</details>

🚀 Features

<details>
<summary>Create TFRecords for tracking datasets vanvalen (602)</summary>

What
* Added functionality to create TFRecords for tracking datasets

Why
* As the training datasets grow in size, they are no longer able to fit in memory (as is the case with image generators). Adding functionality for TFRecords will let us train on larger datasets as they are loaded dynamically from disk during training rather than into memory all at once.

</details>


🐛 Bug Fixes

<details>
<summary>Fix tracking model bug that pinned n_filters, encoder_dim and embedding_dim to 64 vanvalen (606)</summary>

What
* Fixed a bug that required the tracking model to have n_filters, encoder_dim, and embedding_dim be pinned to 64

Why
* Model optimization is going to require us to change these parameters to improve performance and reduce model size. This pull request makes that substantially easier by fixing this bug.

</details>


🧰 Maintenance

<details>
<summary>Bump version to 0.12.1 msschwartz21 (605)</summary>


</details>

<details>
<summary>Expose option for fixed crops in the Track data object vanvalen (607)</summary>

What
* Modify the Track class so that it allows hooks into get_image_features for crop_mode and norm

Why
* Creating the appearance image feature by cropping and resizing removes information about cell size that the model can use to make more accurate tracking predictions. A previous update to deepcell-tracking (https://github.com/vanvalenlab/deepcell-tracking/pull/98) introduced the crop_mode (either 'fixed' or 'resize') and norm flags to get_image_features. This pull request exposes these flags to the Track class.

</details>

<details>
<summary>Bump `deepcell-tracking` to 0.6.0 msschwartz21 (603)</summary>

What
* Bump `deepcell-tracking` to the new minor release
* Update imports to match the reorganization introduced in this release

</details>

<details>
<summary>Update the docstring for `format_output_mesmer` curlup (601)</summary>

What
Doc-string for `format_output_mesmer` is now correctly saying "ValueError: if model output list is not len(4)"

Why
Because `format_output_mesmer` code diverged from the doc in what is expected length of model output list
</details>

🚀 Features

<details>
<summary>Introduce functions for reading and writing TF Records for segmentation data vanvalen (597)</summary>

What
Included functions to save datasets as tfrecords and load them into tf.data.Dataset objects

Why
As our training datasets grow, it is becoming difficult to load full datasets into memory. By introducing support for tfrecords, we can load portions of datasets from disk on the fly during training.

</details>

🧰 Maintenance

<details>
<summary>Update models after retraining on deepcell 0.12.0rc msschwartz21 (599)</summary>

What
* Update models with versions that were trained on tensorflow 2.8 (https://github.com/vanvalenlab/model-registry/pull/17)

Why
* Models should use the same version of tensorflow for predictions as they were trained on

</details>

<details>
<summary>Add option for either batch or layer norm in tracking model msschwartz21 (598)</summary>

What
* Provide the option to select either `BatchNormalization` or `LayerNormalization` in `GNNTrackingModel`

Why
* This option makes it possible to train the model with a batch size of 1 when layer normalization is enabled.

</details>

<details>
<summary>Update TF_VERSION build arg in docker build workflow msschwartz21 (596)</summary>

The TF_VERSION build arg has to be updated manually
</details>

<details>
<summary>Update Tensorflow to 2.8 msschwartz21 (595)</summary>

This PR updates tensorflow to 2.8 and drops support for python 3.6. The following changes were necessary to make this upgrade possible:
- Change imports from `tensorflow.python.keras` to `tensorflow.keras` which was a change introduced with tensorflow 2.6
- Remove convolutional recurrent layers and their functionality from featurenet and panopticnet. Key functions that were used in the convolutional recurrent layer are no longer available in keras.
- Change imports from `tensorflow.keras` to `keras`: `keras_parameterized`, `conv_utils`, `test_utils`
- Drop support for python 3.6

I retrained the nuclear model in the model-registry using this branch of deepcell and the performance was comparable.
</details>

<details>
<summary>Add option for either batch or layer norm in tracking model msschwartz21 (598)</summary>

What
* Provide the option to select either `BatchNormalization` or `LayerNormalization` in `GNNTrackingModel`

Why
* This option makes it possible to train the model with a batch size of 1 when layer normalization is enabled.

</details>

🧰 Maintenance

<details>
<summary>Update TF\_VERSION build arg in docker build workflow msschwartz21 (596)</summary>

The TF_VERSION build arg has to be updated manually
</details>

🧰 Maintenance

<details>
<summary>Update Tensorflow to 2.8 msschwartz21 (595)</summary>

This PR updates tensorflow to 2.8 and drops support for python 3.6. The following changes were necessary to make this upgrade possible:
- Change imports from `tensorflow.python.keras` to `tensorflow.keras` which was a change introduced with tensorflow 2.6
- Remove convolutional recurrent layers and their functionality from featurenet and panopticnet. Key functions that were used in the convolutional recurrent layer are no longer available in keras.
- Change imports from `tensorflow.keras` to `keras`: `keras_parameterized`, `conv_utils`, `test_utils`
- Drop support for python 3.6

I retrained the nuclear model in the model-registry using this branch of deepcell and the performance was comparable.

To dos before reviewing:

- [x] Fix requirements that reference github branches
- [x] Update setup.py with new requirements
- [x] Remove git installation from Dockerfile
</details>