* refactor! more flexible and backwards-compatible with lobe signature files. tf_backend uses the tf 1.15 session and saved model loader rather than the contrib inference library.

* update to tensorflow 2.4

* add onnx backend!

* add note about libjpeg62-turbo for raspbian

* onnxruntime not on arm

* remove tflite install resource (taken care of in setup.py), and add link to trash classifier example

…ement (lobe#18)

Fixes the issues mentioned in lobe#17

Second fix for lobe#17

* add rpi install script
* update readme

Use setuptools extras_require for install options -- [all], [tf], [tflite], and [onnx]

* customize what backends you want installed

* Split out backend package requirements into setup extras for modular installation

* Get the right tflite_runtime wheel url based on the machine, update readme to reflect modular install instructions.

* Update error message for installing modular backends.

* Add release notes for 0.5

…ntime require different versions. If both are installed, use the tflite interpreter built into tensorflow instead of the tflite_runtime one.

…(so that we can publish on PyPi). Add instructions for installing tflite_runtime, and fix the Raspberry Pi install script to reflect those new instructions.

Needed to chmod 755 for good measure.

…ckend base classes (lobe#28)

* Add Grad-CAM++ visualization for exported tensorflow image classification models.

* Clean comments

* Batch-ify the ImageModel.visualize() function so you can pass a list of images.

* Bump Pillow version to latest

The `super()` calls are using the wrong class name. By doing this,
they're skipping the `__init__()` for the actual super class.

This breaks calling `predict()` because `self.lock` has not been
assigned.

The change for the `tflite` backend has been tested. The change for
`onnx` has not been tested, but seems correct.

… install, add tags to setup.py (lobe#33)