tpu VS Mask_RCNN

Found relevant code at https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet + all code implementations here

I'm looking for a generative-LM equivalent of an EfficientNet-Lite, for inference on devices with limited to no VRAM. I know about some popular ones like DistilGPT2. But it's been 2 years after its release. Surely, someone improved their size/performance ratio, right... right?

tensorflow/TPU 'apache' license - https://github.com/tensorflow/tpu/commit/6b3236d0271d2f2c3b2dfdc9d233ff00c4ba21cd

According to original repository, the lite variants:

I was quite surprised to see that some architectures, like efficient net (official implementation: https://github.com/tensorflow/tpu/blob/master/models/official/efficientnet/efficientnet_model.py) don't use bias (bias=False). Does anyone know why is that? Apart from the obvious benefit of reducing parameters, doesn't it make the network less capable of learning certain representations?

The most widely used algorithms for classical feature detection today are "whatever opencv implements"

In terms of tech that's advancing at the moment? https://co-tracker.github.io/ if you want to track individual points, https://github.com/matterport/Mask_RCNN and its descendents if you want to detect, say, the cover of a book.

Mask R-CNN

Click to Learn more...

- Mask RCNN https://github.com/matterport/Mask_RCNN (open source, so also not Palantir's)

pip install git+https://github.com/matterport/Mask_RCNN

It is important to realize that to do its masking procedures, Hent-AI uses the Mask RCNN (MRCNN) package from Matterport. The problem with this version of MRCNN is that it is not compatible with Tensorflow 2.X versions, essentially limiting Hent-AI compatibility to strict Tensorflow 1.X versions. Since Tensorflow 1.15 is the last of the Tensorflow 1.X versions and uses CUDA 10.0, which supports a maximum compute capability of 7.5, this means that the last NVIDIA GPU series that is compatible with the original Hent-AI implementation is the RTX 2000 series. This is, of course, not optimal since it means that RTX 3000 series and later GPUs cannot be used despite their significant computing power and high VRAM.

I read that it could bethe problem with scipy version (https://github.com/matterport/Mask_RCNN/issues/2122) so I downgraded it, I also tried to modify shift = np.array([0, 0, 1., 1.]) in utils.py but nothing helped.

I am assuming you did a pip install of this github repository, or did you run pip install mrcnn. The mrcnn package on pypi is just an example package and doesn't have any useful functionality. In addition, where did you get the code from that you are trying to run, from someone else or did you write it yourself? Reason I am asking is because the import error is to be expected since there indeed is no InferenceConfig class defined in mrcnn.visualize.

Hello AI learners, check out my own development of Mask-RCNN supporting Tensorflow2.7.0 and Keras2.8.0. This is an edit of MRCNN which supports Tensoflow1.0, only.