sophia0509 / coral-basecaller

ONT basecaller running on Coral edge TPU

Setup and installation

First you need Coral TPU accelerator plugged into your USB port.

Before proceeding further, ensure that you have python3 (~3.7) installed and that /bin/env python refers to it. If you are using conda, you can just create environment as conda create --name myenv python=3.7 and then activate myenv. Alternatively, you can try adjusting python to python3 in basecall.py and use pip3 instead of pip.

Then install edge TPU runtime. We strongly recommend using maximum frequency version and plugging the device into USB3 port.

Install requirements via pip install -r requirements.txt. Note that we have fixed version for Linux/Ubuntu so if you are on Mac you should edit requirements to correct platform.

Install tflite runtime.

Running

./basecall.py --model networks/paper_both_init3_f128_k21_r5_edgetpu.tflite --directory your_directory/ --output output.fasta

Use --watch is you wish to run live basecalling while sequencing. You may need to adjust --wait-since-last-write if you suspect that we try to basecall file that is still being written byt MinKNOW. Also, be sure to adjust number of reads being written to the same file when starting the run -- we can't really do live basecalling if we need to wait for 4000 reads to be saved into a single file. A value of 100 should be a good tradeoff between the number of produced fast5 files and processing latency.

On OSX you might need to add --mp-spawn argument to fix some multiprocessing issues.

You can use different tflite files you get different speed/accuracy tradeoff (lower depthwise kernel k means higher speed and smaller accuracy). Use *_edgetpu.tflite files for accelerated Coral inference.

In case you don't have Coral device and want to test our accuracy, you can use original files. Note, however, that Tensorflow Lite is not optimized for x86_64 inference and it will be terribly slow (Google has it on a roadmap https://www.tensorflow.org/lite/guide/roadmap). If you want to go this road, remove experimental_delegates=[tflite.load_delegate(EDGETPU_SHARED_LIB, {})], line in backend.py.

Finally, you can also use non-edgetpu files for visualisation of the architecture (use https://netron.app/ for that).

Reproducing results

If you wish to reproduce results from our paper (https://arxiv.org/abs/2011.04312), head to https://github.com/fmfi-compbio/coral-benchmark and to https://github.com/fmfi-compbio/coral-training/