Re:Show HN: “Is this loss?”: A TensorFlow Lite app to detect...
by eigenloss on Monday April 23, 2018 @06:55PM
(#16908226)
Training was a bit of a process. The architecture I use now is an 8-bit quantized mobilenet_v1_050_224 retrained on about 1000 versions of Loss.jpg (some quite similar, but not exact duplicates) and about 10,000 other images. All the images are 240x240 or thereabouts.
I'll be publishing something soon on what it took to build that 10K image dataset - it was more challenging than I initially anticipated.
I'll be publishing something soon on what it took to build that 10K image dataset - it was more challenging than I initially anticipated.