STT-tensorflow

History

Frederic Rechtenstein 9073aa931f Int16x8 Logistic op : ensure that input_left_shift >= 0

PiperOrigin-RevId: 355075093
Change-Id: I7cd2d71181e3404c8c34209473ee62c4ed16692e

2021-02-01 19:14:20 -08:00

add.h

Add quantized int8 evaluation for squared_difference and rsqrt

2020-11-25 20:09:43 -08:00

conv.h

Add bug for TODO in conv and depthwise conv range check.

2020-11-26 00:36:15 -08:00

depthwise_conv.h

Add bug for TODO in conv and depthwise conv range check.

2020-11-26 00:36:15 -08:00

dequantize.h

Add int16 support to Dequant.

2019-07-19 00:26:49 -07:00

fully_connected.h

Prefer the standard integral types over custom type-aliases.

2020-07-24 10:34:03 -07:00

l2normalization.h

Prefer the standard integral types over custom type-aliases.

2020-07-24 10:34:03 -07:00

log_softmax.h

Create int8 log softmax.

2019-02-19 13:20:27 -08:00

logistic.h

Int16x8 Logistic op : ensure that input_left_shift >= 0

2021-02-01 19:14:20 -08:00

mean.h

Enable tests to pass when running through make

2020-08-19 12:14:42 +02:00

mul.h

Prefer the standard integral types over custom type-aliases.

2020-07-24 10:34:03 -07:00

pooling.h

Prefer the standard integral types over custom type-aliases.

2020-07-24 10:34:03 -07:00

README.md

…

tanh.h

Merge pull request #42671 from wwwind:tanh_logistic_fixes

2021-01-27 00:43:04 -08:00

transpose_conv.h

Port the transpose_conv operator along with with relevant tests.

2021-01-05 16:02:30 -08:00

README.md

This directory contains reference implementations for int8 fully integer kernels.

Weight filters of convs are expected to be symmetric per-channel quantized in the range [-127, 127]. Inputs/activations are expected to be asymmetric per-layer quantized in the range [-128, 127].

THESE ARE EXPERIMENTAL AND PRONE TO CHANGE.