0.9.0
New features
- Core ops:
tf.rsqrt
,tf.mod
,tf.losses.absoluteDifference
,tf.losses.computeWeightedLoss
,tf.losses.squaredDifference
,tf.sign
,tf.floor
,tf.reciprocal
,tf.round
,tf.asinh
,tf.acosh
,tf.atanh
,tf.atan2
. - Core: Centered RMSProp now supported.
- Core:
logicalNot
,cast
supports chaining. - Core:
tf.stack
andtf.concat
allow a single tensor. - Core: Shader programs now work in GLSL 4.1.
- Core:
loadWeights
now supports RequestInit. - Layers: Support dilation in
Conv1D
andConv2D
layers. - Layers: Add
SeparableConv2D
Layer (tf.layers.separableConv2d
).
Bug fixes
- Layers: Map weights to uniquefied tensor names during model loading, to prevent error on repeated loading of the same model with
tf.loadModel()
in the same browser page load. - Layers: Change class registry to
AveragePooling(1|2)D
and add symbol aliases to exports.
Acknowledgments to external contributors:
@manrajgrover for tf.exmp1
, tf.squaredDifference
, tf.rsqrt
, tf.mod
, absolute difference loss, tf.sign
, fixing gradient for tf.floor
, tf.reciprocal
, logicalNot
chaining support, fixing documentation, adding info to issue template.
@Lewuathe for centered RMSProp, unary ops benchmarks, tf.round
. @jgartman for inverse trig hyperbolic functions and tf.atan2
. @adarob for tf.split
, 1-tensor tf.stack
and tf.concat
. @adarabi for fixing GLSL shader compiler for glsl 4.1.