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 and tf.concat allow a single tensor.
• Core: Shader programs now work in GLSL 4.1.
• Core: loadWeights now supports RequestInit.
• Layers: Support dilation in Conv1D and Conv2D 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.