v0.2.0

What's Changed

• Let a transformer block initialize its residual projections by @jessegrabowski in #83
• Draw every parameter at construction, and refuse a constant scheme by @jessegrabowski in #84
• Replace the network-wide init scheme with per-parameter initializers by @jessegrabowski in #85
• Share one implementation between adam and adamw by @jessegrabowski in #86
• Stop the frozen-generator check from rejecting inner-graph inputs by @jessegrabowski in #87
• Skip parameters the loss cannot reach by @jessegrabowski in #88
• Serialize initializers, replacing CustomInitializer with a decorator by @jessegrabowski in #89
• Allow a stored learning rate, with one writer per variable by @jessegrabowski in #90
• Add a reduce-on-plateau learning rate policy by @jessegrabowski in #91
• Add a recurrent layer, split into a cell and a scan wrapper by @jessegrabowski in #92
• Hoist a draw out of a scan, so dropout works in a recurrent cell by @jessegrabowski in #93
• Add a GRU cell and layer by @jessegrabowski in #94
• Add an LSTM, and a direction argument with a Bidirectional wrapper by @jessegrabowski in #95
• Add a mask argument, so a padded batch does not disturb the recurrence by @jessegrabowski in #96
• Add a Conv1D layer, lowered to a patch gather and one matmul by @jessegrabowski in #97
• Dispatch convolution to per-backend kernels, and give the core CI no backend by @jessegrabowski in #98
• Add a Conv2D layer, and stop its gradient going through a batched matmul by @jessegrabowski in #99
• Add max and average pooling, dispatched on every backend by @jessegrabowski in #100
• Add Flatten, and take batch norm's statistics per channel by @jessegrabowski in #101
• Make ConvLayerGrad usable on its own as a transposed convolution by @jessegrabowski in #102
• Add zero, constant, reflection and replication padding layers by @jessegrabowski in #103
• Advance every training clock the compiled function reads by @jessegrabowski in #105
• Add skip_if, a step guard with non-finite and large-step conditions by @jessegrabowski in #107
• Compose a rule with trailing transforms, and scale by a schedule by @jessegrabowski in #108
• Unbreak the release build's metadata check by @jessegrabowski in #110
• Bump the publish action past the metadata 2.5 rejection by @jessegrabowski in #111

Full Changelog: v0.1.0...v0.2.0