Update 2023-03-21: This Release is Deprecated

This release contains a regression that causes some graph models to dispose their tensors too early. See #7504 for details.

Core (4.2.0 ==> 4.3.0)

Features

• Update flags.ts (#7367). Thanks, @Linchenn.
• [op] added support for TensorScatterUpdate op (#7189).

Bug fixes

• Revert "Add a case for atanh when input is 1 " (#7325) (#6964). Thanks, @Linchenn.
• Fix eluGrad's error (#7251). Thanks, @xhcao.

Performance

• Improve assertAndGetBroadcastShape (#7413). Thanks, @Linchenn.

Documentation

• Update tensor.ts (#7366). Thanks, @Linchenn.

Misc

• Update monorepo to 4.3.0. (#7488).
• [wasm] Add Multinomial kernel (#7468). Thanks, @chunnienc.
• Fix gather signature in chained ops (#7465).
• disable offscreen canvas for safari, since gl fencing is not working yet for its technology preview (#7436).
• Propagate errors from indexedDB database.put (#7322).
• Add StaticRegexReplace Op (#7379).
• SharedArrayBuffer is not GPU data (#7344). Thanks, @axinging.
• Support MatrixBandPart and allow numLower to be a tensor (#7351). Thanks, @Linchenn.
• Add a case for atanh when input is 1 (#6964). Thanks, @xhcao.
• [tfjs-node] support einsum in tfjs node (#7321).
• fix comment in tile.ts (#7298). Thanks, @sonmh79.

Data (4.2.0 ==> 4.3.0)

Misc

• Update monorepo to 4.3.0. (#7488).
• build(deps): bump vega in /tfjs-data/demo/boston-housing (#7444). Thanks, @dependabot[bot].
• build(deps): bump vega-functions in /tfjs-data/demo/boston-housing (#7442). Thanks, @dependabot[bot].
• build(deps): bump json5 in /tfjs-data/demo/boston-housing (#7219). Thanks, @dependabot[bot].
• build(deps): bump fast-json-patch in /tfjs-data/demo/boston-housing (#7211). Thanks, @dependabot[bot].

Layers (4.2.0 ==> 4.3.0)

Misc

• Update monorepo to 4.3.0. (#7488).
• Typo fix (#7393). Thanks, @dmotz.
• #7376 Allow seeding the Orthogonal Initializer and use QR instead of gramSchmidt (#7377). Thanks, @cosminc98.

Converter (4.2.0 ==> 4.3.0)

Features

• [converter] add TensorScatterUpdate op to converter and executor (#7246).

Bug fixes

• allow weight only keras H5 file that lacks keras version info to be parsed without error (#7359).

Misc

• Update tfjs_4.3.0 lockfiles (#7492).
• Update monorepo to 4.3.0. (#7488).
• Never pass a NoOp value to an op (#7460).
• Fix topological sort for sync nodes execution order (#7431). Thanks, @chunnienc.
• Remove array.at (#7409). Thanks, @chunnienc.
• Add new tensor diposal algorithm for sync graph execution (#7392). Thanks, @chunnienc.
• Add cache for parseNodeName (#7391). Thanks, @chunnienc.
• Update rules_python to 0.18.1 (#7383).
• Add StaticRegexReplace Op (#7379).
• added converter support for inf and infinite ops (#7365).
• Support MatrixBandPart and allow numLower to be a tensor (#7351). Thanks, @Linchenn.
• Support tfjs converter options in JAX conversion (#7331).
• Update TypeScript to 4.9.4 (#7300).
• Debug converter python code with debugpy (#7283).
• Refactor python converter functions to not call each other (#7281).
• Rename bazel BUILD files (#7253). Thanks, @chunnienc.
• Make converter hashtable test not overly specific (#7249).
• build(deps): bump json5 from 2.1.0 to 2.2.3 in /tfjs-converter/demo (#7240). Thanks, @dependabot[bot].

Node (4.2.0 ==> 4.3.0)

Features

• [op] added support for TensorScatterUpdate op (#7189).

Misc

• Update tfjs_4.3.0 lockfiles (#7492).
• Update monorepo to 4.3.0. (#7488).
• [wasm] Add Multinomial kernel (#7468). Thanks, @chunnienc.
• Add StaticRegexReplace Op (#7379).
• [tfjs-node] support einsum in tfjs node (#7321).
• Update TypeScript to 4.9.4 (#7300).
• Improve npm publishing infra (#7287).
• build(deps): bump json5 from 2.2.0 to 2.2.3 in /tfjs-node-gpu (#7243). Thanks, @dependabot[bot].
• build(deps): bump json5 from 2.2.0 to 2.2.3 in /tfjs-node (#7242). Thanks, @dependabot[bot].
• Do not run docker interactively in tfjs-node publish (#7227).

Wasm (4.2.0 ==> 4.3.0)

Features

• [wasm] Add atan kernel (#7252). Thanks, @chunnienc.
• [op] added support for TensorScatterUpdate op (#7189).

Security

• build(deps): bump json5 in /tfjs-backend-wasm/starter/webpack (#7218). Thanks, @dependabot[bot].

Misc

• Update tfjs_4.3.0 lockfiles (#7492).
• Update monorepo to 4.3.0. (#7488).
• [wasm] Add ArgMin (#7306). Thanks, @chunnienc.
• [wasm] Add Conv3D kernels (#7466). Thanks, @chunnienc.
• Add unique (#7469). Thanks, @chunnienc.
• [wasm] Add Multinomial kernel (#7468). Thanks, @chunnienc.
• Add expm1 (#7467). Thanks, @chunnienc.
• [wasm] Add Bincount kernel (#7412). Thanks, @chunnienc.
• Add LinSpace (#7403). Thanks, @chunnienc.
• Fix license headers added in 2023 (#7353). Thanks, @chunnienc.
• [wasm] Add assorted unary and binary kernels (#7350). Thanks, @chunnienc.
• [wasm] Add ResizeBilinearGrad and ResizeNearestNeighborGrad kernels (#7346). Thanks, @chunnienc.
• Add Log1p (#7328). Thanks, @chunnienc.
• [wasm] Add LRN kernels (#7335). Thanks, @chunnienc.
• [wasm] Add Dilation2D backprop kernels (#7334). Thanks, @chunnienc.
• [wasm] Add Dilation2D (#7327). Thanks, @chunnienc.
• [wasm] Add sparseToDense kernel (#6528). Thanks, @kon72.
• [wasm] Add EluGrad (#7307). Thanks, @chunnienc.
• [wasm] Enable lowerBound and upperBound op tests (#7301). Thanks, @chunnienc.
• [wasm] Add AvgPool3DGrad and MaxPool3DGrad kernels (#7295). Thanks, @chunnienc.
• Add SearchSorted (#7299). Thanks, @chunnienc.
• [wasm] Add AvgPool3D and MaxPool3D kernels (#7294). Thanks, @chunnienc.
• [wasm] Add BroadcastArgs kernel (#7290). Thanks, @chunnienc.
• Upgrade global build param from C++11 to C++17 (#7268). Thanks, @chunnienc.
• [wasm] Add DenseBincount kernel (#7267). Thanks, @chunnienc.
• Add Diag kernel (#7277). Thanks, @chunnienc.
• [wasm] Add asin and asinh kernels (#7262). Thanks, @chunnienc.
• [wasm] Add acos and acosh kernels (#7256). Thanks, @chunnienc.
• Rename bazel BUILD files (#7253). Thanks, @chunnienc.
• Fix test segfaults (#7258). Thanks, @chunnienc.
• build(deps): bump json5 in /tfjs-backend-wasm/starter/parcel (#7222). Thanks, @dependabot[bot].

Cpu (4.2.0 ==> 4.3.0)

Features

• [op] added support for TensorScatterUpdate op (#7189).

Bug fixes

• Fix eluGrad's error (#7251). Thanks, @xhcao.

Misc

• Update monorepo to 4.3.0. (#7488).
• Add unique (#7469). Thanks, @chunnienc.
• Add StaticRegexReplace Op (#7379).
• [WebGPU] Support forwarding FloorDiv to CPU (#7336).

Webgl (4.2.0 ==> 4.3.0)

Features

• [WebGL] Implement packed ScatterND (#7292). Thanks, @Linchenn.
• [op] added support for TensorScatterUpdate op (#7189).

Bug fixes

• Fix eluGrad's error (#7251). Thanks, @xhcao.

Performance

• Remove unnecessary slice (#7414). Thanks, @Linchenn.
• Fine-tune packed BackpropInput (#7386). Thanks, @Linchenn.
• [WebGL] Support packed Conv2DBackpropInput (#7339). Thanks, @Linchenn.

Misc

• Update tfjs_4.3.0 lockfiles (#7492).
• Update monorepo to 4.3.0. (#7488).
• React native update tfjs to 4.2.0 (#7451).
• disable offscreen canvas for safari, since gl fencing is not working yet for its technology preview (#7436).
• [webgl] Reduce JS map lookups in webgl backend (#7387). Thanks, @chunnienc.
• webgl: Add packed op for sin/cos (#7382). Thanks, @qjia7.
• Add StaticRegexReplace Op (#7379).
• Fix license headers added in 2023 (#7353). Thanks, @chunnienc.

Webgpu (0.0.1-alpha.17 ==> 0.0.1-alpha.18)

Misc

• Update monorepo to 4.3.0. (#7488).
• webgpu: Remove the unnecessary conditions (#7470). Thanks, @qjia7.
• webgpu: Optimize broadcast mul/div (#7437). Thanks, @qjia7.
• webgpu: Switch back to old karma reporter (#7449). Thanks, @gyagp.
• [WebGPU] support AvgPool3DGrad kernel (#7440). Thanks, @xhcao.
• [WebGPU] support conv3d backprop kernels (#7434). Thanks, @xhcao.
• [webgpu] Add karma-jasmine-html-reporter to show test status in the b… (#6636). Thanks, @haoyunfeix.
• webgpu: Support any component buffer (#7426). Thanks, @qjia7.
• webgpu: Fix conv2d gradient cases failures (#7433). Thanks, @qjia7.
• [WebGPU] support pool3d kernels (#7405). Thanks, @xhcao.
• webgpu: Optimize Softmax (#7411). Thanks, @qjia7.
• [WebGPU] Support ResizeBilinearGrad kernel (#7385). Thanks, @xhcao.
• webgpu: Skip staticRegexReplace test cases (#7390). Thanks, @gyagp.
• webgpu: Support 2x4 output for Conv2DBackpropInput (#7371). Thanks, @qjia7.
• [WebGPU] support ResizeNearestNeighborGrad kernel (#7354). Thanks, @xhcao.
• webgpu: NaN refactor (#7352). Thanks, @qjia7.
• [WebGPU] Support forwarding FloorDiv to CPU (#7336).
• build(deps): bump ua-parser-js in /tfjs-backend-webgpu (#7312). Thanks, @dependabot[bot].
• webgpu: unify the uniform variables' names of conv (#7279). Thanks, @xhcao.
• webgpu: enable more test cases (#7278). Thanks, @xhcao.
• webgpu: support LRNGrad kernel (#7196). Thanks, @xhcao.
• webgpu: support BroadcastArgs kernel (#7247). Thanks, @xhcao.
• webgpu: support MaxPoolGrad kernel (#7261). Thanks, @xhcao.
• webgpu: add a naive implementation of conv3d (#7016). Thanks, @xhcao.
• build(deps): bump json5 from 2.2.0 to 2.2.3 in /tfjs-backend-webgpu (#7244). Thanks, @dependabot[bot].
• webgpu: support maxPoolWithArgmax kernel (#7176). Thanks, @xhcao.
• webgpu: support EluGrad kernel (#7197). Thanks, @xhcao.
• [webgpu] Skip tensorScatterUpdate unit test (#7255). Thanks, @hujiajie.
• [webgpu] Let vectorized binary op return an overridable value on NaN (#7198). Thanks, @hujiajie.
• webgpu: support DepthwiseConv2dNativeBackprop kernels (#7232). Thanks, @xhcao.
• webgpu: Remove isVectorA check (#7233). Thanks, @qjia7.
• webgpu: support multinomial operator (#7154). Thanks, @xhcao.