v0.9.3

Metalhead v0.9.3

Diff since v0.9.2

Merged pull requests:

• Fix VGG dropout probability (#272) (@adrhill)

Closed issues:

• Default VGG dropout probability changed to zero (#271)

v0.9.2

Metalhead v0.9.2

Diff since v0.9.1

Merged pull requests:

• Deprecate older VGG API (#270) (@theabhirath)

Closed issues:

• VGG pretrained (#268)
• VGG has two "higher-level" functions (#269)

v0.9.1

Metalhead v0.9.1

Diff since v0.9.0

Merged pull requests:

• CompatHelper: add new compat entry for Statistics at version 1, (keep existing compat) (#255) (@github-actions[bot])
• CompatHelper: add new compat entry for LazyArtifacts at version 1, (keep existing compat) (#256) (@github-actions[bot])
• CompatHelper: add new compat entry for Artifacts at version 1, (keep existing compat) (#257) (@github-actions[bot])
• use julia-actions/cache in CI (#260) (@IanButterworth)
• Switch to (Re)TestItems (#262) (@ToucheSir)
• Fix docstring typos (#263) (@christiangnrd)
• Fix bug where MobileNetv2/3 return a MobileNetv1 (#265) (@darsnack)
• Move to Documenter 1.0 (and some miscellaneous fixes) (#266) (@theabhirath)

Closed issues:

• Cannot run EfficientNetv2 on gpu (#261)

v0.9.0

Metalhead v0.9.0

Diff since v0.8.3

Merged pull requests:

• Make CUDA an optional dependency (#253) (@mohamed82008)
• Update use of Flux._isactive in DropBlock to prevent deprecation warning (#254) (@theabhirath)

Closed issues:

• Make CUDA an optional dependency in Julia 1.9 (#250)
• Saving large models with JLD2 fails (#252)

v0.8.3

Metalhead v0.8.3

Diff since v0.8.2

Merged pull requests:

• Adding a more detailed introduction to the Layers module (#249) (@theabhirath)
• CompatHelper: bump compat for CUDA to 5, (keep existing compat) (#251) (@github-actions[bot])

v0.8.2

Metalhead v0.8.2

Diff since v0.8.1

Merged pull requests:

• Remove NNlibCUDA dependency to support Flux 0.14 (#248) (@christiangnrd)

v0.8.1

Metalhead v0.8.1

Diff since v0.8.0

Closed issues:

• No valid algorithm found for EfficientNet and EfficientNetv2 (#245)

Merged pull requests:

• CompatHelper: bump compat for NNlib to 0.9, (keep existing compat) (#244) (@github-actions[bot])
• CompatHelper: bump compat for Flux to 0.14, (keep existing compat) (#246) (@github-actions[bot])

v0.8.0

Metalhead v0.8.0

See this blog post on fluxml.ai about what's new in 0.8.

Diff since v0.7.4

Closed issues:

• Move MHAttention layer to Flux (#141)
• Issues with ViT on the GPU (#165)
• Cannot load pretrained weights for ResNet on master (#206)
• Support for convolution on N-dimensions (#213)
• Fixing deprecated models (#218)
• all links in the Function column in the readme are broken (#220)
• mismatch with torchvision resnets (#228)
• cannot match attention layer output to pytorch's one (#231)
• change ViT constructor's defaults (#232)
• change loadpretrain! to load the model_state (#234)
• pin the artifacts url to specific commit (#236)
• got unsupported keyword argument "inchannels" (#237)
• Release v0.8 (#240)

Merged pull requests:

• CompatHelper: bump compat for CUDA to 4, (keep existing compat) (#215) (@github-actions[bot])
• Natema/issue218 (#219) (@natema)
• Docs restructuring, part 1 (reland) (#221) (@theabhirath)
• Update buildkite (#222) (@CarloLucibello)
• remove some duplicate include (#223) (@CarloLucibello)
• add pytorch/torchvision porting and comparison script (#225) (@CarloLucibello)
• fix the torchvision porting script (#229) (@CarloLucibello)
• fix ViT model output + rewrite attention layer + adapt torchvision script (#230) (@CarloLucibello)
• automatize huggingface upload + change ViT defaults (#233) (@CarloLucibello)
• hook new weights (#235) (@CarloLucibello)
• set specific commit in Artifacts.toml (#239) (@CarloLucibello)
• Docs rehaul for v0.8 (#242) (@theabhirath)

v0.7.4

Metalhead v0.7.4

Diff since v0.7.3

Closed issues:

• MobileNetv1 doesn't match the implementation in the paper (#186)
• Tweak GoogLeNet and Inception family to match the torchvision implementations (#196)
• Unable to load pretrained ResNeXt models (#212)

Merged pull requests:

• Overhaul of ResNet API (#174) (@theabhirath)
• Initial switch to ReTest.jl (#182) (@darsnack)
• Removed extra mobilenetv1 fc layer, fcsize param (#189) (@cluffa)
• Expose a uniform API at the highest level for models (#190) (@theabhirath)
• Res2Net and Res2NeXt, again (#195) (@theabhirath)
• Add pretrained weights on ImageNet for some models (#197) (@theabhirath)
• Implementation of EfficientNetv2 and MNASNet (#198) (@theabhirath)
• Migrate docs to Documenter.jl (#199) (@Saransh-cpp)
• Unify lower level API for EfficientNet and MobileNet model families (#200) (@theabhirath)
• Do not catch orthogonal errors when loading (#203) (@darsnack)
• CompatHelper: bump compat for MLUtils to 0.3, (keep existing compat) (#204) (@github-actions[bot])
• Tweak GoogLeNet to match the torchvision implementations (#205) (@pri1311)
• CompatHelper: bump compat for Functors to 0.4, (keep existing compat) (#207) (@github-actions[bot])
• Readme Links broken (#208) (@ashwani-rathee)
• Adding UNet implementation (#210) (@shivance)
• CompatHelper: bump compat for MLUtils to 0.4, (keep existing compat) (#211) (@github-actions[bot])
• Mark resnet acctests broken (#214) (@ToucheSir)
• update: actions node 12 => node 16 (#217) (@skyleaworlder)

v0.7.3

Metalhead v0.7.3

Diff since v0.7.2

Closed issues:

• Different Models have different preprocessing needs (#32)
• ResNet producing non deterministic and wrong predictions (#72)
• Add steps to create new pre-trained models (#86)
• Add config parameters to network constructors: n_channel_input and output_softmax? (#95)
• Pre-trained models and classify disappeared? (#143)
• Code evaluation errors in the docs (#175)
• MobileNetv2/3 need padding in first layer (#178)

Merged pull requests:

• Add pre-trained models from HuggingFace and add accuracy testing (#164) (@darsnack)
• Implementation of Inceptionv4, InceptionResNetv2 and Xception (#170) (@theabhirath)
• Implement EfficientNet (#171) (@darsnack)
• Fix contributor docs link (#173) (@darsnack)
• Fix quickstart docs and some docstrings (#177) (@darsnack)
• Miscellaneous fixes for MobileNet (#179) (@theabhirath)