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)

v0.7.2

Metalhead v0.7.2

Diff since v0.7.1

Closed issues:

• Addition of the UNet model architecture (#126)
• repeat is not efficient on GPU (#157)
• Code style (#160)

Merged pull requests:

• Try to get gradtests to pass (#152) (@theabhirath)
• Fix docstring for DenseNet (#156) (@theabhirath)
• Repeat fix 2 (#159) (@theabhirath)
• Fix mutation error (#162) (@theabhirath)
• Use MLUtils.ones_like for ClassTokens (#166) (@theabhirath)
• Switch to SciML style, take 3 (#168) (@theabhirath)
• Hotfix for ViT on GPU (#169) (@theabhirath)

v0.7.1

Metalhead v0.7.1

Diff since v0.7.0

Closed issues:

• GSoC 2022 (#148)

Merged pull requests:

• Improved time to first gradient (#151) (@theabhirath)
• Expose nclasses option for DenseNet (#154) (@theabhirath)
• Revert returning Chain for conv_bn (#155) (@theabhirath)

v0.7.0

Metalhead v0.7.0

Diff since v0.6.1

Closed issues:

• Add new Models and provide user with an option to fetch untrained models (#16)
• Feature request: ESRGAN, upconv7, upcunet (#81)
• ResNet50 not defined (#96)
• trainimgs not defined (#98)
• Add EfficientNet with pretrained weights (#106)
• ResNet50: difference between PyTorch and default Metalhead implementation (#110)
• Addition of classic LetNet-5 model (#121)
• cat_channels should support > 2 args (#127)
• Confused about dimensions in computation of multihead attention (#139)

Merged pull requests:

• Simplify construction for some models (#99) (@DhairyaLGandhi)
• Implementation of ResNeXt (#100) (@theabhirath)
• Implementation of MobileNetv2 and MobileNetv3 (#101) (@theabhirath)
• Add back code coverage (#102) (@ToucheSir)
• Implementation of MLPMixer (#103) (@theabhirath)
• Fix docs CI and update docs (#104) (@darsnack)
• Implementation of base ViT model (#105) (@theabhirath)
• CompatHelper: bump compat for NNlib to 0.8, (keep existing compat) (#107) (@github-actions[bot])
• Minor doc fixes for MLPMixer (#108) (@theabhirath)
• Update first conv layer and default connection for ResNet (#111) (@darsnack)
• ResNet: apply stride 2 to conv_bn((3,3)) as in PyTorch (#114) (@Alexander-Barth)
• VGG: use stride=2 for MaxPool like PyTorch (#115) (@Alexander-Barth)
• Update residual connection for ResNet18 and ResNet34 (#116) (@Alexander-Barth)
• Implementation for ConvNeXt (#119) (@theabhirath)
• Refactor SE layer, inverted bottleneck, and MobileNet utilities (#120) (@darsnack)
• Minor fixes for ConNeXt (#122) (@theabhirath)
• CompatHelper: bump compat for MLUtils to 0.2, (keep existing compat) (#124) (@github-actions[bot])
• Implementation of ResMLP and gMLP (with improvements to MLPMixer and PatchEmbedding) (#125) (@theabhirath)
• Let cat_channels accept more args (#128) (@mcabbott)
• Implementation of ConvMixer (#130) (@theabhirath)
• Release 0.7 (#132) (@darsnack)
• Updated README.md (#133) (@ceferisbarov)
• Refactor of ViT models (#135) (@theabhirath)
• Update README.md (#138) (@theabhirath)
• Add MobileNet v1 (#140) (@darsnack)
• Release 0.7 (round 2) (#142) (@darsnack)
• Don't run gradient code in quickstart (#146) (@darsnack)
• Minor tweaks (#147) (@theabhirath)

v0.6.1

Metalhead v0.6.1

Diff since v0.6.0

Closed issues:

• Does Metalhead.classify support pretrained models other than VGG19 ? (#30)
• VGG19 model fails to load with EOFError (#58)

Merged pull requests:

• Add references/links to models in README.md (#92) (@trahflow)
• resnet: fix typo (#97) (@iblis17)

v0.6.0

Metalhead v0.6.0

Diff since v0.5.3

Closed issues:

• Metalhead APIs used in README and datasets/README seem to no longer exist (#8)
• Add a MaxPool object (#13)
• predict not working (#15)
• Float32/Float64 Dilemma (#31)
• load custom dataset (#38)
• Loader For Pascal VOC dataset (#39)
• I met this error when run "feature extraction with MetalHead". (#41)
• VGG example fails on GPU (#42)
• ResNet on MNIST (#47)
• Add support for miniImagenet (#55)
• Resnet50 doesn't seem to work (#61)
• ResNet does not work properly (#62)
• ImageNet dataset not available (#64)
• Metalhead fails on M1 Mac with Julia 1.7 Beta 3 (#85)
• ResNet skip connection definition seems wrong (#89)
• flatten layers appear as closures (#94)

Merged pull requests:

• Refactor repo with more complete models and documentation (#70) (@darsnack)
• add buildkite pipeline (#88) (@DhairyaLGandhi)
• CompatHelper: add new compat entry for Functors at version 0.2, (keep existing compat) (#90) (@github-actions[bot])
• Make ResNet more flexible/generic + minor bug fix (#91) (@darsnack)
• Release fixes (#93) (@darsnack)

v0.5.3

Metalhead v0.5.3

Diff since v0.5.1

Closed issues:

• Incompatible with Flux 0.11 (#65)
• Install (#74)
• [GSoC 2021 Proposal] Segmentation models for Metalhead (#77)

Merged pull requests:

• Install TagBot as a GitHub Action (#57) (@JuliaTagBot)
• Update README.md (#73) (@cyborg1995)
• Bump compat of Colors (#75) (@VarLad)
• CompatHelper: bump compat for "BSON" to "0.3" (#76) (@github-actions[bot])
• CompatHelper: bump compat for "Flux" to "0.12" (#78) (@github-actions[bot])
• Add GitHub Actions CI (#82) (@DhairyaLGandhi)

v0.5.1

bump version

v0.5.0

v0.5.0 (2019-12-21)

Diff since v0.4.0

Merged pull requests:

• CompatHelper: bump compat for "ColorTypes" to "0.9" (#52) (github-actions[bot])
• update Requires compat and add CompatHelper (#51) (ChrisRackauckas)
• Add support for Flux v0.10 and Zygote (#50) (dhairyagandhi96)

v0.4.0

v0.4.0 (2019-11-19)

Diff since v0.3.0

Closed issues:

• while training, get InvalidIRError (#45)
• MethodError: no method matching maxpool(::Array{Float32,4}, ::Tuple{Int64,Int64}) (#44)
• MethodError: no method matching conv(::Array{Float32,4}, ::Array{Float64,4}; ...) (#28)
• Enable code coverage (#26)
• Download CIFAR-10 (#21)

Merged pull requests:

• Add compat entries (#49) (dhairyagandhi96)
• use Pkg3 (#48) (dhairyagandhi96)
• Replace maxpool with MaxPool (#43) (yuehhua)
• Update README.md (#37) (EMCP)
• Update utils.jl (#36) (shashikdm)
• Small readme edits (#35) (sklan)
• allow preprocessing grayscale images (#34) (KristofferC)
• Remove trailing bracket from README.md (#33) (dhairyagandhi96)
• [WIP] Refactor tests a bit (#27) (staticfloat)
• Fix deprecated functions (method_exists, Flux.argmax) (#25) (wookay)
• Update README image (#24) (Keno)
• Fix display of prediction frames (#23) (Keno)
• Automatically download CIFAR10 (#22) (avik-pal)