Hack week planning

[fchollet]

Participants:

• @haifeng-jin
• @nkovela1
• @grasskin
• @hertschuh
• @mattdangerw
• @rchao
• @jbischof
• @divyashreepathihalli
• @ianstenbit
• @qlzh727
• @sampathweb
• @chenmoneygithub

See proposed assignments below.

Some of the assignment buckets may not last the week -- we will reshuffle assignments as we make progress. If we reach a sufficient degree of feature completeness mid-week, we can move on to backwards compatibility testing.

@fchollet

• plot_model / model_to_dot
• explore modernization of plot_model
• stack trace filtering
• add support for legacy h5 weights loading
• JAX rnn function
• FeatureSpace

Layers:

• Wrapper
• Normalization
• Rescaling
• CenterCrop
• Resizing
• RNN / RNNCell
• StackedRNNCells
• SimpleRNN
• LSTM
• GRU
• Lambda
• ConvLSTM1D/2D/3D
• Bidirectional
• TimeDistributed

Wrapped layers:

• Discretization
• TextVectorization
• Hashing
• StringLookup
• IntegerLookup

@chenmoneygithub

• Conv1D/2D/3D
• SeparableConv1D/2D
• DepthwiseConv1D/2D
• Conv1DTranspose/2D/3D

@sampathweb

• Set up code coverage CI test (e.g. codecov)
• Identify missing tests and major TODOs

Callbacks:

• RemoteMonitor
• TerminateOnNaN
• CSVLogger
• ModelCheckpoint
• BackupAndRestore

@qlzh727

• tf.distribute support in TF backend #96
• DTensor support in TF backend keras#18473

@ianstenbit

• Confusion metrics
• Learning rate schedules (keras_core/optimizers/schedules/)

Callbacks:

• LearningRateScheduler
• ReduceLROnPlateau

@divyashreepathihalli

Layers:

• Elu
• ReLU
• LeakyReLU
• PReLU
• Softmax

Wrapped layers:

• RandomCrop
• RandomFlip
• RandomTranslation
• RandomRotation
• RandomZoom

@jbischof

• Implement keras_core/backend/pytorch/numpy.py

@rchao

Layers:

• UpSampling1D/2D/3D

@mattdangerw

Layers:

• Attention
• AdditiveAttention
• MultiHeadAttention

@hertschuh

Layers:

• Flatten
• Reshape
• RepeatVector
• Permute
• Cropping1D/2D/3D
• ZeroPadding1D/2D/3D

@grasskin

Metrics

• R2Score
• F1Score
• FBetaScore
• IoU metrics

Callbacks:

• TensorBoard

@nkovela1

Layers:

• UnitNormalization
• LayerNormalization
• GroupNormalization
• SpectralNormalization
• CategoryEncoding
• RandomContrast
• RandomBrightness

@haifeng-jin

• Implement step fusing in TF Trainer and JAX Trainer

[AakashKumarNain]

Not sure about the best option for explore modernization of plot_model but we can use Rich for modernization of progress bars and all. It's also used by pip now

[jbischof]

@fchollet I might need more details for the pytorch backend. Should I just try to replicate backend/tensorflow/numpy.py line for line? The entire backend/tensorflow folder?

I also don't see any unit tests (at least in the backend/ folder. How do we verify the deliverables?

[fchollet]

I might need more details for the pytorch backend. Should I just try to replicate backend/tensorflow/numpy.py line for line?

Just pytorch/numpy.py for now. Yes, all the same functions as in tensorflow/numpy.py need to be implemented. Ramesh is also taking a look at pytorch/nn.py.

I also don't see any unit tests (at least in the backend/ folder. How do we verify the deliverables?

This is indirectly tested by unit tests in keras_core/operations/. I think you should be able to set the backend as "pytorch" (we'll need to edit backend/__init__.py to make it work) and run those tests.

We could also add a keras_core/backend/tests/ directory to cover cross-backend function tests without relying on any other Keras functionality. It would probably be a good move since it makes it easier for folks to add new backends.

[ianstenbit]

Jonathan, I think I may have not enough on my list for the hack week so I can help out with the PyTorch if you’d like (sounds like fun to me) I’ve added a few things to backend/ and ops/ so I think I have a good sense of what needs to happen.

…

On Sun, May 7, 2023 at 10:46 AM François Chollet ***@***.***> wrote: I might need more details for the pytorch backend. Should I just try to replicate backend/tensorflow/numpy.py <https://github.com/keras-team/keras-core/blob/main/keras_core/backend/tensorflow/numpy.py> line for line? Just pytorch/numpy.py for now. Yes, all the same functions as in tensorflow/numpy.py need to be implemented. Ramesh is also taking a look at pytorch/nn.py. I also don't see any unit tests (at least in the backend/ folder. How do we verify the deliverables? This is indirectly tested by unit tests in keras_core/operations/. I think you should be able to set the backend as "pytorch" (we'll need to edit backend/__init__.py to make it work) and run those tests. We could also add a keras_core/backend/tests/ directory to cover cross-backend function tests without relying on any other Keras functionality. It would probably be a good move since it makes it easier for folks to add new backends. — Reply to this email directly, view it on GitHub <#83 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAXOHB3ZXB7DRS4WFWE2UR3XE7GXLANCNFSM6AAAAAAXVCRMQE> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[fchollet]

@AakashKumarNain are you still working on the PyTorch dataloader?

[AakashKumarNain]

@fchollet didn't get much time to work during weekdays. Was planning to work on it over the weekend but if someone else wants to pick it up right aways, I am okay with it

[fchollet]

It's still yours if you want to do it!

[AakashKumarNain]

I will work on it and will update you on the status on Monday

[AakashKumarNain]

@fchollet now that torch dataloader adapter has been merged, is there any unassigned task that I can work on in the coming weeks?

[fchollet]

@fchollet now that torch dataloader adapter has been merged, is there any unassigned task that I can work on in the coming weeks?

Yes, thanks for offering! The next big sprint is going to be porting Keras Applications. We have 12 separate architectures to port. You could be working on e.g. ResNet, ResNetV2, MobileNet, etc. I will port the first application (likely VGG) within a few days.

[AakashKumarNain]

Thanks. I will start with ResNet

[fchollet]

Thanks a lot for the help! 👍

[AakashKumarNain]

No problem. This project is way too close to my heart 🥂