8000 Releases · analokmaus/kuma_utils · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

Releases: analokmaus/kuma_utils

v0.7.0

13 Jan 07:10
@analokmaus analokmaus
10000 v0.7.0
2641645
Compare
Choose a tag to compare
v0.7.0 Latest
Latest

What's new

  • Renewed implementation of TorchLogger and LGBMLogger. The new implementation uses Python's logging module to handle stream output and file writing. The arguments and usage remain the same as before.
Assets 2
Loading

v0.7.0b

13 Jan 05:49
@analokmaus analokmaus
v0.7.0b
6c8d6ca
Compare
Choose a tag to compare
v0.7.0b Pre-release
Pre-release 
update settings related to rye
Loading

v0.6.2

12 Nov 07:35
@analokmaus analokmaus
v0.6.2
1ea041a
Compare
Choose a tag to compare
v0.6.2

What's new

  • Updated pytorch dependencies for security reasons
Loading

v0.6.1

12 Nov 06:44
@analokmaus analokmaus
v0.6.1
d0757e0
Compare
Choose a tag to compare
v0.6.1

What's new

  • Updated dependencies
Loading

v0.6.0

31 Oct 03:57
@analokmaus analokmaus
v0.6.0
a8a602d
Compare
Choose a tag to compare
v0.6.0

What's New

  • TorchTrainer can now accept models in which loss functions are included (e.g., ArcFace, LLMs). You must pass criterion = None to the TorchTrainer.train() and calculate the loss inside your training hook.

Sample code

def forward_train(self, trainer, inputs):
    embeddings, loss = trainer.model.*inputs)
    return loss, embeddings.detach()
Loading

v0.5.1

24 Oct 05:42
@analokmaus analokmaus
v0.5.1
c02e376
Compare
Choose a tag to compare
v0.5.1

What's new

  • fixed the implementation of make_demographic_table() in kuma_utils.stats to hide the first category when the variable is binary.
  • added loveplot() to visualize the balance of variables before/after the propensity score matching.
  • Examples
Loading

v0.5.0

20 May 10:06
@analokmaus analokmaus
v0.5.0
dd452fd
Compare
Choose a tag to compare
v0.5.0

What's new

  • The backpropagation part can now be freely customized, enabling support for a wider range of learning tasks.
  • See implementation of TrainHook
Loading

v0.4.2

25 Feb 06:58
@analokmaus analokmaus
v0.4.2
3b997bc
Compare
Choose a tag to compare
v0.4.2

What's new

  • Wandb run name will be set to TorchTrainer.serial when it is not specified.
Loading

v0.4.1

21 Feb 00:37
@analokmaus analokmaus
v0.4.1
864601a
Compare
Choose a tag to compare
v0.4.1

What's new

  • Bugfixes for TorchTrainer
  • Wandb run name will be TorchTrainer.serial by default. Specify name in wandb_params to override.
Loading

v0.4.0

14 Feb 06:30
@analokmaus analokmaus
v0.4.0
f33232f
Compare
Choose a tag to compare
v0.4.0

Update

LGBMImputer

  • Cross validation support (see tutorial notebook)
  • New argument target_cols, in case you want to fit columns with no missing values.
  • Internal label encoder is added. Now you don't have to encode categorical variables in advance.

TemperatureScaler

  • Moved to kuma_utils.torch.modules.calibration

Others

  • Bugfixes
Loading
0