perming-1.6.0

new traits:

• more optional parameters are set to enable users to revise options of solver and lr_scheduler for consistent with various versions of torch.optim, see _solver and _scheduler in _utils/BaseModel.
• enrich README to release it on PyPI Project Profile, users can read meaning and notices of any module before start to train features and target.
• maintain any crucial process as 1.5.* and 1.4.*, like data_loader, _set_container, early_stop (control by if), _pack_info, and so on, users can adapt to assertions and conditional context for more compatible use for any defined and preprocessed learning task.

example (multi-outputs):

import perming
main = perming.Box(10, 3, (30,), batch_size=8, activation='relu', inplace_on=True, solver='sgd', criterion='MultiLabelSoftMarginLoss', learning_rate_init=0.01)
main.data_loader(X, y, random_seed=0)
main.train_val(num_epochs=60, interval=25, tolerance=1e-4, patience=10, early_stop=True)
main.test()

take main.unique and main.indices to establish bidirectional O(1) convert when tackled with target shape at (n,), and main.model stores the pretrained parameters of model instance.

download:

!pip install perming==1.6.0 # in jupyter
pip install perming==1.6.0 # in cmd

perming-1.5.2

revised details:

• fixed bug about target with shape (n,1) pass through data_loader module when filter self.is_target_2d, if users need to train target with shape (n,1), it might take 'MSELoss' and 'is_float_type=True' to make the execution of such task possible in v1.5.1
• replace show with con in controling whether to print model.state_dict() in save and load module

download:

!pip install perming==1.5.2 # in jupyter
pip install perming==1.5.2 # in cmd

perming-1.5.1

update assertions:

• add assertion under self.is_target_2d to detect and indicate users to convert target with (n,1) to (n,) for subsequent uninterrupted process and suitable type_of_problems.

download:

!pip install perming==1.5.1 # in jupyter
pip install perming==1.5.1 # in cmd

perming-1.5.0

updated traits:

• move accuracy in console print to _pack_info module for test indicators of regression and multi-outputs task are separated with accuracy, users can refer to https://pytorch.org/torcheval/main/ for more evaluation tools, like r2_score.
• init self.is_task_c1d in data_loader ahead of train_val and test module to record the state of 1d task in classification, so reduce combined bool value times than v1.4.*
• sustain any process in train_val module, include early_stop with O(1) method.

download:

!pip install perming==1.5.0 # in jupyter
pip install perming==1.5.0 # in cmd

perming-1.4.3

updated details:

• modified __doc__ for all classes from common and general, including the MLP model and TabularDataset.
• set the left endpoint of the learning_rate interval to 1e-6 to avoid users setting too low value for training.

download:

!pip install perming==1.4.3 # in jupyter
pip install perming==1.4.3 # in cmd

same parameters configuration and interactive process as v1.4.2

perming-1.4.2

vital revise:

• set break in single epoch between next console print [interval, 2inertval] after self.stop_iter -> True.

download:

!pip install perming==1.4.2 # in jupyter
pip install perming==1.4.2 # in cmd

tests for 1.4.* and refer to updated tests in main branch

perming-1.4.1

update traits:

• drop bool_val_first with val_counts==1 to detect start value of val_loss and store it with val_loss_pre.
• set val_pos_ini to note the previous lowest loss to make length of judgement about patience interval more concise.
• maintain O(1) method while optimized tolerance between val_loss_pre and self.val_loss (noted by val_counts) compared to v1.4.0

download:

!pip install perming==1.4.1 # in jupyter
pip install perming==1.4.1 # in cmd

perming-1.4.0

new traits:

• support early-stopping by given tolerance and patience in hyper parameters of train_val.
• adopt O(1) method to reduce the runtime burden after val_loss was accumulated and val_counts was renewed.
• enable users to define the predefined enough epochs and appropriate tolerance while enbale early_stop.

download:

!pip install perming==1.4.0 # in jupyter
pip install perming==1.4.0 # in cmd

perming-1.3.1

fix bugs:

• replace _packing module with _pack_info to make meaning of module more reasonable.
• more comprehensive annotation and explanation of initialization parameters, like activation, criterion.
• improved __doc__ for the test module in the parent class BaseModel to distinguish correct_class and loss of 3 stages.

download:

!pip install perming==1.3.1 # in jupyter
pip install perming==1.3.1 # in cmd

perming-1.3.0

new release:

• support multi-outputs input, like np.array([[0, 1, 1], [0, 1, 0], [0, 0, 1]]) and np.array([[1.1, 2.3, 3.4], [1.6, 2.8, 4.2]]).
• release Ranker to support most multi-outputs task, see test case in tests/Multi-outputs Task.
• optimize data_loader to make multi-outputs input more original, and enable users to establish bidirectional conversion of target.

download:

!pip install perming==1.3.0 # in jupyter
pip install perming==1.3.0 # in cmd