mkkb473 / mwutils

Introduction

mwutils is a package developed to be used in the ModelWhale platform for machine learning experiments tracking. It is now under active development.

At the moment, mwutils supports keras and PyTorch mwutils has experimental support for Tensorflow

Usage

class Run: methods:

• init(self, name="lab_run", user_id="user1", lab_id="lab1", org_id="", flush_interval_seconds=5, sys_stat_sample_size=5, sys_stat_sample_interval=5, local_path='', write_logs_to_local=False, remote_path='', buffer_all_logs=False):

  • name: name for current experiment
  • user_id, lab_id, org_id: user shall get these information from the notebook tools panel in ModelWhale Notebook interface.
  • flush_interval_seconds: time interval to flush buffered logs, default is 5
  • sys_stat_sample_size: maximum data size for system metrics, default is 5
  • sys_stat_sample_interval: sampling rate for system metrics, default is 5
  • write_logs_to_local: if logs should be stored locally
  • local_path: local path to store logs
  • remote_path: kesci api
  • buffer_all_logs: if all logs should be buffered in RAM (use with caution)

• init_ml():

• start_ml():

• log_ml(step=None, epoch=None, batch=None, loss=None, acc=None, phase="train"):

  • step, epoch, batch: starting from 1;
  • loss: numpy.float32 woulde be converted to float type;
  • acc: numpy.float32 woulde be converted to float type;
  • phase: default is train, test, val, system

• conclude(show_memoize=True): call function after training end

  • show_memoize:

• add_memoize_funcs_to_logger: testing;

Keras:

provide MWCustomCallback method

example

from mwutils.keras import MWCustomCallback
from mwutils.run import Run

r = Run(name=RUN_NAME,
        user_id = $user_id,
        lab_id = $lab_id,
        org_id = $org_id,
        flush_interval_seconds=30,
        sys_stat_sample_size=5,
        sys_stat_sample_interval=5,
        local_path='',
        write_logs_to_local=False,
        remote_path= $remote_path,
        buffer_all_logs=True)
callBack = MWCustomCallback()
callBack.set_run(r)        
history = model.fit(X_train, y_train,
                    epochs = 40,
                    batch_size = 32,
                    validation_data=(X_test,y_test),
                    shuffle=True,
                    callbacks=[callBack])
r.conclude()

PyTorch:

provide LoggerHook method

example

from mwutils.torch_utils import LoggerHook
from mwutils.run import Run

r = Run(name=RUN_NAME,
        user_id = $user_id,
        lab_id = $lab_id,
        org_id = $org_id,
        flush_interval_seconds=30,
        sys_stat_sample_size=5,
        sys_stat_sample_interval=5,
        local_path='',
        write_logs_to_local=False,
        remote_path= $remote_path,
        buffer_all_logs=True)
loggerHook = LoggerHook()
loggerHook.set_run(r)
criterion.register_forward_hook(loggerHook.torch_loss_hook)

r.conclude()