Source code for

from typing import Dict, List, Tuple, Optional, Callable
from pathlib import Path
from fastprogress import master_bar, progress_bar
from fastprogress.fastprogress import IN_NOTEBOOK
import pickle
import timeit
import numpy as np
import os
from functools import partial
from fastcore.all import is_listy

from import FoldYielder
from ..models.model_builder import ModelBuilder
from ..models.model import Model
from ..callbacks.pred_handlers import PredHandler
from ..callbacks.monitors import EarlyStopping, SaveBest, MetricLogger
from ...utils.statistics import uncert_round
from ..metrics.eval_metric import EvalMetric
from import plot_train_history
from ...plotting.plot_settings import PlotSettings

import matplotlib.pyplot as plt

__all__ = ['train_models']

[docs]def train_models(fy:FoldYielder, n_models:int, bs:int, model_builder:ModelBuilder, n_epochs:int, patience:Optional[int]=None, loss_is_meaned:bool=True, cb_partials:Optional[List[partial]]=None, eval_metrics:Optional[Dict[str,EvalMetric]]=None, pred_cb:Callable[[],PredHandler]=PredHandler, train_on_weights:bool=True, bulk_move:bool=True, start_mode_id:int=0, live_fdbk:bool=IN_NOTEBOOK, live_fdbk_first_only:bool=False, live_fdbk_extra:bool=True, live_fdbk_extra_first_only:bool=False, savepath:Path=Path('train_weights'), plot_settings:PlotSettings=PlotSettings()) \ -> Tuple[List[Dict[str,float]],List[Dict[str,List[float]]],List[Dict[str,float]]]: r''' Main training method for :class:`~lumin.nn.models.model.Model`. Trains a specified numer of models created by a :class:`~lumin.nn.models.model_builder.ModelBuilder` on data provided by a :class:``, and saves them to `savepath`. Note, this does not return trained models, instead they are saved and must be loaded later. Instead this method returns results of model training. Each :class:`~lumin.nn.models.model.Model` is trained on N-1 folds, for a :class:`` with N folds, and the remaining fold is used as validation data. Depending on the live_fdbk arguments, live plots of losses and other metrics may be shown during training, if running in Jupyter. Showing the live plot slightly slows down the training, but can help highlight problems without having to wait to the end. If not running in Jupyter, then losses are printed to the terminal. Once training is finished, the state with the lowest validation loss is loaded, evaluated, and saved. Arguments: fy: :class:`` interfacing ot training data n_models: number of models to train bs: batch size. Number of data points per iteration model_builder: :class:`~lumin.nn.models.model_builder.ModelBuilder` creating the networks to train n_epochs: maximum number of epochs for which to train patience: if not `None`, sets the number of epochs or cycles to train without decrease in validation loss before ending training (early stopping) loss_is_meaned: if the batch loss value has been averaged over the number of elements in the batch, this should be true cb_partials: optional list of functools.partial, each of which will a instantiate :class:`~lumin.nn.callbacks.callback.Callback` when called eval_metrics: list of instantiated :class:`~lumin.nn.metric.eval_metric.EvalMetric`. At the end of training, validation data and model predictions will be passed to each, and the results printed and saved pred_cb: pred_cb: :class:`~lumin.nn.callbacks.pred_handlers.PredHandler` callback to determin how predictions are computed. Default simply returns the model predictions. Other uses could be e.g. running argmax on a multiclass classifier train_on_weights: If weights are present in training data, whether to pass them to the loss function during training bulk_move: if true, will optimise for speed by using more RAM and VRAM start_mode_id: model ID at whcih to start training, i.e. if training was interupted, this can be set to resume training form the last model which was trained live_fdbk: whether or not to show any live feedback at all during training (slightly slows down training, but helps spot problems) live_fdbk_first_only: whether to only show live feedback for the first model trained (trade off between time and problem spotting) live_fdbk_extra: whether to show extra information live feedback (further slows training) live_fdbk_extra_first_only: whether to only show extra live feedback information for the first model trained (trade off between time and information) savepath: path to to which to save model weights and results plot_settings: :class:`~lumin.plotting.plot_settings.PlotSettings` class to control figure appearance Returns: - results list of validation losses and other eval_metrics results, ordered by model training. Can be used to create an :class:`~lumin.nn.ensemble.ensemble.Ensemble`. - histories list of loss histories, ordered by model training - cycle_losses if an :class:`~lumin.nn.callbacks.cyclic_callbacks.AbsCyclicCallback` was passed, lists validation losses at the end of each cycle, ordered by model training. Can be passed to :class:`~lumin.nn.ensemble.ensemble.Ensemble`. ''' results,histories,cycle_losses,savepath = [],[],[],Path(savepath) if cb_partials is None: cb_partials = [] if not is_listy(cb_partials): cb_partials = [cb_partials] model_rng = range(start_mode_id, n_models) for i in model_rng: os.system(f"rm -r {savepath}/model_id_{i}") model_bar = master_bar(model_rng) if IN_NOTEBOOK else progress_bar(model_rng) train_tmr = timeit.default_timer() for model_num in (model_bar): if IN_NOTEBOOK: val_idx = model_num % fy.n_folds print(f"Training model {model_num+1} / {n_models}, Val ID = {val_idx}") if model_num == 1: if live_fdbk_first_only: live_fdbk = False # Only show fdbk for first training elif live_fdbk_extra_first_only: live_fdbk_extra = False # Only show full fdbk info for first training model_dir = savepath/f'model_id_{model_num}' model_dir.mkdir(parents=True) model = Model(model_builder) cbs = [] for c in cb_partials: cbs.append(c()) save_best = SaveBest(auto_reload=True, loss_is_meaned=loss_is_meaned) metric_log = MetricLogger(show_plots=live_fdbk, extra_detail=live_fdbk_extra) cbs += [save_best,metric_log] if patience is not None: cbs.append(EarlyStopping(patience=patience, loss_is_meaned=loss_is_meaned)) for c in cbs: c.set_plot_settings(plot_settings) model_tmr = timeit.default_timer(), fy=fy, bs=bs, bulk_move=bulk_move, train_on_weights=train_on_weights, val_idx=val_idx, cbs=cbs, cb_savepath=model_dir) print(f"Model took {timeit.default_timer()-model_tmr:.3f}s\n")'train_{model_num}.h5') histories.append(metric_log.get_loss_history()) cycle_losses.append([]) for c in cbs: if hasattr(c, 'cycle_save') and c.cycle_save: cycle_losses[-1] = c.cycle_losses results.append({}) results[-1]['loss'] = save_best.min_loss if eval_metrics is not None and len(eval_metrics) > 0: y_pred = model.predict(fy[val_idx]['inputs'], bs=bs if not bulk_move else None) for m in eval_metrics: results[-1][m] = eval_metrics[m].evaluate(fy, val_idx, y_pred) print(f"Scores are: {results[-1]}") results[-1]['path'] = model_dir with open(savepath/'results_file.pkl', 'wb') as fout: pickle.dump(results, fout) with open(savepath/'cycle_file.pkl', 'wb') as fout: pickle.dump(cycle_losses, fout) plt.clf() print("\n______________________________________") print("Training finished") print(f"Cross-validation took {timeit.default_timer()-train_tmr:.3f}s ") plot_train_history(histories, savepath/'loss_history', settings=plot_settings, show=IN_NOTEBOOK, log_y='regress' in model_builder.objective) for score in results[0]: if score == 'path': continue mean = uncert_round(np.mean([x[score] for x in results]), np.std([x[score] for x in results])/np.sqrt(len(results))) print(f"Mean {score} = {mean[0]}±{mean[1]}") print("______________________________________\n") return results, histories, cycle_losses
Read the Docs v: v0.7.0
On Read the Docs
Project Home

Free document hosting provided by Read the Docs.


Access comprehensive developer and user documentation for LUMIN

View Docs


Get tutorials for beginner and advanced researchers demonstrating many of the features of LUMIN

View Tutorials