base_trainer.py 13.8 KB
"""trainer code"""
import copy
import logging
import os
from typing import List, Dict, Optional, Callable, Union

import dill
import numpy as np
import torch
from torch.utils.tensorboard import SummaryWriter

from lib.utils.logging import loss_logger_helper

logger = logging.getLogger()


class Trainer:
    # This is like skorch but instead of callbacks we use class functions (looks less magic)
    # this is an evolving template
    def __init__(
            self,
            model: torch.nn.Module,
            optimizer: torch.optim,
            scheduler: torch.optim.lr_scheduler,
            result_dir: Optional[str],
            statefile: Optional[str] = None,
            log_every: int = 100,
            save_strategy: Optional[List] = None,
            patience: int = 20,
            max_epoch: int = 100,
            gradient_norm_clip=-1,
            stopping_criteria_direction: str = "bigger",
            stopping_criteria: Optional[Union[str, Callable]] = "accuracy",
            evaluations=None,
            **kwargs,
    ):
        """
            stopping_criteria : can be a function, string or none. If string it should match one
            of the keys in aux_loss or should be loss, if none we don't invoke early stopping
        """
        super().__init__()

        self.result_dir = result_dir
        self.model = model
        self.optimizer = optimizer
        self.scheduler = scheduler
        self.evaluations = evaluations
        self.gradient_norm_clip = gradient_norm_clip

        # training state related params
        self.epoch = 0
        self.step = 0
        self.best_criteria = None
        self.best_epoch = -1

        # config related param
        self.log_every = log_every
        self.save_strategy = save_strategy
        self.patience = patience
        self.max_epoch = max_epoch
        self.stopping_criteria_direction = stopping_criteria_direction
        self.stopping_criteria = stopping_criteria

        # TODO: should save config and see if things have changed?
        if statefile is not None:
            self.load(statefile)

        # init best model
        self.best_model = self.model.state_dict()

        # logging stuff
        if result_dir is not None:
            # we do not need to purge. Purging can delete the validation result
            self.summary_writer = SummaryWriter(log_dir=result_dir)

    def load(self, fname: str) -> Dict:
        """
            fname: file name to load data from
        """

        data = torch.load(open(fname, "rb"), pickle_module=dill, map_location=self.model.device)

        if getattr(self, "model", None) and data.get("model") is not None:
            state_dict = self.model.state_dict()
            state_dict.update(data["model"])
            self.model.load_state_dict(state_dict)

        if getattr(self, "optimizer", None) and data.get("optimizer") is not None:
            optimizer_dict = self.optimizer.state_dict()
            optimizer_dict.update(data["optimizer"])
            self.optimizer.load_state_dict(optimizer_dict)

        if getattr(self, "scheduler", None) and data.get("scheduler") is not None:
            scheduler_dict = self.scheduler.state_dict()
            scheduler_dict.update(data["scheduler"])
            self.scheduler.load_state_dict(scheduler_dict)

        self.epoch = data["epoch"]
        self.step = data["step"]
        self.best_criteria = data["best_criteria"]
        self.best_epoch = data["best_epoch"]
        return data

    def save(self, fname: str, **kwargs):
        """
        fname: file name to save to
        kwargs: more arguments that we may want to save.

        By default we
            - save,
            - model,
            - optimizer,
            - epoch,
            - step,
            - best_criteria,
            - best_epoch
        """
        # NOTE: Best model is maintained but is saved automatically depending on save strategy,
        # So that It could be loaded outside of the training process
        kwargs.update({
                "model"        : self.model.state_dict(),
                "optimizer"    : self.optimizer.state_dict(),
                "epoch"        : self.epoch,
                "step"         : self.step,
                "best_criteria": self.best_criteria,
                "best_epoch"   : self.best_epoch,
        })

        if self.scheduler is not None:
            kwargs.update({"scheduler": self.scheduler.state_dict()})

        torch.save(kwargs, open(fname, "wb"), pickle_module=dill)

    # todo : allow to extract predictions
    def run_iteration(self, batch, training: bool = True, reduce: bool = True):
        """
            batch : batch of data, directly passed to model as is
            training: if training set to true else false
            reduce: whether to compute loss mean or return the raw vector form
        """
        pred = self.model(batch)
        loss, aux_loss = self.model.loss(pred, batch, reduce=reduce)
        print(pred)

        if training:
            print(pred)
            loss.backward()
            if self.gradient_norm_clip > 0:
                torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.gradient_norm_clip)
            self.optimizer.step()
            self.optimizer.zero_grad()

        return loss, aux_loss

    def compute_criteria(self, loss, aux_loss):
        stopping_criteria = self.stopping_criteria
        if stopping_criteria is None:
            return loss

        if callable(stopping_criteria):
            return stopping_criteria(loss, aux_loss)

        if stopping_criteria == "loss":
            return loss

        if aux_loss.get(stopping_criteria) is not None:
            return aux_loss[stopping_criteria]

        raise Exception(f"{stopping_criteria} not found")

    def train_batch(self, batch, *args, **kwargs):
        # This trains the batch
        loss, aux_loss = self.run_iteration(batch, training=True, reduce=True)
        loss_logger_helper(loss, aux_loss, writer=self.summary_writer, step=self.step,
                           epoch=self.epoch,
                           log_every=self.log_every, string="train")

    def train_epoch(self, train_loader, *args, **kwargs):
        # This trains the epoch and also calls on batch begin and on batch end
        # before and after calling train_batch respectively
        self.model.train()
        for i, batch in enumerate(train_loader):
            self.on_batch_begin(i, batch, *args, **kwargs)
            self.train_batch(batch, *args, **kwargs)
            self.on_batch_end(i, batch, *args, **kwargs)
            self.step += 1
        self.model.eval()

    def on_train_begin(self, train_loader, valid_loader, *args, **kwargs):
        # this could be used to add things to class object like scheduler etc
        if "init" in self.save_strategy:
            if self.epoch == 0:
                self.save(f"{self.result_dir}/init_model.pt")

    def on_epoch_begin(self, train_loader, valid_loader, *args, **kwargs):
        # This is called when epoch begins
        pass

    def on_batch_begin(self, epoch_step, batch, *args, **kwargs):
        # This is called when batch begins
        pass

    def on_train_end(self, train_loader, valid_loader, *args, **kwargs):
        # Called when training finishes. For base trainer we just save the last model
        if "last" in self.save_strategy:
            logger.info("Saving the last model")
            self.save(f"{self.result_dir}/last_model.pt")

    def on_epoch_end(self, train_loader, valid_loader, *args, **kwargs):
        # called when epoch ends
        # we call validation, scheduler here
        # also check if we have a new best model and save model if needed

        # call validate
        loss, aux_loss = self.validate(train_loader, valid_loader, *args, **kwargs)
        loss_logger_helper(loss, aux_loss, writer=self.summary_writer, step=self.step,
                           epoch=self.epoch, log_every=self.log_every, string="val",
                           force_print=True)

        # do scheduler step
        if self.scheduler is not None:
            prev_lr = [group['lr'] for group in self.optimizer.param_groups]
            if isinstance(self.scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau):
                criteria = self.compute_criteria(loss, aux_loss)
                self.scheduler.step(criteria)
            else:
                self.scheduler.step()
            new_lr = [group['lr'] for group in self.optimizer.param_groups]

        # if you don't pass a criteria, it won't be computed and best model won't be saved.
        # on the contrary if you pass a stopping criteria, best model would be saved.
        # You can pass a large patience to get rid of early stopping
        if self.stopping_criteria is not None:
            criteria = self.compute_criteria(loss, aux_loss)

            if (
                    (self.best_criteria is None)
                    or (
                    self.stopping_criteria_direction == "bigger" and self.best_criteria < criteria)
                    or (
                    self.stopping_criteria_direction == "lower" and self.best_criteria > criteria)
            ):
                self.best_criteria = criteria
                self.best_epoch = self.epoch
                self.best_model = copy.deepcopy(
                    {k: v.cpu() for k, v in self.model.state_dict().items()})

                if "best" in self.save_strategy:
                    logger.info(f"Saving best model at epoch {self.epoch}")
                    self.save(f"{self.result_dir}/best_model.pt")

        if "epoch" in self.save_strategy:
            logger.info(f"Saving model at epoch {self.epoch}")
            self.save(f"{self.result_dir}/{self.epoch}_model.pt")

        if "current" in self.save_strategy:
            logger.info(f"Saving model at epoch {self.epoch}")
            self.save(f"{self.result_dir}/current_model.pt")

        # logic to load best model on reduce lr
        if self.scheduler is not None and not (all(a == b for (a, b) in zip(prev_lr, new_lr))):
            if getattr(self.scheduler, 'load_on_reduce', None) == "best":
                logger.info(f"Loading best model at epoch {self.epoch}")
                # we want to preserve the scheduler
                old_lrs = list(map(lambda x: x['lr'], self.optimizer.param_groups))
                old_scheduler_dict = copy.deepcopy(self.scheduler.state_dict())

                best_model_path = None
                if os.path.exists(f"{self.result_dir}/best_model.pt"):
                    best_model_path = f"{self.result_dir}/best_model.pt"
                else:
                    d = "/".join(self.result_dir.split("/")[:-1])
                    for directory in os.listdir(d):
                        if os.path.exists(f"{d}/{directory}/best_model.pt"):
                            best_model_path = self.load(f"{d}/{directory}/best_model.pt")

                if best_model_path is None:
                    raise FileNotFoundError(
                        f"Best Model not found in {self.result_dir}, please copy if it exists in "
                        f"other folder")

                self.load(best_model_path)
                # override scheduler to keep old one and also keep reduced learning rates
                self.scheduler.load_state_dict(old_scheduler_dict)
                for idx, lr in enumerate(old_lrs):
                    self.optimizer.param_groups[idx]['lr'] = lr
                logger.info(f"loaded best model and restarting from end of {self.epoch}")

    def on_batch_end(self, epoch_step, batch, *args, **kwargs):
        # called after a batch is trained
        pass

    def train(self, train_loader, valid_loader, *args, **kwargs):

        self.on_train_begin(train_loader, valid_loader, *args, **kwargs)
        while self.epoch < self.max_epoch:
            # NOTE: +1 here is more convenient, as now we don't need to do +1 before saving model
            # If we don't do +1 before saving model, we will have to redo the last epoch
            # So +1 here makes life easy, if we load model at end of e epoch, we will load model
            # and start with e+1... smooth
            self.epoch += 1
            self.on_epoch_begin(train_loader, valid_loader, *args, **kwargs)
            logger.info(f"Starting epoch {self.epoch}")
            self.train_epoch(train_loader, *args, **kwargs)
            self.on_epoch_end(train_loader, valid_loader, *args, **kwargs)

            if self.epoch - self.best_epoch > self.patience:
                logger.info(f"Patience reached stopping training after {self.epoch} epochs")
                break

        self.on_train_end(train_loader, valid_loader, *args, **kwargs)

    def validate(self, train_loader, valid_loader, *args, **kwargs):
        """
        we expect validate to return mean and other aux losses that we want to log
        """
        losses = []
        aux_losses = {}

        self.model.eval()
        with torch.no_grad():
            for i, batch in enumerate(valid_loader):
                loss, aux_loss = self.run_iteration(batch, training=False, reduce=False)
                losses.extend(loss.cpu().tolist())

                if i == 0:
                    for k, v in aux_loss.items():
                        # when we can't return sample wise statistics, we need to do this
                        if len(v.shape) == 0:
                            aux_losses[k] = [v.cpu().tolist()]
                        else:
                            aux_losses[k] = v.cpu().tolist()
                else:
                    for k, v in aux_loss.items():
                        if len(v.shape) == 0:
                            aux_losses[k].append(v.cpu().tolist())
                        else:
                            aux_losses[k].extend(v.cpu().tolist())
        return np.mean(losses), {k: np.mean(v) for (k, v) in aux_losses.items()}

    def test(self, train_loader, test_loader, *args, **kwargs):
        return self.validate(train_loader, test_loader, *args, **kwargs)