train_with_prioritised_buffer.py

from typing import Callable, Any, Optional, List

import torch.optim.optimizer
from tqdm import tqdm
import numpy as np
import matplotlib.pyplot as plt
import pathlib
from time import time
import os

from fab.utils.logging import Logger, ListLogger
from fab.core import FABModel
from fab.utils.prioritised_replay_buffer import PrioritisedReplayBuffer


lr_scheduler = Any  # a learning rate schedular from torch.optim.lr_scheduler
Plotter = Callable[[FABModel], List[plt.Figure]]


class PrioritisedBufferTrainer:
    """A trainer for the FABModel for use with a prioritised replay buffer, and a different form
    of loss. In this training loop we target p^\alpha / q^(\alpha - 1) instead of p."""
    def __init__(self,
                 model: FABModel,
                 optimizer: torch.optim.Optimizer,
                 buffer: PrioritisedReplayBuffer,
                 alpha: float,
                 n_batches_buffer_sampling: int = 2,
                 optim_schedular: Optional[lr_scheduler] = None,
                 logger: Logger = ListLogger(),
                 plot: Optional[Plotter] = None,
                 max_gradient_norm: Optional[float] = 5.0,
                 w_adjust_max_clip: Optional[float] = 10.0,
                 w_adjust_in_buffer_after_update: bool = False,
                 save_path: str = "",
                 ):
        self.model = model
        self.alpha = alpha

        # Ensure we have p^\alpha q^{1-\alpha} as the AIS target distribution.
        self.model.p_target = False
        self.model.annealed_importance_sampler.p_target = False

        self.optimizer = optimizer
        self.optim_schedular = optim_schedular
        self.logger = logger
        self.plot = plot
        # if no gradient clipping set max_gradient_norm to inf
        self.max_gradient_norm = max_gradient_norm if max_gradient_norm else float("inf")
        self.save_dir = save_path
        self.plots_dir = os.path.join(self.save_dir, f"plots")
        self.checkpoints_dir = os.path.join(self.save_dir, f"model_checkpoints")
        self.buffer = buffer
        self.n_batches_buffer_sampling = n_batches_buffer_sampling
        self.flow_device = next(model.flow.parameters()).device
        self.max_adjust_w_clip = w_adjust_max_clip
        self.w_adjust_in_buffer_after_update = w_adjust_in_buffer_after_update

    def save_checkpoint(self, i):
        checkpoint_path = os.path.join(self.checkpoints_dir, f"iter_{i}/")
        pathlib.Path(checkpoint_path).mkdir(exist_ok=False)
        self.model.save(os.path.join(checkpoint_path, "model.pt"))
        torch.save(self.optimizer.state_dict(),
                   os.path.join(checkpoint_path, 'optimizer.pt'))
        self.buffer.save(os.path.join(checkpoint_path, 'buffer.pt'))
        if self.optim_schedular:
            torch.save(self.optim_schedular.state_dict(),
                       os.path.join(self.checkpoints_dir, 'scheduler.pt'))

    def make_and_save_plots(self, i, save):
        figures = self.plot(self.model)
        for j, figure in enumerate(figures):
            if save:
                figure.savefig(os.path.join(self.plots_dir, f"{j}_iter_{i}.png"))
            else:
                plt.show()
            plt.close(figure)

    def perform_eval(self, i, eval_batch_size, batch_size):
        # Set ais distribution to target for evaluation of ess, freeze transition operator params.
        self.model.annealed_importance_sampler.transition_operator.set_eval_mode(True)
        eval_info_true_target = self.model.get_eval_info(outer_batch_size=eval_batch_size,
                                                         inner_batch_size=batch_size,
                                                         set_p_target=True)
        # Double check the ais distribution has been set back to p^\alpha q^{1-\alpha}.
        assert self.model.annealed_importance_sampler.p_target is False
        assert self.model.annealed_importance_sampler.transition_operator.p_target is False
        # Evaluation with the AIS ESS with target set as p^\alpha q^{1-\alpha}.
        eval_info_practical_target = self.model.get_eval_info(outer_batch_size=eval_batch_size,
                                                              inner_batch_size=batch_size,
                                                              set_p_target=False,
                                                              ais_only=True)
        self.model.annealed_importance_sampler.transition_operator.set_eval_mode(False)


        eval_info = {}
        eval_info.update({key + "_p_target": val for key, val in eval_info_true_target.items()})
        eval_info.update(
            {key + "_min_var_target": val for key, val in eval_info_practical_target.items()})

        eval_info.update(step=i)
        self.logger.write(eval_info)



    def run(self,
            n_iterations: int,
            batch_size: int,
            eval_batch_size: Optional[int] = None,
            n_eval: Optional[int] = None,
            n_plot: Optional[int] = None,
            n_checkpoints: Optional[int] = None,
            save: bool = True,
            tlimit: Optional[float] = None,
            start_time: Optional[float] = None,
            start_iter: Optional[int] = 0) -> None:
        if save:
            pathlib.Path(self.plots_dir).mkdir(exist_ok=True)
            pathlib.Path(self.checkpoints_dir).mkdir(exist_ok=True)
        if n_checkpoints:
            checkpoint_iter = list(np.linspace(1, n_iterations, n_checkpoints, dtype="int"))
        if n_eval is not None:
            eval_iter = list(np.linspace(1, n_iterations, n_eval, dtype="int"))
            assert eval_batch_size is not None
        if n_plot is not None:
            plot_iter = list(np.linspace(1, n_iterations, n_plot, dtype="int"))
        if tlimit is not None:
            assert n_checkpoints is not None, "Time limited specified but not checkpoints are " \
                                          "being saved."
        if start_time is not None:
            start_time = time()

        if start_iter >= n_iterations:
            raise Exception("Not running training as start_iter >= total training iterations")

        pbar = tqdm(range(n_iterations - start_iter))
        max_it_time = 0.0
        for pbar_iter in pbar:
            i = pbar_iter + start_iter + 1
            it_start_time = time()
            self.optimizer.zero_grad()
            # collect samples and log weights with AIS and add to the buffer
            point_ais, log_w_ais = self.model.\
                annealed_importance_sampler.sample_and_log_weights(batch_size)
            x_ais = point_ais.x.detach()
            log_w_ais = log_w_ais.detach()
            log_q_x_ais = point_ais.log_q.detach()
            self.buffer.add(x_ais.detach(), log_w_ais.detach(),
                            log_q_x_ais.detach())

            # we log info from the step of the recently generated ais points.
            info = self.model.get_iter_info()

            # We now take self.n_batches_buffer_sampling gradient steps using
            # data from the replay buffer.
            mini_dataset = self.buffer.sample_n_batches(
                    batch_size=batch_size, n_batches=self.n_batches_buffer_sampling)
            for (x, log_w, log_q_old, indices) in mini_dataset:
                x, log_w, log_q_old, indices = x.to(self.flow_device), log_w.to(self.flow_device), \
                                               log_q_old.to(self.flow_device), indices.to(self.flow_device)
                self.optimizer.zero_grad()
                log_q_x = self.model.flow.log_prob(x)
                # adjustment to account for change to theta since sample was last added/adjusted
                log_w_adjust = (1-self.alpha) * (log_q_x.detach() - log_q_old)
                w_adjust_pre_clip = torch.exp(log_w_adjust)  # no grad
                if self.max_adjust_w_clip is not None:
                    w_adjust = torch.clip(w_adjust_pre_clip, max=self.max_adjust_w_clip)
                else:
                    w_adjust = w_adjust_pre_clip
                # manually calculate the new form of the loss
                loss = - torch.mean(w_adjust * log_q_x)
                if not torch.isnan(loss) and not torch.isinf(loss):
                    loss.backward()
                    grad_norm = torch.nn.utils.clip_grad_norm_(self.model.parameters(),
                                                               self.max_gradient_norm)
                    if torch.isfinite(grad_norm):
                        self.optimizer.step()
                    else:
                        print("nan grad norm in replay step")
                else:
                    print("nan loss in replay step")

                # Adjust log weights in the buffer on the fly.
                if not self.w_adjust_in_buffer_after_update:
                    with torch.no_grad():
                        self.buffer.adjust(log_w_adjust, log_q_x, indices)


            info.update(loss=loss.cpu().detach().item(),
                        step=i,
                        grad_norm=grad_norm.cpu().detach().item(),
                        sampled_log_w_std=torch.std(log_w).detach().cpu().item(),
                        sampled_log_w_mean=torch.mean(log_w).detach().cpu().item(),
                        w_adjust_mean=torch.mean(w_adjust_pre_clip).detach().cpu().item(),
                        w_adjust_min=torch.min(w_adjust_pre_clip).detach().cpu().item(),
                        w_adjust_max=torch.max(w_adjust_pre_clip).detach().cpu().item(),
                        log_q_x_mean=torch.mean(log_q_x).cpu().item()
                        )

            if self.w_adjust_in_buffer_after_update:
                with torch.no_grad():
                    for (x, log_w, log_q_old, indices) in mini_dataset:
                        """Adjust importance weights in the buffer for the points in the 
                        `mini_dataset` to account for the updated theta."""
                        x, log_w, log_q_old, indices = x.to(self.flow_device), log_w.to(
                            self.flow_device), log_q_old.to(self.flow_device), indices.to(
                            self.flow_device)
                        log_q_new = self.model.flow.log_prob(x)
                        log_w_adjust_insert = (1 - self.alpha) * (log_q_new - log_q_old)
                        self.buffer.adjust(log_w_adjust_insert, log_q_new, indices)
                    info.update(
                        log_w_adjust_insert_mean = torch.mean
                        (log_w_adjust_insert).detach().cpu().item(),
                        log_q_mean = torch.mean(log_q_new).detach().cpu().item())

            self.logger.write(info)
            pbar.set_description(f"loss: {loss.cpu().detach().item()}, ess base: {info['ess_base']},"
                                 f"ess ais: {info['ess_ais']}")

            if n_eval is not None:
                if i in eval_iter:
                    self.perform_eval(i, eval_batch_size, batch_size)

            if n_plot is not None:
                if i in plot_iter:
                    self.make_and_save_plots(i, save)

            if n_checkpoints is not None:
                if i in checkpoint_iter:
                    self.save_checkpoint(i)


            max_it_time = max(max_it_time, time() - it_start_time)

            # End job if necessary
            if tlimit is not None:
                time_past = (time() - start_time) / 3600
                if (time_past + max_it_time/3600) > tlimit:
                    # self.perform_eval(i, eval_batch_size, batch_size)
                    # self.make_and_save_plots(i, save)
                    if i not in checkpoint_iter:
                        self.save_checkpoint(i)
                    self.logger.close()
                    print(f"\nEnding training at iteration {i}, after training for {time_past:.2f} "
                          f"hours as timelimit {tlimit:.2f} hours has been reached.\n")
                    return


        if tlimit is None:
            print("Timelimit not set")
        else:
            print(f"\n Run completed in {(time() - start_time) / 3600:.2f} hours \n")
            print(f"Run finished before timelimit of {tlimit:.2f} hours was reached. \n")

        self.logger.close()