engine/solver.py

import os
import sys
import time
import torch
import numpy as np

from pathlib import Path
from tqdm.auto import tqdm
from ema_pytorch import EMA
from torch.optim import Adam
from torch.nn.utils import clip_grad_norm_
from utils.io_utils import instantiate_from_config, get_model_parameters_info


sys.path.append(os.path.join(os.path.dirname(__file__), "../"))


def cycle(dl):
    while True:
        for data in dl:
            yield data


class Trainer(object):
    def __init__(self, config, args, model, dataloader, logger=None):
        super().__init__()

        if os.getenv("WANDB_ENABLED") == "true":
            import wandb

            self.run = wandb.init(project="tiffusion-revenue", config=config)
        else:
            self.run = None

        self.model = model
        self.device = self.model.betas.device
        self.train_num_steps = config["solver"]["max_epochs"]
        self.gradient_accumulate_every = config["solver"]["gradient_accumulate_every"]
        self.save_cycle = config["solver"]["save_cycle"]
        self.dl = cycle(dataloader["dataloader"])
        self.step = 0
        self.milestone = 0
        self.args = args
        self.logger = logger

        self.results_folder = Path(
            config["solver"]["results_folder"] + f"_{model.seq_length}"
        )
        os.makedirs(self.results_folder, exist_ok=True)

        start_lr = config["solver"].get("base_lr", 1.0e-4)
        ema_decay = config["solver"]["ema"]["decay"]
        ema_update_every = config["solver"]["ema"]["update_interval"]

        self.opt = Adam(
            filter(lambda p: p.requires_grad, self.model.parameters()),
            lr=start_lr,
            betas=[0.9, 0.96],
        )
        self.ema = EMA(self.model, beta=ema_decay, update_every=ema_update_every).to(
            self.device
        )

        sc_cfg = config["solver"]["scheduler"]
        sc_cfg["params"]["optimizer"] = self.opt
        self.sch = instantiate_from_config(sc_cfg)

        if self.logger is not None:
            self.logger.log_info(str(get_model_parameters_info(self.model)))
        self.log_frequency = 100

    def save(self, milestone, verbose=False):
        if self.logger is not None and verbose:
            self.logger.log_info(
                "Save current model to {}".format(
                    str(self.results_folder / f"checkpoint-{milestone}.pt")
                )
            )
        data = {
            "step": self.step,
            "model": self.model.state_dict(),
            "ema": self.ema.state_dict(),
            "opt": self.opt.state_dict(),
        }
        torch.save(data, str(self.results_folder / f"checkpoint-{milestone}.pt"))

    def load(self, milestone, verbose=False, from_folder=None):
        if self.logger is not None and verbose:
            self.logger.log_info(
                "Resume from {}".format(
                    os.path.join(from_folder, f"checkpoint-{milestone}.pt")
                )
            )
        device = self.device
        data = torch.load(
            os.path.join(from_folder,f"checkpoint-{milestone}.pt") if from_folder else str(self.results_folder / f"checkpoint-{milestone}.pt"), map_location=device, weights_only=True
        )
        self.model.load_state_dict(data["model"], )
        self.step = data["step"]
        self.opt.load_state_dict(data["opt"])
        self.ema.load_state_dict(data["ema"])
        self.milestone = milestone

    def train(self):
        device = self.device
        step = 0
        if self.logger is not None:
            tic = time.time()
            self.logger.log_info(
                "{}: start training...".format(self.args.name), check_primary=False
            )

        with tqdm(initial=step, total=self.train_num_steps) as pbar:
            while step < self.train_num_steps:
                total_loss = 0.0
                for _ in range(self.gradient_accumulate_every):
                    data = next(self.dl).to(device)
                    loss = self.model(data, target=data)
                    loss = loss / self.gradient_accumulate_every
                    loss.backward()
                    total_loss += loss.item()

                pbar.set_description(
                    f'loss: {total_loss:.6f} lr: {self.opt.param_groups[0]["lr"]:.6f}'
                )
                if self.run is not None:
                    wandb.log(
                        {
                            "step": step,
                            "loss": total_loss,
                            "lr": self.opt.param_groups[0]["lr"],
                        },
                        step=self.step,
                    )

                clip_grad_norm_(self.model.parameters(), 1.0)
                self.opt.step()
                self.sch.step(total_loss)
                self.opt.zero_grad()
                self.step += 1
                step += 1
                self.ema.update()

                with torch.no_grad():
                    if self.step != 0 and self.step % self.save_cycle == 0:
                        self.milestone += 1
                        self.save(self.milestone)
                        # self.logger.log_info('saved in {}'.format(str(self.results_folder / f'checkpoint-{self.milestone}.pt')))

                    if self.logger is not None and self.step % self.log_frequency == 0:
                        # info = '{}: train'.format(self.args.name)
                        # info = info + ': Epoch {}/{}'.format(self.step, self.train_num_steps)
                        # info += ' ||'
                        # info += '' if loss_f == 'none' else ' Fourier Loss: {:.4f}'.format(loss_f.item())
                        # info += '' if loss_r == 'none' else ' Reglarization: {:.4f}'.format(loss_r.item())
                        # info += ' | Total Loss: {:.6f}'.format(total_loss)
                        # self.logger.log_info(info)
                        self.logger.add_scalar(
                            tag="train/loss",
                            scalar_value=total_loss,
                            global_step=self.step,
                        )

                pbar.update(1)

        print("training complete")
        if self.logger is not None:
            self.logger.log_info(
                "Training done, time: {:.2f}".format(time.time() - tic)
            )

    def sample(self, num, size_every, shape=None):
        if self.logger is not None:
            tic = time.time()
            self.logger.log_info("Begin to sample...")
        samples = np.empty([0, shape[0], shape[1]])
        num_cycle = int(num // size_every) + 1

        for _ in range(num_cycle):
            sample = self.ema.ema_model.generate_mts(batch_size=size_every)
            samples = np.row_stack([samples, sample.detach().cpu().numpy()])
            torch.cuda.empty_cache()

        if self.logger is not None:
            self.logger.log_info(
                "Sampling done, time: {:.2f}".format(time.time() - tic)
            )
        return samples

    def control_sample(self, num, size_every, shape=None, model_kwargs={}, target=None, partial_mask=None):
        samples = np.empty([0, shape[0], shape[1]])
        import math
        num_cycle = math.ceil(num / size_every)
        assert not ((target is None) ^ (partial_mask is None)), "target and partial_mask should be provided"
        if self.logger is not None:
            tic = time.time()
            self.logger.log_info("Begin to infill sample...")

        target = torch.tensor(target).to(self.device) if target is not None else torch.zeros(shape).to(self.device)
        target = target.repeat(size_every, 1, 1) if len(target.shape) == 2 else target
        partial_mask = torch.tensor(partial_mask).to(self.device) if partial_mask is not None else torch.zeros(shape).to(self.device)
        partial_mask = partial_mask.repeat(size_every, 1, 1) if len(partial_mask.shape) == 2 else partial_mask

        for _ in range(num_cycle):
            sample = self.ema.ema_model.generate_mts_infill(target, partial_mask, model_kwargs=model_kwargs)
            samples = np.row_stack([samples, sample.detach().cpu().numpy()])
            torch.cuda.empty_cache()

        if self.logger is not None:
            self.logger.log_info(
                "Sampling done, time: {:.2f}".format(time.time() - tic)
            )
        return samples

    def predict(

        self,

        observed_points: torch.Tensor,

        coef=1e-1,

        stepsize=1e-1,

        sampling_steps=50,

        **kargs,

    ):
        model_kwargs = {}
        model_kwargs["coef"] = coef
        model_kwargs["learning_rate"] = stepsize
        model_kwargs = {**model_kwargs, **kargs}
        assert len(observed_points.shape) == 2, "observed_points should be 2D, batch size = 1"
        x = observed_points.unsqueeze(0)
        t_m = x != 0
        x = x * 2 - 1 # normalize
        x, t_m = x.to(self.device), t_m.to(self.device)
        if sampling_steps == self.model.num_timesteps:
            print("normal sampling")
            sample = self.ema.ema_model.sample_infill(
                shape=x.shape,
                target=x * t_m,
                partial_mask=t_m,
                model_kwargs=model_kwargs,
            )
            # x: partially noise : (batch_size, seq_length, feature_dim)
        else:
            print("fast sampling")
            sample = self.ema.ema_model.fast_sample_infill(
                shape=x.shape,
                target=x * t_m,
                partial_mask=t_m,
                model_kwargs=model_kwargs,
                sampling_timesteps=sampling_steps,
            )
        # unnormalize
        sample = (sample + 1) / 2
        return sample.squeeze(0).detach().cpu().numpy()
    
    
    def predict_weighted_points(

        self,

        observed_points: torch.Tensor,

        observed_mask: torch.Tensor,

        coef=1e-1,

        stepsize=1e-1,

        sampling_steps=50,

        **kargs,

    ):
        model_kwargs = {}
        model_kwargs["coef"] = coef
        model_kwargs["learning_rate"] = stepsize
        model_kwargs = {**model_kwargs, **kargs}
        assert len(observed_points.shape) == 2, "observed_points should be 2D, batch size = 1"
        x = observed_points.unsqueeze(0)
        float_mask = observed_mask.unsqueeze(0) # x != 0, 1 for observed, 0 for missing, bool tensor
        binary_mask = float_mask.clone()
        binary_mask[binary_mask > 0] = 1

        x = x * 2 - 1 # normalize
        x, float_mask, binary_mask = x.to(self.device), float_mask.to(self.device), binary_mask.to(self.device)
        if sampling_steps == self.model.num_timesteps:
            print("normal sampling")
            raise NotImplementedError
            sample = self.ema.ema_model.sample_infill_float_mask(
                shape=x.shape,
                target=x * binary_mask, # x * t_m, 1 for observed, 0 for missing
                partial_mask=float_mask,
                model_kwargs=model_kwargs,
            )
            # x: partially noise : (batch_size, seq_length, feature_dim)
        else:
            print("fast sampling")
            sample = self.ema.ema_model.fast_sample_infill_float_mask(
                shape=x.shape,
                target=x * binary_mask, # x * t_m, 1 for observed, 0 for missing
                partial_mask=float_mask,
                model_kwargs=model_kwargs,
                sampling_timesteps=sampling_steps,
            )

        # unnormalize
        sample = (sample + 1) / 2
        return sample.squeeze(0).detach().cpu().numpy()
    
    def restore(

        self,

        raw_dataloader,

        shape=None,

        coef=1e-1,

        stepsize=1e-1,

        sampling_steps=50,

        **kargs,

    ):
        if self.logger is not None:
            tic = time.time()
            self.logger.log_info("Begin to restore...")
        model_kwargs = {}
        model_kwargs["coef"] = coef
        model_kwargs["learning_rate"] = stepsize
        model_kwargs = {**model_kwargs, **kargs}
        test = kargs.get("test", False)

        samples = np.empty([0, shape[0], shape[1]])  # seq_length, feature_dim
        reals = np.empty([0, shape[0], shape[1]])
        masks = np.empty([0, shape[0], shape[1]])

        for idx, (x, t_m) in enumerate(raw_dataloader):
            # # take first 5 example
            # # x, t_m = x[:5], t_m[:5]
            # # x[~t_m] = 0
            # # print(x, t_m)
            
            # # 1M 2021/2/10 9
            # # 2M 2021/2/16 6+9
            # # 3M 2021/2/19 9+9
            # # 4M 2021/2/24 14+
            # # 5M 2021/3/3 20+9
            # x = torch.zeros_like(x)[:1]
            # # x[0, 0, 0] = 0.03
            # # x[0, 9, 0] = 0.16
            # # x[0, 15, 0] = 0.25
            # # x[0, 18, 0] = 0.22
            # # x[0, 24, 0] = 0.21
            # # x[0, 33, 0] = 0.16
            # x[0, 0, 0] = 0.04
            # x[0, 2, 0] = 0.58
            # x[0, 6, 0] = 0.27
            # x[0, 58, 0] = 1.
            # x[0, -1, 0] = 0.05
            # # x[0, 0, 0] = 0.01
            
            # # x[0, -1, 0] = 0.01
            # # x[0, -20, 0] = 0.01
            # # x[0, -100, 0] = 0.01
            # # x[0, -50, 0] = 0.01
            # # x[0, -120, 0] = 0.01
            # # import math
            # # for i in range(35, 240, 2):
            # #     x[0, i, 0] = max(0.01, math.exp(-0.01*i) / 10)

            # # import matplotlib.pyplot as plt
            # # plt.plot(x[0, :, 0].detach().cpu().numpy())
            # # plt.show()
            t_m = x == 0 # x != 0, 1 for observed, 0 for missing, bool tensor
            # # 
            if test:
                t_m = t_m.type_as(x)
                binary_mask = t_m.clone()
                binary_mask[binary_mask > 0] = 1
            else:
                binary_mask = t_m

            # x = x * 2 - 1
            x, t_m = x.to(self.device), t_m.to(self.device)
            binary_mask = binary_mask.to(self.device)

            if sampling_steps == self.model.num_timesteps:
                print("normal sampling")
                sample = self.ema.ema_model.sample_infill(
                    shape=x.shape,
                    target=x * t_m,
                    partial_mask=t_m,
                    model_kwargs=model_kwargs,
                )
                # x: partially noise : (batch_size, seq_length, feature_dim)
            else:
                print("fast sampling")
                if test:
                    sample = self.ema.ema_model.fast_sample_infill_float_mask(
                        shape=x.shape,
                        target=x * binary_mask, # x * t_m, 1 for observed, 0 for missing
                        partial_mask=t_m,
                        model_kwargs=model_kwargs,
                        sampling_timesteps=sampling_steps,
                    )
                else:
                    sample = self.ema.ema_model.fast_sample_infill(
                        shape=x.shape,
                        target=x * t_m,
                        partial_mask=t_m,
                        model_kwargs=model_kwargs,
                        sampling_timesteps=sampling_steps,
                    )

            samples = np.row_stack([samples, sample.detach().cpu().numpy()])
            reals = np.row_stack([reals, x.detach().cpu().numpy()])
            masks = np.row_stack([masks, t_m.detach().cpu().numpy()])
            break

        if self.logger is not None:
            self.logger.log_info(
                "Imputation done, time: {:.2f}".format(time.time() - tic)
            )
        return samples, reals, masks
        # return samples