experiment.py

import copy
import json
import os
import re,sys

import numpy as np
import pandas as pd
import pytorch_lightning as pl
import torch
from numpy.lib.function_base import flip
from pytorch_lightning import loggers as pl_loggers
from pytorch_lightning.callbacks import *
from torch import nn
from torch.cuda import amp
from torch.distributions import Categorical
from torch.optim.optimizer import Optimizer
from torch.utils.data.dataset import ConcatDataset, TensorDataset
from torchvision.utils import make_grid, save_image

from config import *
from dataset import *
from dist_utils import *
from renderer import *


class Rendering_Model(pl.LightningModule):
    def __init__(self, conf: TrainConfig):
        super().__init__()
        assert conf.train_mode != TrainMode.manipulate
        if conf.seed is not None:
            pl.seed_everything(conf.seed)

        self.save_hyperparameters(conf.as_dict_jsonable())

        self.conf = conf

        self.model = conf.make_model_conf().make_model(conf.N_frames)
        self.ema_model = copy.deepcopy(self.model)
        self.ema_model.requires_grad_(False)
        self.ema_model.eval()

        model_size = 0
        for param in self.model.parameters():
            model_size += param.data.nelement()
        print('Model params: %.2f M' % (model_size / 1024 / 1024))

        self.sampler = conf.make_diffusion_conf().make_sampler()
        self.eval_sampler = conf.make_eval_diffusion_conf().make_sampler()

        # this is shared for both model and latent
        self.T_sampler = conf.make_T_sampler()

        if conf.train_mode.use_latent_net():
            self.latent_sampler = conf.make_latent_diffusion_conf(
            ).make_sampler()
            self.eval_latent_sampler = conf.make_latent_eval_diffusion_conf(
            ).make_sampler()
        else:
            self.latent_sampler = None
            self.eval_latent_sampler = None

        # initial variables for consistent sampling
        self.register_buffer(
            'x_T',
            torch.randn(conf.sample_size, 3, conf.img_size, conf.img_size))

        if conf.pretrain is not None:
            selfState = self.state_dict()

            print(f'loading pretrain ... {conf.pretrain}')
            state = torch.load(conf.pretrain, map_location='cpu')
            print('step:', state['global_step'])
            state['state_dict']['x_T'] = torch.randn(conf.sample_size, 3, conf.img_size, conf.img_size)
            state['state_dict']['sample_size']= conf.sample_size

            state = state['state_dict']

            for name, param in state.items():
                origName = name;
                if name not in selfState:
                    name = name.replace("module.", "")
                    if name not in selfState:
                        print("%s is not in the model."%origName)
                        continue
                if selfState[name].size() != state[origName].size():
                    print("Wrong parameter length: %s, model: %s, loaded: %s"%(origName, selfState[name].size(), state[origName].size()))
                    continue
                selfState[name].copy_(param)

        self.conds_mean = None
        self.conds_std = None

    def normalize(self, cond):
        cond = (cond - self.conds_mean.to(self.device)) / self.conds_std.to(
            self.device)
        return cond
    
    def normalize(self, cond, device):
        cond = (cond - self.conds_mean.to(device)) / self.conds_std.to(
            device)
        return cond

    def denormalize(self, cond):
        cond = (cond * self.conds_std.to(self.device)) + self.conds_mean.to(
            self.device)
        return cond

    def sample(self, N, device, T=None, T_latent=None):
        if T is None:
            sampler = self.eval_sampler
            latent_sampler = self.latent_sampler
        else:
            sampler = self.conf._make_diffusion_conf(T).make_sampler()
            latent_sampler = self.conf._make_latent_diffusion_conf(T_latent).make_sampler()

        noise = torch.randn(N,
                            3,
                            self.conf.img_size,
                            self.conf.img_size,
                            device=device)
        pred_img = render_uncondition(
            self.conf,
            self.ema_model,
            noise,
            sampler=sampler,
            latent_sampler=latent_sampler,
            conds_mean=self.conds_mean,
            conds_std=self.conds_std,
        )
        pred_img = (pred_img + 1) / 2
        return pred_img

    def render(self, noise, x_start, x_ref, cond=None, T=None, mask=None):
        if T is None:
            sampler = self.eval_sampler
        else:
            sampler = self.conf._make_diffusion_conf(T).make_sampler()
        # print(cond.shape)
        pred_img = render_condition(self.conf,
                                        self.ema_model,
                                        noise,
                                        sampler=sampler,
                                        x_start=x_start,
                                        x_ref=x_ref,
                                        mask=mask,
                                        cond=cond)
        pred_img = (pred_img + 1) / 2
        return pred_img

    def encode(self, x):
        assert self.conf.model_type.has_autoenc()
        cond = self.ema_model.encoder.forward(x)
        return cond

    def forward(self, noise=None, x_start=None, ema_model: bool = False):
        with amp.autocast(False):
            if ema_model:
                model = self.ema_model
            else:
                model = self.model
            gen = self.eval_sampler.sample(model=model,
                                           noise=noise,
                                           x_start=x_start)
            return gen

    def setup(self, stage=None) -> None:
        """
        make datasets & seeding each worker separately
        """
        ##############################################
        # NEED TO SET THE SEED SEPARATELY HERE
        if self.conf.seed is not None:
            seed = self.conf.seed * get_world_size() + self.global_rank
            np.random.seed(seed)
            torch.manual_seed(seed)
            torch.cuda.manual_seed(seed)
            print('local seed:', seed)
        ##############################################

        self.train_data = self.conf.make_dataset()
        print('train data:', len(self.train_data))
        self.val_data = self.train_data
        print('val data:', len(self.val_data))

    def _train_dataloader(self, drop_last=True):
        """
        really make the dataloader
        """
        # make sure to use the fraction of batch size
        # the batch size is global!
        conf = self.conf.clone()
        conf.batch_size = self.batch_size

        dataloader = conf.make_loader(self.train_data,
                                      shuffle=True,
                                      drop_last=drop_last)
        return dataloader

    def train_dataloader(self):
        """
        return the dataloader, if diffusion mode => return image dataset
        if latent mode => return the inferred latent dataset
        """
        print('on train dataloader start ...')
        if self.conf.train_mode.require_dataset_infer():
            if self.conds is None:
                # usually we load self.conds from a file
                # so we do not need to do this again!
                self.conds = self.infer_whole_dataset()
                # need to use float32! unless the mean & std will be off!
                # (1, c)
                self.conds_mean.data = self.conds.float().mean(dim=0,
                                                               keepdim=True)
                self.conds_std.data = self.conds.float().std(dim=0,
                                                             keepdim=True)
            print('mean:', self.conds_mean.mean(), 'std:',
                  self.conds_std.mean())

            # return the dataset with pre-calculated conds
            conf = self.conf.clone()
            conf.batch_size = self.batch_size
            data = TensorDataset(self.conds)
            return conf.make_loader(data, shuffle=True)
        else:
            return self._train_dataloader()

    @property
    def batch_size(self):
        """
        local batch size for each worker
        """
        ws = get_world_size()
        assert self.conf.batch_size % ws == 0
        return self.conf.batch_size // ws

    @property
    def num_samples(self):
        """
        (global) batch size * iterations
        """
        # batch size here is global!
        # global_step already takes into account the accum batches
        return self.global_step * self.conf.batch_size_effective

    def is_last_accum(self, batch_idx):
        """
        is it the last gradient accumulation loop? 
        used with gradient_accum > 1 and to see if the optimizer will perform "step" in this iteration or not
        """
        return (batch_idx + 1) % self.conf.accum_batches == 0

    def training_step(self, batch, batch_idx):
        """
        given an input, calculate the loss function
        no optimization at this stage.
        """
        with amp.autocast(False):
            # batch size here is local!
            # forward
            if self.conf.train_mode.require_dataset_infer():
                # this mode as pre-calculated cond
                cond = batch[0]
                if self.conf.latent_znormalize:
                    cond = (cond - self.conds_mean.to(
                        self.device)) / self.conds_std.to(self.device)
            else:
                if 'ref' in batch.keys():
                    imgs, idxs, condtion_tensor, mouth_masked, ref_imgs = batch['img'], batch['index'], batch['condtion'], batch['mouth_masked'], batch['ref']
                else:
                    ref_imgs = None
                    imgs, idxs, condtion_tensor, mouth_masked = batch['img'], batch['index'], batch['condtion'], batch['mouth_masked']
                x_start = imgs
            if self.conf.train_mode == TrainMode.diffusion:
                """
                main training mode!!!
                """
                # with numpy seed we have the problem that the sample t's are related!
                t, weight = self.T_sampler.sample(len(x_start), x_start.device)
                losses = self.sampler.training_losses(model=self.model,
                                                      x_start=x_start,
                                                      ref_imgs=ref_imgs,
                                                      x_aug_cond=condtion_tensor,
                                                      mask=mouth_masked,
                                                      t=t)
            elif self.conf.train_mode.is_latent_diffusion():
                """
                training the latent variables!
                """
                # diffusion on the latent
                t, weight = self.T_sampler.sample(len(cond), cond.device)
                latent_losses = self.latent_sampler.training_losses(
                    model=self.model.latent_net, x_start=cond, t=t)
                # train only do the latent diffusion
                losses = {
                    'latent': latent_losses['loss'],
                    'loss': latent_losses['loss']
                }
            else:
                raise NotImplementedError()

            loss = losses['loss'].mean()
            # divide by accum batches to make the accumulated gradient exact!
            for key in ['loss', 'vae', 'latent', 'mmd', 'chamfer', 'arg_cnt']:
                if key in losses:
                    losses[key] = self.all_gather(losses[key]).mean()

            if self.global_rank == 0:
                self.logger.experiment.add_scalar('loss', losses['loss'],
                                                  self.num_samples)
                for key in ['vae', 'latent', 'mmd', 'chamfer', 'arg_cnt']:
                    if key in losses:
                        self.logger.experiment.add_scalar(
                            f'loss/{key}', losses[key], self.num_samples)

        return {'loss': loss}

    def on_train_batch_end(self, outputs, batch, batch_idx: int,
                           dataloader_idx: int) -> None:
        """
        after each training step ...
        """
        if self.is_last_accum(batch_idx):
            # only apply ema on the last gradient accumulation step,
            # if it is the iteration that has optimizer.step()
            if self.conf.train_mode == TrainMode.latent_diffusion:
                # it trains only the latent hence change only the latent
                ema(self.model.latent_net, self.ema_model.latent_net,
                    self.conf.ema_decay)
            else:
                ema(self.model, self.ema_model, self.conf.ema_decay)

    def on_before_optimizer_step(self, optimizer: Optimizer,
                                 optimizer_idx: int) -> None:
        # fix the fp16 + clip grad norm problem with pytorch lightinng
        # this is the currently correct way to do it
        if self.conf.grad_clip > 0:
            # from trainer.params_grads import grads_norm, iter_opt_params
            params = [
                p for group in optimizer.param_groups for p in group['params']
            ]
            # print('before:', grads_norm(iter_opt_params(optimizer)))
            torch.nn.utils.clip_grad_norm_(params,
                                           max_norm=self.conf.grad_clip)
            # print('after:', grads_norm(iter_opt_params(optimizer)))
            
    def configure_optimizers(self):
        out = {}
        if self.conf.optimizer == OptimizerType.adam:
            optim = torch.optim.Adam(self.model.parameters(),
                                     lr=self.conf.lr,
                                     weight_decay=self.conf.weight_decay)
        elif self.conf.optimizer == OptimizerType.adamw:
            optim = torch.optim.AdamW(self.model.parameters(),
                                      lr=self.conf.lr,
                                      weight_decay=self.conf.weight_decay)
        else:
            raise NotImplementedError()
        out['optimizer'] = optim
        if self.conf.warmup > 0:
            sched = torch.optim.lr_scheduler.LambdaLR(optim,
                                                      lr_lambda=WarmupLR(
                                                          self.conf.warmup))
            out['lr_scheduler'] = {
                'scheduler': sched,
                'interval': 'step',
            }
        return out

    def split_tensor(self, x):
        """
        extract the tensor for a corresponding "worker" in the batch dimension

        Args:
            x: (n, c)

        Returns: x: (n_local, c)
        """
        n = len(x)
        rank = self.global_rank
        world_size = get_world_size()
        # print(f'rank: {rank}/{world_size}')
        per_rank = n // world_size
        return x[rank * per_rank:(rank + 1) * per_rank]

def ema(source, target, decay):
    source_dict = source.state_dict()
    target_dict = target.state_dict()
    for key in source_dict.keys():
        target_dict[key].data.copy_(target_dict[key].data * decay +
                                    source_dict[key].data * (1 - decay))

class WarmupLR:
    def __init__(self, warmup) -> None:
        self.warmup = warmup

    def __call__(self, step):
        return min(step, self.warmup) / self.warmup


def is_time(num_samples, every, step_size):
    closest = (num_samples // every) * every
    return num_samples - closest < step_size


def train(conf: TrainConfig, gpus, nodes=1, mode: str = 'train'):
    print('conf:', conf.name)
    print('gpus:', gpus)
    # assert not (conf.fp16 and conf.grad_clip > 0
    #             ), 'pytorch lightning has bug with amp + gradient clipping'
    model = Rendering_Model(conf)

    if not os.path.exists(conf.logdir):
        os.makedirs(conf.logdir)
    checkpoint = ModelCheckpoint(dirpath=f'{conf.logdir}',
                                 save_last=True,
                                 save_top_k=-1,
                                 every_n_train_steps=conf.save_every_samples //
                                 conf.batch_size_effective)
    checkpoint_path = f'{conf.logdir}/last.ckpt'
    print('ckpt path:', checkpoint_path)
    if os.path.exists(checkpoint_path):
        resume = checkpoint_path
        print('resume!')
    else:
        if conf.continue_from is not None:
            # continue from a checkpoint
            resume = conf.continue_from.path
        else:
            resume = None

    tb_logger = pl_loggers.TensorBoardLogger(save_dir=conf.logdir,
                                             name=None,
                                             version='')

    # from pytorch_lightning.

    plugins = []
    if len(gpus) == 1 and nodes == 1:
        accelerator = None
    else:
        accelerator = 'dp'
        print(accelerator)
        from pytorch_lightning.plugins import DDPPlugin

        # important for working with gradient checkpoint
        plugins.append(DDPPlugin(find_unused_parameters=False))

    trainer = pl.Trainer(
        max_steps=conf.total_samples // conf.batch_size_effective,
        resume_from_checkpoint=resume,
        gpus=gpus,
        num_nodes=nodes,
        accelerator=accelerator,
        precision=16 if conf.fp16 else 32,
        callbacks=[
            checkpoint,
            LearningRateMonitor(),
        ],
        # clip in the model instead
        # gradient_clip_val=conf.grad_clip,
        replace_sampler_ddp=True,
        logger=tb_logger,
        accumulate_grad_batches=conf.accum_batches,
        plugins=plugins,
    )

    if mode == 'train':
        trainer.fit(model)
    elif mode == 'eval':
        # load the latest checkpoint
        # perform lpips
        # dummy loader to allow calling "test_step"
        dummy = DataLoader(TensorDataset(torch.tensor([0.] * conf.batch_size)),
                           batch_size=conf.batch_size)
        eval_path = conf.eval_path or checkpoint_path
        # conf.eval_num_images = 50
        print('loading from:', eval_path)
        state = torch.load(eval_path, map_location='cpu')
        print('step:', state['global_step'])
        model.load_state_dict(state['state_dict'])
        # trainer.fit(model)
        out = trainer.test(model, dataloaders=dummy)
        # first (and only) loader
        out = out[0]
        print(out)

        if get_rank() == 0:
            # save to tensorboard
            for k, v in out.items():
                tb_logger.experiment.add_scalar(
                    k, v, state['global_step'] * conf.batch_size_effective)

            # # save to file
            # # make it a dict of list
            # for k, v in out.items():
            #     out[k] = [v]
            tgt = f'evals/{conf.name}.txt'
            dirname = os.path.dirname(tgt)
            if not os.path.exists(dirname):
                os.makedirs(dirname)
            with open(tgt, 'a') as f:
                f.write(json.dumps(out) + "\n")
            # pd.DataFrame(out).to_csv(tgt)
    elif mode == 'load':
        # load the latest checkpoint
        # perform lpips
        # dummy loader to allow calling "test_step"
        dummy = DataLoader(TensorDataset(torch.tensor([0.] * conf.batch_size)),
                           batch_size=conf.batch_size)
        eval_path = conf.eval_path or checkpoint_path
        # conf.eval_num_images = 50
        print('loading from:', eval_path)
        state = torch.load(eval_path, map_location='cpu')
        print('step:', state['global_step'])
        model.load_state_dict(state['state_dict'])
        return trainer, model, dummy
    else:
        raise NotImplementedError()