nsr/train_util.py

import copy
import functools
import json
import os
from pathlib import Path
from pdb import set_trace as st

import matplotlib.pyplot as plt
import traceback
import blobfile as bf
import imageio
import numpy as np
# from sympy import O
import torch as th
import torch.distributed as dist
import torchvision
from PIL import Image
from torch.nn.parallel.distributed import DistributedDataParallel as DDP
from torch.optim import AdamW
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm

from guided_diffusion import dist_util, logger
from guided_diffusion.fp16_util import MixedPrecisionTrainer
from guided_diffusion.nn import update_ema
from guided_diffusion.resample import LossAwareSampler, UniformSampler
from guided_diffusion.train_util import (calc_average_loss,
                                         find_ema_checkpoint,
                                         find_resume_checkpoint,
                                         get_blob_logdir, log_rec3d_loss_dict,
                                         parse_resume_step_from_filename)

from .camera_utils import LookAtPoseSampler, FOV_to_intrinsics

# from ..guided_diffusion.train_util import TrainLoop


def flip_yaw(pose_matrix):
    flipped = pose_matrix.clone()
    flipped[:, 0, 1] *= -1
    flipped[:, 0, 2] *= -1
    flipped[:, 1, 0] *= -1
    flipped[:, 2, 0] *= -1
    flipped[:, 0, 3] *= -1
    # st()
    return flipped


# basic reconstruction model
class TrainLoopBasic:

    def __init__(
            self,
            *,
            rec_model,
            loss_class,
            # diffusion,
            data,
            eval_data,
            batch_size,
            microbatch,
            lr,
            ema_rate,
            log_interval,
            eval_interval,
            save_interval,
            resume_checkpoint,
            use_fp16=False,
            fp16_scale_growth=1e-3,
            # schedule_sampler=None,
            weight_decay=0.0,
            lr_anneal_steps=0,
            iterations=10001,
            load_submodule_name='',
            ignore_resume_opt=False,
            model_name='rec',
            use_amp=False,
            compile=False,
            **kwargs):
        self.pool_512 = th.nn.AdaptiveAvgPool2d((512, 512))
        self.pool_256 = th.nn.AdaptiveAvgPool2d((256, 256))
        self.pool_128 = th.nn.AdaptiveAvgPool2d((128, 128))
        self.pool_64 = th.nn.AdaptiveAvgPool2d((64, 64))
        self.rec_model = rec_model
        self.loss_class = loss_class
        # self.diffusion = diffusion
        # self.schedule_sampler = schedule_sampler or UniformSampler(diffusion)
        self.data = data
        self.eval_data = eval_data
        self.batch_size = batch_size
        self.microbatch = microbatch if microbatch > 0 else batch_size
        self.lr = lr
        self.ema_rate = ([ema_rate] if isinstance(ema_rate, float) else
                         [float(x) for x in ema_rate.split(",")])
        self.log_interval = log_interval
        self.eval_interval = eval_interval
        self.save_interval = save_interval
        self.iterations = iterations
        self.resume_checkpoint = resume_checkpoint
        self.use_fp16 = use_fp16
        self.fp16_scale_growth = fp16_scale_growth
        self.weight_decay = weight_decay
        self.lr_anneal_steps = lr_anneal_steps

        self.step = 0
        self.resume_step = 0
        # self.global_batch = self.batch_size * dist.get_world_size()
        self.global_batch = self.batch_size * dist_util.get_world_size()

        self.sync_cuda = th.cuda.is_available()

        # self._load_and_sync_parameters(load_submodule_name)
        self._load_and_sync_parameters()

        self.mp_trainer_rec = MixedPrecisionTrainer(
            model=self.rec_model,
            use_fp16=self.use_fp16,
            fp16_scale_growth=fp16_scale_growth,
            model_name=model_name,
            use_amp=use_amp)
        self.writer = SummaryWriter(log_dir=f'{logger.get_dir()}/runs')

        self.opt = AdamW(self._init_optim_groups(kwargs))

        if dist_util.get_rank() == 0:
            logger.log(self.opt)

        if self.resume_step:
            if not ignore_resume_opt:
                self._load_optimizer_state()
            else:
                logger.warn("Ignoring optimizer state from checkpoint.")
            # Model was resumed, either due to a restart or a checkpoint
            # being specified at the command line.
            # self.ema_params = [
            #     self._load_ema_parameters(rate, load_submodule_name) for rate in self.ema_rate
            # ]

            self.ema_params = [
                self._load_ema_parameters(
                    rate,
                    self.rec_model,
                    self.mp_trainer_rec,
                    model_name=self.mp_trainer_rec.model_name)
                for rate in self.ema_rate
            ]
        else:
            self.ema_params = [
                copy.deepcopy(self.mp_trainer_rec.master_params)
                for _ in range(len(self.ema_rate))
            ]

        # compile
        if compile:
            logger.log('compiling... ignore vit_decoder')
            # self.rec_model.encoder  = th.compile(self.rec_model.encoder)
            self.rec_model.decoder.decoder_pred = th.compile(
                self.rec_model.decoder.decoder_pred)
            # self.rec_model.decoder.triplane_decoder  = th.compile(self.rec_model.decoder.triplane_decoder)
            for module_k, sub_module in self.rec_model.decoder.superresolution.items(
            ):
                self.rec_model.decoder.superresolution[module_k] = th.compile(
                    sub_module)

        if th.cuda.is_available():
            self.use_ddp = True

            self.rec_model = th.nn.SyncBatchNorm.convert_sync_batchnorm(
                self.rec_model)

            self.rec_model = DDP(
                self.rec_model,
                device_ids=[dist_util.dev()],
                output_device=dist_util.dev(),
                broadcast_buffers=False,
                bucket_cap_mb=128,
                find_unused_parameters=False,
            )
        else:
            if dist_util.get_world_size() > 1:
                logger.warn("Distributed training requires CUDA. "
                            "Gradients will not be synchronized properly!")
            self.use_ddp = False
            self.rec_model = self.rec_model

        self.novel_view_poses = None
        th.cuda.empty_cache()

    def _init_optim_groups(self, kwargs):
        raise NotImplementedError('')

    def _load_and_sync_parameters(self, submodule_name=''):
        # resume_checkpoint, self.resume_step = find_resume_checkpoint() or self.resume_checkpoint
        resume_checkpoint = self.resume_checkpoint  # * default behaviour
        # logger.log('resume_checkpoint', resume_checkpoint, self.resume_checkpoint)

        if resume_checkpoint:
            self.resume_step = parse_resume_step_from_filename(
                resume_checkpoint)
            if dist_util.get_rank() == 0:
                logger.log(
                    f"loading model from checkpoint: {resume_checkpoint}...")
                map_location = {
                    'cuda:%d' % 0: 'cuda:%d' % dist_util.get_rank()
                }  # configure map_location properly

                resume_state_dict = dist_util.load_state_dict(
                    resume_checkpoint, map_location=map_location)
                if submodule_name != '':
                    model_state_dict = getattr(self.rec_model,
                                               submodule_name).state_dict()
                    if dist_util.get_rank() == 0:
                        logger.log('loading submodule: ', submodule_name)
                else:
                    model_state_dict = self.rec_model.state_dict()

                model = self.rec_model

                # for k, v in resume_state_dict.items():
                #     if k in model_state_dict.keys() and v.size(
                #     ) == model_state_dict[k].size():
                #         model_state_dict[k] = v
                #     else:
                #         logger.log('!!!! ignore key: ', k, ": ", v.size())

                for k, v in resume_state_dict.items():
                    if '._orig_mod' in k:  # prefix in torch.compile
                        k = k.replace('._orig_mod', '')
                    if k in model_state_dict.keys():
                        if v.size() == model_state_dict[k].size():
                            model_state_dict[k] = v
                            # model_state_dict[k].copy_(v)
                        else:
                            # if v.ndim > 1:
                            #     model_state_dict[k][:v.shape[0], :v.shape[1], ...] = v # load the decoder
                            #     model_state_dict[k][v.shape[0]:, v.shape[1]:, ...] = 0
                            # else:
                            #     model_state_dict[k][:v.shape[0], ...] = v # load the decoder
                            #     model_state_dict[k][v.shape[0]:, ...] = 0
                            # logger.log('!!!! size mismatch, partially load: ', k, ": ", v.size(), "state_dict: ", model_state_dict[k].size())
                            logger.log('!!!! size mismatch, ignore: ', k, ": ",
                                       v.size(), "state_dict: ",
                                       model_state_dict[k].size())

                    elif 'decoder.vit_decoder.blocks' in k:
                        # st()
                        # load from 2D ViT pre-trained into 3D ViT blocks.
                        assert len(model.decoder.vit_decoder.blocks[0].vit_blks
                                   ) == 2  # assert depth=2 here.
                        fusion_ca_depth = len(
                            model.decoder.vit_decoder.blocks[0].vit_blks)
                        vit_subblk_index = int(k.split('.')[3])
                        vit_blk_keyname = ('.').join(k.split('.')[4:])
                        fusion_blk_index = vit_subblk_index // fusion_ca_depth
                        fusion_blk_subindex = vit_subblk_index % fusion_ca_depth
                        model_state_dict[
                            f'decoder.vit_decoder.blocks.{fusion_blk_index}.vit_blks.{fusion_blk_subindex}.{vit_blk_keyname}'] = v
                        logger.log('load 2D ViT weight: {}'.format(
                            f'decoder.vit_decoder.blocks.{fusion_blk_index}.vit_blks.{fusion_blk_subindex}.{vit_blk_keyname}'
                        ))

                    else:
                        logger.log(
                            '!!!! ignore key, not in the model_state_dict: ',
                            k, ": ", v.size())

                logger.log('model loading finished')

                if submodule_name != '':
                    getattr(self.rec_model,
                            submodule_name).load_state_dict(model_state_dict,
                                                            strict=True)
                else:
                    self.rec_model.load_state_dict(model_state_dict,
                                                   strict=False)
                    #    strict=True)

        if dist_util.get_world_size() > 1:
            # dist_util.sync_params(self.model.named_parameters())
            dist_util.sync_params(self.rec_model.parameters())
            logger.log('synced params')

    def _load_ema_parameters(self,
                             rate,
                             model=None,
                             mp_trainer=None,
                             model_name='ddpm'):

        if mp_trainer is None:
            mp_trainer = self.mp_trainer_rec
        if model is None:
            model = self.rec_model

        ema_params = copy.deepcopy(mp_trainer.master_params)

        # main_checkpoint, _ = find_resume_checkpoint(
        #     self.resume_checkpoint, model_name) or self.resume_checkpoint

        main_checkpoint = self.resume_checkpoint
        ema_checkpoint = find_ema_checkpoint(main_checkpoint, self.resume_step,
                                             rate, model_name)
        if ema_checkpoint and model_name == 'ddpm':

            if dist_util.get_rank() == 0:

                if not Path(ema_checkpoint).exists():
                    logger.log(
                        f"failed to load EMA from checkpoint: {ema_checkpoint}, not exist"
                    )
                    return

                logger.log(f"loading EMA from checkpoint: {ema_checkpoint}...")

                map_location = {
                    'cuda:%d' % 0: 'cuda:%d' % dist_util.get_rank()
                }  # configure map_location properly

                state_dict = dist_util.load_state_dict(
                    ema_checkpoint, map_location=map_location)

                model_ema_state_dict = model.state_dict()

                for k, v in state_dict.items():
                    if k in model_ema_state_dict.keys() and v.size(
                    ) == model_ema_state_dict[k].size():
                        model_ema_state_dict[k] = v

                    elif 'IN' in k and getattr(model, 'decomposed_IN', False):
                        model_ema_state_dict[k.replace(
                            'IN', 'IN.IN')] = v  # decomposed IN

                    else:
                        logger.log('ignore key: ', k, ": ", v.size())

                ema_params = mp_trainer.state_dict_to_master_params(
                    model_ema_state_dict)

                del state_dict

        # logger.log('ema mark 3, ', model_name, )

        # ! debugging, remove to check which key fails.
        if dist_util.get_world_size() > 1:
            dist_util.sync_params(ema_params)

        # logger.log('ema mark 4, ', model_name, )
        # del ema_params
        return ema_params

    def _load_optimizer_state(self):
        main_checkpoint, _ = find_resume_checkpoint() or self.resume_checkpoint
        opt_checkpoint = bf.join(bf.dirname(main_checkpoint),
                                 f"opt{self.resume_step:06}.pt")
        if bf.exists(opt_checkpoint):
            logger.log(
                f"loading optimizer state from checkpoint: {opt_checkpoint}")

            map_location = {
                'cuda:%d' % 0: 'cuda:%d' % dist_util.get_rank()
            }  # configure map_location properly

            state_dict = dist_util.load_state_dict(opt_checkpoint,
                                                   map_location=map_location)
            self.opt.load_state_dict(state_dict)
            # self.opt.load_state_dict({k: v for k, v in state_dict.items() if k in self.opt.state_dict()})

            del state_dict

    def run_loop(self, batch=None):
        while (not self.lr_anneal_steps
               or self.step + self.resume_step < self.lr_anneal_steps):

            # let all processes sync up before starting with a new epoch of training
            dist_util.synchronize()

            # batch, cond = next(self.data)
            # if batch is None:
            if isinstance(self.data, list):
                if self.step <= self.data[2]:
                    batch = next(self.data[1])
                else:
                    batch = next(self.data[0])
            else:
                batch = next(self.data)

            # batch = next(self.data)
            if self.novel_view_poses is None:
                self.novel_view_poses = th.roll(batch['c'], 1, 0).to(
                    dist_util.dev())  # save for eval visualization use

            self.run_step(batch)

            if self.step % 1000 == 0:
                dist_util.synchronize()
                th.cuda.empty_cache()  # avoid memory leak

            if self.step % self.log_interval == 0 and dist_util.get_rank(
            ) == 0:
                out = logger.dumpkvs()
                # * log to tensorboard
                for k, v in out.items():
                    self.writer.add_scalar(f'Loss/{k}', v,
                                           self.step + self.resume_step)

            if self.step % self.eval_interval == 0 and self.step != 0:
                # if self.step % self.eval_interval == 0 and (self.step +
                #                                             self.resume_step) != 0:
                # if self.step % self.eval_interval == 0:  # ! for debugging
                # if self.step % self.eval_interval == 0:
                if dist_util.get_rank() == 0:
                    try:
                        self.eval_loop()
                    except Exception as e:
                        logger.log(e)
                    # self.eval_novelview_loop()
                    # let all processes sync up before starting with a new epoch of training
                dist_util.synchronize()

            if self.step % self.save_interval == 0 and self.step != 0:
                self.save()
                dist_util.synchronize()
                # Run for a finite amount of time in integration tests.
                if os.environ.get("DIFFUSION_TRAINING_TEST",
                                  "") and self.step > 0:
                    return

            self.step += 1

            if self.step > self.iterations:
                logger.log('reached maximum iterations, exiting')

                # Save the last checkpoint if it wasn't already saved.
                if (self.step -
                        1) % self.save_interval != 0 and self.step != 1:
                    self.save()

                exit()

        # Save the last checkpoint if it wasn't already saved.
        if (self.step - 1) % self.save_interval != 0 and self.step != 1:
            self.save()

    @th.no_grad()
    def eval_loop(self):
        raise NotImplementedError('')

    def run_step(self, batch, *args):
        self.forward_backward(batch)
        took_step = self.mp_trainer_rec.optimize(self.opt)
        if took_step:
            self._update_ema()
        self._anneal_lr()
        self.log_step()

    def forward_backward(self, batch, *args, **kwargs):
        # th.cuda.empty_cache()
        raise NotImplementedError('')

    def _update_ema(self):
        for rate, params in zip(self.ema_rate, self.ema_params):
            update_ema(params, self.mp_trainer_rec.master_params, rate=rate)

    def _anneal_lr(self):
        if not self.lr_anneal_steps:
            return
        frac_done = (self.step + self.resume_step) / self.lr_anneal_steps
        lr = self.lr * (1 - frac_done)
        for param_group in self.opt.param_groups:
            param_group["lr"] = lr

    def log_step(self):
        logger.logkv("step", self.step + self.resume_step)
        logger.logkv("samples",
                     (self.step + self.resume_step + 1) * self.global_batch)

    def save(self):

        def save_checkpoint(rate, params):
            state_dict = self.mp_trainer_rec.master_params_to_state_dict(
                params)
            if dist_util.get_rank() == 0:
                logger.log(f"saving model {rate}...")
                if not rate:
                    filename = f"model_rec{(self.step+self.resume_step):07d}.pt"
                else:
                    filename = f"ema_{rate}_{(self.step+self.resume_step):07d}.pt"
                with bf.BlobFile(bf.join(get_blob_logdir(), filename),
                                 "wb") as f:
                    th.save(state_dict, f)

        save_checkpoint(
            0, self.mp_trainer_rec.master_params)  # avoid OOM when saving ckpt
        for rate, params in zip(self.ema_rate, self.ema_params):
            save_checkpoint(rate, params)
        th.cuda.empty_cache()

        dist.barrier()


class TrainLoop3DRec(TrainLoopBasic):

    def __init__(
            self,
            *,
            rec_model,
            loss_class,
            # diffusion,
            data,
            eval_data,
            batch_size,
            microbatch,
            lr,
            ema_rate,
            log_interval,
            eval_interval,
            save_interval,
            resume_checkpoint,
            use_fp16=False,
            fp16_scale_growth=1e-3,
            # schedule_sampler=None,
            weight_decay=0.0,
            lr_anneal_steps=0,
            iterations=10001,
            load_submodule_name='',
            ignore_resume_opt=False,
            model_name='rec',
            use_amp=False,
            compile=False,
            **kwargs):
        super().__init__(rec_model=rec_model,
                         loss_class=loss_class,
                         data=data,
                         eval_data=eval_data,
                         batch_size=batch_size,
                         microbatch=microbatch,
                         lr=lr,
                         ema_rate=ema_rate,
                         log_interval=log_interval,
                         eval_interval=eval_interval,
                         save_interval=save_interval,
                         resume_checkpoint=resume_checkpoint,
                         use_fp16=use_fp16,
                         fp16_scale_growth=fp16_scale_growth,
                         weight_decay=weight_decay,
                         lr_anneal_steps=lr_anneal_steps,
                         iterations=iterations,
                         load_submodule_name=load_submodule_name,
                         ignore_resume_opt=ignore_resume_opt,
                         model_name=model_name,
                         use_amp=use_amp,
                         compile=compile,
                         **kwargs)

        # self.rec_model = self.ddp_model
        # self._prepare_nvs_pose() # for eval novelview visualization

        self.triplane_scaling_divider = 1.0
        self.latent_name = 'latent_normalized_2Ddiffusion'  # normalized triplane latent
        self.render_latent_behaviour = 'decode_after_vae'  # directly render using triplane operations

        th.cuda.empty_cache()

    @th.inference_mode()
    def render_video_given_triplane(self,
                                    planes,
                                    rec_model,
                                    name_prefix='0',
                                    save_img=False,
                                    render_reference=None,
                                    save_mesh=False):

        planes *= self.triplane_scaling_divider  # if setting clip_denoised=True, the sampled planes will lie in [-1,1]. Thus, values beyond [+- std] will be abandoned in this version. Move to IN for later experiments.

        # sr_w_code = getattr(self.ddp_rec_model.module.decoder, 'w_avg', None)
        # sr_w_code = None
        batch_size = planes.shape[0]

        # if sr_w_code is not None:
        #     sr_w_code = sr_w_code.reshape(1, 1,
        #                                   -1).repeat_interleave(batch_size, 0)

        # used during diffusion sampling inference
        # if not save_img:

        # ! mesh

        if planes.shape[1] == 16:  # ffhq/car
            ddpm_latent = {
                self.latent_name: planes[:, :12],
                'bg_plane': planes[:, 12:16],
            }
        else:
            ddpm_latent = {
                self.latent_name: planes,
            }

        ddpm_latent.update(
            rec_model(latent=ddpm_latent,
                      behaviour='decode_after_vae_no_render'))

        # if export_mesh:
        # if True:
        if save_mesh:
            # mesh_size = 512
            mesh_size = 256
            # mesh_size = 384
            # mesh_size = 320
            # mesh_thres = 3 # TODO, requires tuning
            # mesh_thres = 5 # TODO, requires tuning
            mesh_thres = 10  # TODO, requires tuning
            import mcubes
            import trimesh
            dump_path = f'{logger.get_dir()}/mesh/'

            os.makedirs(dump_path, exist_ok=True)

            grid_out = rec_model(
                latent=ddpm_latent,
                grid_size=mesh_size,
                behaviour='triplane_decode_grid',
            )

            vtx, faces = mcubes.marching_cubes(
                grid_out['sigma'].squeeze(0).squeeze(-1).cpu().numpy(),
                mesh_thres)
            vtx = vtx / (mesh_size - 1) * 2 - 1

            # vtx_tensor = th.tensor(vtx, dtype=th.float32, device=dist_util.dev()).unsqueeze(0)
            # vtx_colors = self.model.synthesizer.forward_points(planes, vtx_tensor)['rgb'].squeeze(0).cpu().numpy()  # (0, 1)
            # vtx_colors = (vtx_colors * 255).astype(np.uint8)

            # mesh = trimesh.Trimesh(vertices=vtx, faces=faces, vertex_colors=vtx_colors)
            mesh = trimesh.Trimesh(
                vertices=vtx,
                faces=faces,
            )

            mesh_dump_path = os.path.join(dump_path, f'{name_prefix}.ply')
            mesh.export(mesh_dump_path, 'ply')

            print(f"Mesh dumped to {dump_path}")
            del grid_out, mesh
            th.cuda.empty_cache()
            # return

        video_out = imageio.get_writer(
            f'{logger.get_dir()}/triplane_{name_prefix}.mp4',
            mode='I',
            fps=15,
            codec='libx264')

        if planes.shape[1] == 16:  # ffhq/car
            ddpm_latent = {
                self.latent_name: planes[:, :12],
                'bg_plane': planes[:, 12:16],
            }
        else:
            ddpm_latent = {
                self.latent_name: planes,
            }

        ddpm_latent.update(
            rec_model(latent=ddpm_latent,
                      behaviour='decode_after_vae_no_render'))

        # planes = planes.repeat_interleave(micro['c'].shape[0], 0)

        # for i in range(0, len(c_list), 1): # TODO, larger batch size for eval
        # micro_batchsize = 2
        # micro_batchsize = batch_size

        if render_reference is None:
            render_reference = self.eval_data  # compat
        else:  # use train_traj
            for key in ['ins', 'bbox', 'caption']:
                if key in render_reference:
                    render_reference.pop(key)
            # render_reference.pop('bbox')
            # render_reference.pop('caption')

            # compat lst for enumerate
            render_reference = [{
                k: v[idx:idx + 1]
                for k, v in render_reference.items()
            } for idx in range(40)]

        # for i, batch in enumerate(tqdm(self.eval_data)):
        for i, batch in enumerate(tqdm(render_reference)):
            micro = {
                k: v.to(dist_util.dev()) if isinstance(v, th.Tensor) else v
                for k, v in batch.items()
            }
            # micro = {'c': batch['c'].to(dist_util.dev()).repeat_interleave(batch_size, 0)}

            # all_pred = []
            pred = rec_model(
                img=None,
                c=micro['c'],
                latent=ddpm_latent,
                # latent={
                #     # k: v.repeat_interleave(micro['c'].shape[0], 0) if v is not None else None
                #     k: v.repeat_interleave(micro['c'].shape[0], 0) if v is not None else None
                #     for k, v in ddpm_latent.items()
                # },
                behaviour='triplane_dec')

            # if True:
            pred_depth = pred['image_depth']
            pred_depth = (pred_depth - pred_depth.min()) / (pred_depth.max() -
                                                            pred_depth.min())

            # save viridis_r depth
            pred_depth = pred_depth.cpu()[0].permute(1, 2, 0).numpy()
            pred_depth = (plt.cm.viridis(pred_depth[..., 0])[..., :3]) * 2 - 1
            pred_depth = th.from_numpy(pred_depth).to(
                pred['image_raw'].device).permute(2, 0, 1).unsqueeze(0)
            # st()
            # pred_depth =

            if 'image_sr' in pred:

                gen_img = pred['image_sr']

                if pred['image_sr'].shape[-1] == 512:

                    pred_vis = th.cat([
                        micro['img_sr'],
                        self.pool_512(pred['image_raw']), gen_img,
                        self.pool_512(pred_depth).repeat_interleave(3, dim=1)
                    ],
                                      dim=-1)

                elif pred['image_sr'].shape[-1] == 128:

                    pred_vis = th.cat([
                        micro['img_sr'],
                        self.pool_128(pred['image_raw']), pred['image_sr'],
                        self.pool_128(pred_depth).repeat_interleave(3, dim=1)
                    ],
                                      dim=-1)

            else:
                gen_img = pred['image_raw']

                pred_vis = th.cat(
                    [
                        # self.pool_128(micro['img']),
                        self.pool_128(gen_img),
                        # self.pool_128(pred_depth.repeat_interleave(3, dim=1))
                        self.pool_128(pred_depth)
                    ],
                    dim=-1)  # B, 3, H, W

            if save_img:
                for batch_idx in range(gen_img.shape[0]):
                    sampled_img = Image.fromarray(
                        (gen_img[batch_idx].permute(1, 2, 0).cpu().numpy() *
                         127.5 + 127.5).clip(0, 255).astype(np.uint8))
                    if sampled_img.size != (512, 512):
                        sampled_img = sampled_img.resize(
                            (128, 128), Image.HAMMING)  # for shapenet
                    sampled_img.save(logger.get_dir() +
                                     '/FID_Cals/{}_{}.png'.format(
                                         int(name_prefix) * batch_size +
                                         batch_idx, i))
                    # print('FID_Cals/{}_{}.png'.format(int(name_prefix)*batch_size+batch_idx, i))

            vis = pred_vis.permute(0, 2, 3, 1).cpu().numpy()
            vis = vis * 127.5 + 127.5
            vis = vis.clip(0, 255).astype(np.uint8)

            # if vis.shape[0] > 1:
            #     vis = np.concatenate(np.split(vis, vis.shape[0], axis=0),
            #                          axis=-3)

            # if not save_img:
            for j in range(vis.shape[0]
                           ):  # ! currently only export one plane at a time
                video_out.append_data(vis[j])

        # if not save_img:
        video_out.close()
        del video_out
        print('logged video to: ',
              f'{logger.get_dir()}/triplane_{name_prefix}.mp4')

        del vis, pred_vis, micro, pred,

    def _init_optim_groups(self, kwargs):
        if kwargs.get('decomposed', False):  # AE

            optim_groups = [
                # vit encoder
                {
                    'name': 'encoder',
                    'params': self.mp_trainer_rec.model.encoder.parameters(),
                    'lr': kwargs['encoder_lr'],
                    'weight_decay': kwargs['encoder_weight_decay']
                },

                # vit decoder backbone
                {
                    'name':
                    'decoder.vit_decoder',
                    'params':
                    self.mp_trainer_rec.model.decoder.vit_decoder.parameters(),
                    'lr':
                    kwargs['vit_decoder_lr'],
                    'weight_decay':
                    kwargs['vit_decoder_wd']
                },

                # triplane decoder, may include bg synthesis network
                {
                    'name':
                    'decoder.triplane_decoder',
                    'params':
                    self.mp_trainer_rec.model.decoder.triplane_decoder.
                    parameters(),
                    'lr':
                    kwargs['triplane_decoder_lr'],
                    # 'weight_decay': self.weight_decay
                },
            ]

            if self.mp_trainer_rec.model.decoder.superresolution is not None:
                optim_groups.append({
                    'name':
                    'decoder.superresolution',
                    'params':
                    self.mp_trainer_rec.model.decoder.superresolution.
                    parameters(),
                    'lr':
                    kwargs['super_resolution_lr'],
                })

            if self.mp_trainer_rec.model.dim_up_mlp is not None:
                optim_groups.append({
                    'name':
                    'dim_up_mlp',
                    'params':
                    self.mp_trainer_rec.model.dim_up_mlp.parameters(),
                    'lr':
                    kwargs['encoder_lr'],
                    # 'weight_decay':
                    # self.weight_decay
                })

            # add 3D aware operators
            if self.mp_trainer_rec.model.decoder.decoder_pred_3d is not None:
                optim_groups.append({
                    'name':
                    'decoder_pred_3d',
                    'params':
                    self.mp_trainer_rec.model.decoder.decoder_pred_3d.
                    parameters(),
                    'lr':
                    kwargs['vit_decoder_lr'],
                    'weight_decay':
                    kwargs['vit_decoder_wd']
                })

            if self.mp_trainer_rec.model.decoder.transformer_3D_blk is not None:
                optim_groups.append({
                    'name':
                    'decoder_transformer_3D_blk',
                    'params':
                    self.mp_trainer_rec.model.decoder.transformer_3D_blk.
                    parameters(),
                    'lr':
                    kwargs['vit_decoder_lr'],
                    'weight_decay':
                    kwargs['vit_decoder_wd']
                })

            if self.mp_trainer_rec.model.decoder.logvar is not None:
                optim_groups.append({
                    'name':
                    'decoder_logvar',
                    'params':
                    self.mp_trainer_rec.model.decoder.logvar,
                    'lr':
                    kwargs['vit_decoder_lr'],
                    'weight_decay':
                    kwargs['vit_decoder_wd']
                })

            if self.mp_trainer_rec.model.decoder.decoder_pred is not None:
                optim_groups.append(
                    # MLP triplane SR
                    {
                        'name':
                        'decoder.decoder_pred',
                        'params':
                        self.mp_trainer_rec.model.decoder.decoder_pred.
                        parameters(),
                        'lr':
                        kwargs['vit_decoder_lr'],
                        # 'weight_decay': 0
                        'weight_decay':
                        kwargs['vit_decoder_wd']
                    }, )

            if self.mp_trainer_rec.model.confnet is not None:
                optim_groups.append({
                    'name':
                    'confnet',
                    'params':
                    self.mp_trainer_rec.model.confnet.parameters(),
                    'lr':
                    1e-5,  # as in unsup3d
                })

            # self.opt = AdamW(optim_groups)

            if dist_util.get_rank() == 0:
                logger.log('using independent optimizer for each components')
        else:
            optim_groups = [
                dict(name='mp_trainer.master_params',
                     params=self.mp_trainer_rec.master_params,
                     lr=self.lr,
                     weight_decay=self.weight_decay)
            ]

        logger.log(optim_groups)

        return optim_groups

    @th.no_grad()
    # def eval_loop(self, c_list:list):
    def eval_novelview_loop(self):
        # novel view synthesis given evaluation camera trajectory
        video_out = imageio.get_writer(
            f'{logger.get_dir()}/video_novelview_{self.step+self.resume_step}.mp4',
            mode='I',
            fps=60,
            codec='libx264')

        all_loss_dict = []
        novel_view_micro = {}

        # for i in range(0, len(c_list), 1): # TODO, larger batch size for eval
        for i, batch in enumerate(tqdm(self.eval_data)):
            # for i in range(0, 8, self.microbatch):
            # c = c_list[i].to(dist_util.dev()).reshape(1, -1)
            micro = {k: v.to(dist_util.dev()) for k, v in batch.items()}

            if i == 0:
                novel_view_micro = {
                    k:
                    v[0:1].to(dist_util.dev()).repeat_interleave(
                        micro['img'].shape[0], 0)
                    if isinstance(v, th.Tensor) else v[0:1]
                    for k, v in batch.items()
                }
            else:
                # if novel_view_micro['c'].shape[0] < micro['img'].shape[0]:
                novel_view_micro = {
                    k:
                    v[0:1].to(dist_util.dev()).repeat_interleave(
                        micro['img'].shape[0], 0)
                    for k, v in novel_view_micro.items()
                }

            pred = self.rec_model(img=novel_view_micro['img_to_encoder'],
                                  c=micro['c'])  # pred: (B, 3, 64, 64)
            # target = {
            #     'img': micro['img'],
            #     'depth': micro['depth'],
            #     'depth_mask': micro['depth_mask']
            # }
            # targe

            _, loss_dict = self.loss_class(pred, micro, test_mode=True)
            all_loss_dict.append(loss_dict)

            # ! move to other places, add tensorboard

            # pred_vis = th.cat([
            #     pred['image_raw'],
            #     -pred['image_depth'].repeat_interleave(3, dim=1)
            # ],
            #                   dim=-1)

            # normalize depth
            # if True:
            pred_depth = pred['image_depth']
            pred_depth = (pred_depth - pred_depth.min()) / (pred_depth.max() -
                                                            pred_depth.min())
            if 'image_sr' in pred:

                if pred['image_sr'].shape[-1] == 512:

                    pred_vis = th.cat([
                        micro['img_sr'],
                        self.pool_512(pred['image_raw']), pred['image_sr'],
                        self.pool_512(pred_depth).repeat_interleave(3, dim=1)
                    ],
                                      dim=-1)

                elif pred['image_sr'].shape[-1] == 256:

                    pred_vis = th.cat([
                        micro['img_sr'],
                        self.pool_256(pred['image_raw']), pred['image_sr'],
                        self.pool_256(pred_depth).repeat_interleave(3, dim=1)
                    ],
                                      dim=-1)

                else:
                    pred_vis = th.cat([
                        micro['img_sr'],
                        self.pool_128(pred['image_raw']),
                        self.pool_128(pred['image_sr']),
                        self.pool_128(pred_depth).repeat_interleave(3, dim=1)
                    ],
                                      dim=-1)

            else:
                # pred_vis = th.cat([
                #     self.pool_64(micro['img']), pred['image_raw'],
                #     pred_depth.repeat_interleave(3, dim=1)
                # ],
                #                   dim=-1)  # B, 3, H, W

                pred_vis = th.cat([
                    self.pool_128(micro['img']),
                    self.pool_128(pred['image_raw']),
                    self.pool_128(pred_depth).repeat_interleave(3, dim=1)
                ],
                                  dim=-1)  # B, 3, H, W

            vis = pred_vis.permute(0, 2, 3, 1).cpu().numpy()
            vis = vis * 127.5 + 127.5
            vis = vis.clip(0, 255).astype(np.uint8)

            for j in range(vis.shape[0]):
                video_out.append_data(vis[j])

        video_out.close()

        val_scores_for_logging = calc_average_loss(all_loss_dict)
        with open(os.path.join(logger.get_dir(), 'scores_novelview.json'),
                  'a') as f:
            json.dump({'step': self.step, **val_scores_for_logging}, f)

        # * log to tensorboard
        for k, v in val_scores_for_logging.items():
            self.writer.add_scalar(f'Eval/NovelView/{k}', v,
                                   self.step + self.resume_step)
        del video_out
        # del pred_vis
        # del pred

        th.cuda.empty_cache()

    # @th.no_grad()
    # def eval_loop(self, c_list:list):
    @th.inference_mode()
    def eval_loop(self):
        # novel view synthesis given evaluation camera trajectory
        video_out = imageio.get_writer(
            f'{logger.get_dir()}/video_{self.step+self.resume_step}.mp4',
            mode='I',
            fps=60,
            codec='libx264')
        all_loss_dict = []
        self.rec_model.eval()

        # for i in range(0, len(c_list), 1): # TODO, larger batch size for eval
        for i, batch in enumerate(tqdm(self.eval_data)):
            # for i in range(0, 8, self.microbatch):
            # c = c_list[i].to(dist_util.dev()).reshape(1, -1)
            micro = {
                k: v.to(dist_util.dev()) if isinstance(v, th.Tensor) else v
                for k, v in batch.items()
            }

            pred = self.rec_model(img=micro['img_to_encoder'],
                                  c=micro['c'])  # pred: (B, 3, 64, 64)
            # target = {
            #     'img': micro['img'],
            #     'depth': micro['depth'],
            #     'depth_mask': micro['depth_mask']
            # }

            # if last_batch or not self.use_ddp:
            #     loss, loss_dict = self.loss_class(pred, target)
            # else:
            #     with self.ddp_model.no_sync():  # type: ignore
            _, loss_dict = self.loss_class(pred, micro, test_mode=True)
            all_loss_dict.append(loss_dict)

            # ! move to other places, add tensorboard
            # gt_vis = th.cat([micro['img'], micro['img']], dim=-1) # TODO, fail to load depth. range [0, 1]
            # pred_vis = th.cat([
            #     pred['image_raw'],
            #     -pred['image_depth'].repeat_interleave(3, dim=1)
            # ],
            #                   dim=-1)
            # vis = th.cat([gt_vis, pred_vis], dim=-2)[0].permute(1,2,0).cpu().numpy() # ! pred in range[-1, 1]

            # normalize depth
            # if True:
            pred_depth = pred['image_depth']
            pred_depth = (pred_depth - pred_depth.min()) / (pred_depth.max() -
                                                            pred_depth.min())

            if 'image_sr' in pred:

                if pred['image_sr'].shape[-1] == 512:

                    pred_vis = th.cat([
                        micro['img_sr'],
                        self.pool_512(pred['image_raw']), pred['image_sr'],
                        self.pool_512(pred_depth).repeat_interleave(3, dim=1)
                    ],
                                      dim=-1)

                elif pred['image_sr'].shape[-1] == 256:
                    pred_vis = th.cat([
                        micro['img_sr'],
                        self.pool_256(pred['image_raw']), pred['image_sr'],
                        self.pool_256(pred_depth).repeat_interleave(3, dim=1)
                    ],
                                      dim=-1)

                else:
                    pred_vis = th.cat([
                        micro['img_sr'],
                        self.pool_128(pred['image_raw']),
                        self.pool_128(pred['image_sr']),
                        self.pool_128(pred_depth).repeat_interleave(3, dim=1)
                    ],
                                      dim=-1)

            else:
                pred_vis = th.cat([
                    self.pool_128(micro['img']),
                    self.pool_128(pred['image_raw']),
                    self.pool_128(pred_depth).repeat_interleave(3, dim=1)
                ],
                                  dim=-1)  # B, 3, H, W

            vis = pred_vis.permute(0, 2, 3, 1).cpu().numpy()
            vis = vis * 127.5 + 127.5
            vis = vis.clip(0, 255).astype(np.uint8)

            for j in range(vis.shape[0]):
                video_out.append_data(vis[j])

        video_out.close()

        val_scores_for_logging = calc_average_loss(all_loss_dict)
        with open(os.path.join(logger.get_dir(), 'scores.json'), 'a') as f:
            json.dump({'step': self.step, **val_scores_for_logging}, f)

        # * log to tensorboard
        for k, v in val_scores_for_logging.items():
            self.writer.add_scalar(f'Eval/Rec/{k}', v,
                                   self.step + self.resume_step)

        th.cuda.empty_cache()
        # if 'SuperresolutionHybrid8X' in self.rendering_kwargs: # ffhq/afhq
        #     self.eval_novelview_loop_trajectory()
        # else:
        self.eval_novelview_loop()
        self.rec_model.train()

    @th.inference_mode()
    def eval_novelview_loop_trajectory(self):
        # novel view synthesis given evaluation camera trajectory
        # for i in range(0, len(c_list), 1): # TODO, larger batch size for eval
        for i, batch in enumerate(tqdm(self.eval_data)):
            micro = {k: v.to(dist_util.dev()) for k, v in batch.items()}

            video_out = imageio.get_writer(
                f'{logger.get_dir()}/video_novelview_{self.step+self.resume_step}_batch_{i}.mp4',
                mode='I',
                fps=60,
                codec='libx264')

            for idx, c in enumerate(self.all_nvs_params):
                pred = self.rec_model(img=micro['img_to_encoder'],
                                      c=c.unsqueeze(0).repeat_interleave(
                                          micro['img'].shape[0],
                                          0))  # pred: (B, 3, 64, 64)
                #   c=micro['c'])  # pred: (B, 3, 64, 64)

                # normalize depth
                # if True:
                pred_depth = pred['image_depth']
                pred_depth = (pred_depth - pred_depth.min()) / (
                    pred_depth.max() - pred_depth.min())
                if 'image_sr' in pred:

                    if pred['image_sr'].shape[-1] == 512:

                        pred_vis = th.cat([
                            micro['img_sr'],
                            self.pool_512(pred['image_raw']), pred['image_sr'],
                            self.pool_512(pred_depth).repeat_interleave(3,
                                                                        dim=1)
                        ],
                                          dim=-1)

                    elif pred['image_sr'].shape[-1] == 256:

                        pred_vis = th.cat([
                            micro['img_sr'],
                            self.pool_256(pred['image_raw']), pred['image_sr'],
                            self.pool_256(pred_depth).repeat_interleave(3,
                                                                        dim=1)
                        ],
                                          dim=-1)

                    else:
                        pred_vis = th.cat([
                            micro['img_sr'],
                            self.pool_128(pred['image_raw']),
                            self.pool_128(pred['image_sr']),
                            self.pool_128(pred_depth).repeat_interleave(3,
                                                                        dim=1)
                        ],
                                          dim=-1)

                else:

                    # st()
                    pred_vis = th.cat([
                        self.pool_128(micro['img']),
                        self.pool_128(pred['image_raw']),
                        self.pool_128(pred_depth).repeat_interleave(3, dim=1)
                    ],
                                      dim=-1)  # B, 3, H, W

                    # ! cooncat h dim
                    pred_vis = pred_vis.permute(0, 2, 3, 1).flatten(0,
                                                                    1)  # H W 3

                # vis = pred_vis.permute(0, 2, 3, 1).cpu().numpy()
                # vis = pred_vis.permute(1,2,0).cpu().numpy()
                vis = pred_vis.cpu().numpy()
                vis = vis * 127.5 + 127.5
                vis = vis.clip(0, 255).astype(np.uint8)

                # for j in range(vis.shape[0]):
                # video_out.append_data(vis[j])
                video_out.append_data(vis)

            video_out.close()

        th.cuda.empty_cache()

    def _prepare_nvs_pose(self):

        device = dist_util.dev()

        fov_deg = 18.837  # for ffhq/afhq
        intrinsics = FOV_to_intrinsics(fov_deg, device=device)

        all_nvs_params = []

        pitch_range = 0.25
        yaw_range = 0.35
        num_keyframes = 10  # how many nv poses to sample from
        w_frames = 1

        cam_pivot = th.Tensor(
            self.rendering_kwargs.get('avg_camera_pivot')).to(device)
        cam_radius = self.rendering_kwargs.get('avg_camera_radius')

        for frame_idx in range(num_keyframes):

            cam2world_pose = LookAtPoseSampler.sample(
                3.14 / 2 + yaw_range * np.sin(2 * 3.14 * frame_idx /
                                              (num_keyframes * w_frames)),
                3.14 / 2 - 0.05 +
                pitch_range * np.cos(2 * 3.14 * frame_idx /
                                     (num_keyframes * w_frames)),
                cam_pivot,
                radius=cam_radius,
                device=device)

            camera_params = th.cat(
                [cam2world_pose.reshape(-1, 16),
                 intrinsics.reshape(-1, 9)], 1)

            all_nvs_params.append(camera_params)

        self.all_nvs_params = th.cat(all_nvs_params, 0)

    def forward_backward(self, batch, *args, **kwargs):
        # th.cuda.empty_cache()
        self.mp_trainer_rec.zero_grad()
        batch_size = batch['img_to_encoder'].shape[0]

        for i in range(0, batch_size, self.microbatch):

            micro = {
                k: v[i:i + self.microbatch].to(dist_util.dev())
                for k, v in batch.items()
            }

            last_batch = (i + self.microbatch) >= batch_size

            # wrap forward within amp
            with th.autocast(device_type='cuda',
                             dtype=th.float16,
                             enabled=self.mp_trainer_rec.use_amp):

                pred = self.rec_model(img=micro['img_to_encoder'],
                                      c=micro['c'])  # pred: (B, 3, 64, 64)
                target = micro

                # ! only enable in ffhq dataset
                conf_sigma_percl = None
                conf_sigma_percl_flip = None
                if 'conf_sigma' in pred:
                    # all_conf_sigma_l1, all_conf_sigma_percl = pred['conf_sigma']
                    # all_conf_sigma_l1  = pred['conf_sigma']
                    all_conf_sigma_l1 = th.nn.functional.interpolate(
                        pred['conf_sigma'],
                        size=pred['image_raw'].shape[-2:],
                        mode='bilinear'
                    )  # dynamically resize to target img size
                    conf_sigma_l1 = all_conf_sigma_l1[:, :1]
                    conf_sigma_l1_flip = all_conf_sigma_l1[:, 1:]
                    # conf_sigma_percl = all_conf_sigma_percl[:,:1]
                    # conf_sigma_percl_flip = all_conf_sigma_percl[:,1:]
                else:
                    conf_sigma = None
                    conf_sigma_l1 = None
                    conf_sigma_l1_flip = None

                with self.rec_model.no_sync():  # type: ignore
                    loss, loss_dict, fg_mask = self.loss_class(
                        pred,
                        target,
                        step=self.step + self.resume_step,
                        test_mode=False,
                        return_fg_mask=True,
                        conf_sigma_l1=conf_sigma_l1,
                        conf_sigma_percl=conf_sigma_percl)

                if self.loss_class.opt.symmetry_loss:
                    loss_dict['conf_sigma_log'] = conf_sigma_l1.log()
                    pose, intrinsics = micro['c'][:, :16].reshape(
                        -1, 4, 4), micro['c'][:, 16:]
                    flipped_pose = flip_yaw(pose)
                    mirror_c = th.cat(
                        [flipped_pose.reshape(-1, 16), intrinsics], -1)

                    nvs_pred = self.rec_model(latent={
                        k: v
                        for k, v in pred.items() if 'latent' in k
                    },
                                              c=mirror_c,
                                              behaviour='triplane_dec',
                                              return_raw_only=True)

                    # concat data for supervision
                    nvs_gt = {
                        k: th.flip(target[k], [-1])
                        for k in
                        ['img']  # fliplr leads to wrong color; B 3 H W shape
                    }
                    flipped_fg_mask = th.flip(fg_mask, [-1])

                    # if 'conf_sigma' in pred:
                    #     conf_sigma = th.flip(pred['conf_sigma'], [-1])
                    #     conf_sigma = th.nn.AdaptiveAvgPool2d(fg_mask.shape[-2:])(conf_sigma) # dynamically resize to target img size
                    # else:
                    #     conf_sigma=None

                    with self.rec_model.no_sync():  # type: ignore
                        loss_symm, loss_dict_symm = self.loss_class.calc_2d_rec_loss(
                            nvs_pred['image_raw'],
                            nvs_gt['img'],
                            flipped_fg_mask,
                            # test_mode=True,
                            test_mode=False,
                            step=self.step + self.resume_step,
                            # conf_sigma=conf_sigma,
                            conf_sigma_l1=conf_sigma_l1_flip,
                            conf_sigma_percl=conf_sigma_percl_flip)
                        # )
                        loss += (loss_symm * 1.0)  # as in unsup3d
                        # loss += (loss_symm * 0.5) # as in unsup3d
                        # loss += (loss_symm * 0.01) # as in unsup3d
                        # if conf_sigma is not None:
                        #     loss += th.nn.functional.mse_loss(conf_sigma, flipped_fg_mask) * 0.001 # a log that regularizes all confidence to 1
                        for k, v in loss_dict_symm.items():
                            loss_dict[f'{k}_symm'] = v
                        loss_dict[
                            'flip_conf_sigma_log'] = conf_sigma_l1_flip.log()

                # ! add density-reg in eg3d, tv-loss

                if self.loss_class.opt.density_reg > 0 and self.step % self.loss_class.opt.density_reg_every == 0:

                    initial_coordinates = th.rand(
                        (batch_size, 1000, 3),
                        device=dist_util.dev()) * 2 - 1  # [-1, 1]
                    perturbed_coordinates = initial_coordinates + th.randn_like(
                        initial_coordinates
                    ) * self.loss_class.opt.density_reg_p_dist
                    all_coordinates = th.cat(
                        [initial_coordinates, perturbed_coordinates], dim=1)

                    sigma = self.rec_model(
                        latent=pred['latent'],
                        coordinates=all_coordinates,
                        directions=th.randn_like(all_coordinates),
                        behaviour='triplane_renderer',
                    )['sigma']

                    sigma_initial = sigma[:, :sigma.shape[1] // 2]
                    sigma_perturbed = sigma[:, sigma.shape[1] // 2:]

                    TVloss = th.nn.functional.l1_loss(
                        sigma_initial,
                        sigma_perturbed) * self.loss_class.opt.density_reg

                    loss_dict.update(dict(tv_loss=TVloss))
                    loss += TVloss

            self.mp_trainer_rec.backward(loss)
            log_rec3d_loss_dict(loss_dict)

            # for name, p in self.rec_model.named_parameters():
            #     if p.grad is None:
            #         logger.log(f"found rec unused param: {name}")

            if dist_util.get_rank() == 0 and self.step % 500 == 0:
                with th.no_grad():
                    # gt_vis = th.cat([batch['img'], batch['depth']], dim=-1)

                    def norm_depth(pred_depth):  # to [-1,1]
                        # pred_depth = pred['image_depth']
                        pred_depth = (pred_depth - pred_depth.min()) / (
                            pred_depth.max() - pred_depth.min())
                        return -(pred_depth * 2 - 1)

                    pred_img = pred['image_raw']
                    gt_img = micro['img']

                    # infer novel view also
                    if self.loss_class.opt.symmetry_loss:
                        pred_nv_img = nvs_pred
                    else:
                        pred_nv_img = self.rec_model(
                            img=micro['img_to_encoder'],
                            c=self.novel_view_poses)  # pred: (B, 3, 64, 64)

                    # if 'depth' in micro:
                    gt_depth = micro['depth']
                    if gt_depth.ndim == 3:
                        gt_depth = gt_depth.unsqueeze(1)
                    gt_depth = norm_depth(gt_depth)
                    # gt_depth = (gt_depth - gt_depth.min()) / (gt_depth.max() -
                    #                                           gt_depth.min())
                    # if True:
                    fg_mask = pred['image_mask'] * 2 - 1  # 0-1
                    nv_fg_mask = pred_nv_img['image_mask'] * 2 - 1  # 0-1
                    if 'image_depth' in pred:
                        pred_depth = norm_depth(pred['image_depth'])
                        pred_nv_depth = norm_depth(pred_nv_img['image_depth'])
                    else:
                        pred_depth = th.zeros_like(gt_depth)
                        pred_nv_depth = th.zeros_like(gt_depth)

                    if 'image_sr' in pred:
                        if pred['image_sr'].shape[-1] == 512:
                            pred_img = th.cat(
                                [self.pool_512(pred_img), pred['image_sr']],
                                dim=-1)
                            gt_img = th.cat(
                                [self.pool_512(micro['img']), micro['img_sr']],
                                dim=-1)
                            pred_depth = self.pool_512(pred_depth)
                            gt_depth = self.pool_512(gt_depth)

                        elif pred['image_sr'].shape[-1] == 256:
                            pred_img = th.cat(
                                [self.pool_256(pred_img), pred['image_sr']],
                                dim=-1)
                            gt_img = th.cat(
                                [self.pool_256(micro['img']), micro['img_sr']],
                                dim=-1)
                            pred_depth = self.pool_256(pred_depth)
                            gt_depth = self.pool_256(gt_depth)

                        else:
                            pred_img = th.cat(
                                [self.pool_128(pred_img), pred['image_sr']],
                                dim=-1)
                            gt_img = th.cat(
                                [self.pool_128(micro['img']), micro['img_sr']],
                                dim=-1)
                            gt_depth = self.pool_128(gt_depth)
                            pred_depth = self.pool_128(pred_depth)
                    else:
                        gt_img = self.pool_128(gt_img)
                        gt_depth = self.pool_128(gt_depth)

                    pred_vis = th.cat([
                        pred_img,
                        pred_depth.repeat_interleave(3, dim=1),
                        fg_mask.repeat_interleave(3, dim=1),
                    ],
                                      dim=-1)  # B, 3, H, W

                    if 'conf_sigma' in pred:
                        conf_sigma_l1 = (1 / (conf_sigma_l1 + 1e-7)
                                         ).repeat_interleave(3, dim=1) * 2 - 1
                        pred_vis = th.cat([
                            pred_vis,
                            conf_sigma_l1,
                        ], dim=-1)  # B, 3, H, W

                    pred_vis_nv = th.cat([
                        pred_nv_img['image_raw'],
                        pred_nv_depth.repeat_interleave(3, dim=1),
                        nv_fg_mask.repeat_interleave(3, dim=1),
                    ],
                                         dim=-1)  # B, 3, H, W

                    if 'conf_sigma' in pred:
                        # conf_sigma_for_vis = (1/conf_sigma).repeat_interleave(3, dim=1)
                        # conf_sigma_for_vis = (conf_sigma_for_vis / conf_sigma_for_vis.max() ) * 2 - 1 # normalize to [-1,1]
                        conf_sigma_for_vis_flip = (
                            1 / (conf_sigma_l1_flip + 1e-7)).repeat_interleave(
                                3, dim=1) * 2 - 1
                        pred_vis_nv = th.cat(
                            [
                                pred_vis_nv,
                                conf_sigma_for_vis_flip,
                                # th.cat([conf_sigma_for_vis, flipped_fg_mask*2-1], -1)
                            ],
                            dim=-1)  # B, 3, H, W

                    pred_vis = th.cat([pred_vis, pred_vis_nv],
                                      dim=-2)  # cat in H dim

                    gt_vis = th.cat(
                        [
                            gt_img,
                            gt_depth.repeat_interleave(3, dim=1),
                            th.zeros_like(gt_img)
                        ],
                        dim=-1)  # TODO, fail to load depth. range [0, 1]

                    if 'conf_sigma' in pred:
                        gt_vis = th.cat([gt_vis, fg_mask],
                                        dim=-1)  # placeholder

                    # vis = th.cat([gt_vis, pred_vis], dim=-2)[0].permute(
                    # st()
                    vis = th.cat([gt_vis, pred_vis], dim=-2)
                    # .permute(
                    #     0, 2, 3, 1).cpu()
                    vis_tensor = torchvision.utils.make_grid(
                        vis, nrow=vis.shape[-1] // 64)  # HWC
                    torchvision.utils.save_image(
                        vis_tensor,
                        f'{logger.get_dir()}/{self.step+self.resume_step}.jpg',
                        value_range=(-1, 1),
                        normalize=True)
                    # vis = vis.numpy() * 127.5 + 127.5
                    # vis = vis.clip(0, 255).astype(np.uint8)

                    # Image.fromarray(vis).save(
                    #     f'{logger.get_dir()}/{self.step+self.resume_step}.jpg')

                    logger.log(
                        'log vis to: ',
                        f'{logger.get_dir()}/{self.step+self.resume_step}.jpg')

                    # self.writer.add_image(f'images',
                    #                       vis,
                    #                       self.step + self.resume_step,
                    #                       dataformats='HWC')
            return pred


class TrainLoop3DTriplaneRec(TrainLoop3DRec):

    def __init__(self,
                 *,
                 rec_model,
                 loss_class,
                 data,
                 eval_data,
                 batch_size,
                 microbatch,
                 lr,
                 ema_rate,
                 log_interval,
                 eval_interval,
                 save_interval,
                 resume_checkpoint,
                 use_fp16=False,
                 fp16_scale_growth=0.001,
                 weight_decay=0,
                 lr_anneal_steps=0,
                 iterations=10001,
                 load_submodule_name='',
                 ignore_resume_opt=False,
                 model_name='rec',
                 use_amp=False,
                 compile=False,
                 **kwargs):
        super().__init__(rec_model=rec_model,
                         loss_class=loss_class,
                         data=data,
                         eval_data=eval_data,
                         batch_size=batch_size,
                         microbatch=microbatch,
                         lr=lr,
                         ema_rate=ema_rate,
                         log_interval=log_interval,
                         eval_interval=eval_interval,
                         save_interval=save_interval,
                         resume_checkpoint=resume_checkpoint,
                         use_fp16=use_fp16,
                         fp16_scale_growth=fp16_scale_growth,
                         weight_decay=weight_decay,
                         lr_anneal_steps=lr_anneal_steps,
                         iterations=iterations,
                         load_submodule_name=load_submodule_name,
                         ignore_resume_opt=ignore_resume_opt,
                         model_name=model_name,
                         use_amp=use_amp,
                         compile=compile,
                         **kwargs)

    @th.inference_mode()
    def eval_loop(self):
        # novel view synthesis given evaluation camera trajectory
        video_out = imageio.get_writer(
            f'{logger.get_dir()}/video_{self.step+self.resume_step}.mp4',
            mode='I',
            fps=60,
            codec='libx264')
        all_loss_dict = []
        self.rec_model.eval()

        device = dist_util.dev()

        # to get intrinsics
        demo_pose = next(self.data)
        intrinsics = demo_pose['c'][0][16:25].to(device)

        video_out = imageio.get_writer(
            f'{logger.get_dir()}/video_{self.step+self.resume_step}.mp4',
            mode='I',
            fps=24,
            bitrate='10M',
            codec='libx264')

        # for i in range(0, len(c_list), 1): # TODO, larger batch size for eval
        # for i, batch in enumerate(tqdm(self.eval_data)):

        cam_pivot = th.tensor([0, 0, 0], device=dist_util.dev())
        cam_radius = 1.8

        pitch_range = 0.45
        yaw_range = 3.14  # 0.35
        frames = 72

        # TODO, use PanoHead trajectory
        # for frame_idx in range(frames):

        for pose_idx, (angle_y, angle_p) in enumerate(
                # zip(np.linspace(-0.4, 0.4, 72), [-0.2] * 72)):
                # zip(np.linspace(-1.57, 1.57, 72), [-1.57] * 72)):
                # zip(np.linspace(0,3.14, 72), [0] * 72)): # check canonical pose
                zip([0.2] * 72, np.linspace(-3.14, 3.14, 72))):

            # cam2world_pose = LookAtPoseSampler.sample(3.14/2 + yaw_range * np.cos(2 * 3.14 * frame_idx / (frames)),
            #                                         3.14/2 -0.05 + pitch_range * np.sin(2 * 3.14 * frame_idx / (frames)),
            #                                         cam_pivot,
            #                                         radius=cam_radius, device=device)

            cam2world_pose = LookAtPoseSampler.sample(
                np.pi / 2 + angle_y,
                np.pi / 2 + angle_p,
                # angle_p,
                cam_pivot,
                #    horizontal_stddev=0.1, # 0.25
                #    vertical_stddev=0.125, # 0.35,
                radius=cam_radius,
                device=device)

            camera_params = th.cat(
                [cam2world_pose.reshape(-1, 16),
                 intrinsics.reshape(-1, 9)], 1).to(dist_util.dev())

            # micro = {k: v.to(dist_util.dev()) for k, v in batch.items()}
            micro = {'c': camera_params}

            pred = self.rec_model(c=micro['c'])

            # normalize depth
            # if True:
            pred_depth = pred['image_depth']
            pred_depth = (pred_depth - pred_depth.min()) / (pred_depth.max() -
                                                            pred_depth.min())

            pred_vis = th.cat([
                self.pool_128(pred['image_raw']),
                self.pool_128(pred_depth).repeat_interleave(3, dim=1)
            ],
                              dim=-1)  # B, 3, H, W

            vis = pred_vis.permute(0, 2, 3, 1).cpu().numpy()
            vis = vis * 127.5 + 127.5
            vis = vis.clip(0, 255).astype(np.uint8)

            for j in range(vis.shape[0]):
                video_out.append_data(vis[j])

        video_out.close()

        self.rec_model.train()


class TrainLoop3DRecTrajVis(TrainLoop3DRec):

    def __init__(self,
                 *,
                 rec_model,
                 loss_class,
                 data,
                 eval_data,
                 batch_size,
                 microbatch,
                 lr,
                 ema_rate,
                 log_interval,
                 eval_interval,
                 save_interval,
                 resume_checkpoint,
                 use_fp16=False,
                 fp16_scale_growth=0.001,
                 weight_decay=0,
                 lr_anneal_steps=0,
                 iterations=10001,
                 load_submodule_name='',
                 ignore_resume_opt=False,
                 model_name='rec',
                 use_amp=False,
                 **kwargs):
        super().__init__(rec_model=rec_model,
                         loss_class=loss_class,
                         data=data,
                         eval_data=eval_data,
                         batch_size=batch_size,
                         microbatch=microbatch,
                         lr=lr,
                         ema_rate=ema_rate,
                         log_interval=log_interval,
                         eval_interval=eval_interval,
                         save_interval=save_interval,
                         resume_checkpoint=resume_checkpoint,
                         use_fp16=use_fp16,
                         fp16_scale_growth=fp16_scale_growth,
                         weight_decay=weight_decay,
                         lr_anneal_steps=lr_anneal_steps,
                         iterations=iterations,
                         load_submodule_name=load_submodule_name,
                         ignore_resume_opt=ignore_resume_opt,
                         model_name=model_name,
                         use_amp=use_amp,
                         **kwargs)
        self.rendering_kwargs = self.rec_model.module.decoder.triplane_decoder.rendering_kwargs  # type: ignore
        self._prepare_nvs_pose()  # for eval novelview visualization

    @th.inference_mode()
    def eval_novelview_loop(self):
        # novel view synthesis given evaluation camera trajectory
        # for i in range(0, len(c_list), 1): # TODO, larger batch size for eval
        for i, batch in enumerate(tqdm(self.eval_data)):
            micro = {k: v.to(dist_util.dev()) for k, v in batch.items()}

            video_out = imageio.get_writer(
                f'{logger.get_dir()}/video_novelview_{self.step+self.resume_step}_batch_{i}.mp4',
                mode='I',
                fps=60,
                codec='libx264')

            for idx, c in enumerate(self.all_nvs_params):
                pred = self.rec_model(img=micro['img_to_encoder'],
                                      c=c.unsqueeze(0).repeat_interleave(
                                          micro['img'].shape[0],
                                          0))  # pred: (B, 3, 64, 64)
                #   c=micro['c'])  # pred: (B, 3, 64, 64)

                # normalize depth
                # if True:
                pred_depth = pred['image_depth']
                pred_depth = (pred_depth - pred_depth.min()) / (
                    pred_depth.max() - pred_depth.min())
                if 'image_sr' in pred:

                    if pred['image_sr'].shape[-1] == 512:

                        pred_vis = th.cat([
                            micro['img_sr'],
                            self.pool_512(pred['image_raw']), pred['image_sr'],
                            self.pool_512(pred_depth).repeat_interleave(3,
                                                                        dim=1)
                        ],
                                          dim=-1)

                    elif pred['image_sr'].shape[-1] == 256:

                        pred_vis = th.cat([
                            micro['img_sr'],
                            self.pool_256(pred['image_raw']), pred['image_sr'],
                            self.pool_256(pred_depth).repeat_interleave(3,
                                                                        dim=1)
                        ],
                                          dim=-1)

                    else:
                        pred_vis = th.cat([
                            micro['img_sr'],
                            self.pool_128(pred['image_raw']),
                            self.pool_128(pred['image_sr']),
                            self.pool_128(pred_depth).repeat_interleave(3,
                                                                        dim=1)
                        ],
                                          dim=-1)

                else:

                    # st()
                    pred_vis = th.cat([
                        self.pool_128(micro['img']),
                        self.pool_128(pred['image_raw']),
                        self.pool_128(pred_depth).repeat_interleave(3, dim=1)
                    ],
                                      dim=-1)  # B, 3, H, W

                    # ! cooncat h dim
                    pred_vis = pred_vis.permute(0, 2, 3, 1).flatten(0,
                                                                    1)  # H W 3

                # vis = pred_vis.permute(0, 2, 3, 1).cpu().numpy()
                # vis = pred_vis.permute(1,2,0).cpu().numpy()
                vis = pred_vis.cpu().numpy()
                vis = vis * 127.5 + 127.5
                vis = vis.clip(0, 255).astype(np.uint8)

                # for j in range(vis.shape[0]):
                # video_out.append_data(vis[j])
                video_out.append_data(vis)

            video_out.close()

        th.cuda.empty_cache()

    def _prepare_nvs_pose(self):

        device = dist_util.dev()

        fov_deg = 18.837  # for ffhq/afhq
        intrinsics = FOV_to_intrinsics(fov_deg, device=device)

        all_nvs_params = []

        pitch_range = 0.25
        yaw_range = 0.35
        num_keyframes = 10  # how many nv poses to sample from
        w_frames = 1

        cam_pivot = th.Tensor(
            self.rendering_kwargs.get('avg_camera_pivot')).to(device)
        cam_radius = self.rendering_kwargs.get('avg_camera_radius')

        for frame_idx in range(num_keyframes):

            cam2world_pose = LookAtPoseSampler.sample(
                3.14 / 2 + yaw_range * np.sin(2 * 3.14 * frame_idx /
                                              (num_keyframes * w_frames)),
                3.14 / 2 - 0.05 +
                pitch_range * np.cos(2 * 3.14 * frame_idx /
                                     (num_keyframes * w_frames)),
                cam_pivot,
                radius=cam_radius,
                device=device)

            camera_params = th.cat(
                [cam2world_pose.reshape(-1, 16),
                 intrinsics.reshape(-1, 9)], 1)

            all_nvs_params.append(camera_params)

        self.all_nvs_params = th.cat(all_nvs_params, 0)