run_nerf.py

import os, sys
import numpy as np
import imageio
import json
import random
import time
import torch
import torch.nn as nn
import torch.nn.functional as F
from tqdm import tqdm, trange

import lpips as LPI

from skimage.metrics import structural_similarity as ssim_metric
from config import config_parser
import matplotlib.pyplot as plt

from ray_utils import *
from run_nerf_helpers import *

from load_llff import load_llff_data
from load_deepvoxels import load_dv_data
from load_blender import load_blender_data
from load_LINEMOD import load_LINEMOD_data


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
np.random.seed(0)
DEBUG = False


def batchify(fn, chunk):
    """Constructs a version of 'fn' that applies to smaller batches.
    """
    if chunk is None:
        return fn
    def ret(inputs, view_dirs):
        return torch.cat([fn(inputs[i:i+chunk], view_dirs[i:i+chunk]) for i in range(0, inputs.shape[0], chunk)], 0)
    return ret


def run_network(inputs, viewdirs, fn, embed_fn, embeddirs_fn, netchunk=1024*64):
    """Prepares inputs and applies network 'fn'.
    """
    # input shape: (2, Batch_size=1024, num_sampels=64, 3) 2: means+covs
    means_flag, covs_flag = torch.reshape(inputs, [2, -1, inputs.shape[-1]])
    embedded = embed_fn(means_flag, covs_flag)

    if viewdirs is not None:
        input_dirs = viewdirs[:,None].expand(inputs.shape[1:])
        input_dirs_flat = torch.reshape(input_dirs, [-1, input_dirs.shape[-1]])
        embedded_dirs = embeddirs_fn(input_dirs_flat)
        embedded = torch.cat([embedded, embedded_dirs], -1)

    outputs_flat = batchify(fn, netchunk)(embedded, input_dirs_flat)
    outputs = torch.reshape(outputs_flat, list(inputs.shape[1:-1]) + [outputs_flat.shape[-1]])
    return outputs


def batchify_rays(rays_flat, rays_radii, chunk=1024*32, **kwargs):
    """Render rays in smaller minibatches to avoid OOM.
    """
    all_ret = {}
    for i in range(0, rays_flat.shape[0], chunk):
        ret = render_rays(rays_flat[i:i+chunk], rays_radii[i:i+chunk], **kwargs)
        for k in ret:
            if k not in all_ret:
                all_ret[k] = []
            all_ret[k].append(ret[k])

    all_ret = {k : torch.cat(all_ret[k], 0) for k in all_ret}
    return all_ret


def render(H, W, K, chunk=1024*32, rays=None, c2w=None, ndc=True, batch_radii=None,
                  near=0., far=1.,
                  use_viewdirs=False, c2w_staticcam=None,
                  **kwargs):
    """Render rays
    Args:
      H: int. Height of image in pixels.
      W: int. Width of image in pixels.
      focal: float. Focal length of pinhole camera.
      chunk: int. Maximum number of rays to process simultaneously. Used to
        control maximum memory usage. Does not affect final results.
      rays: array of shape [2, batch_size, 3]. Ray origin and direction for
        each example in batch.
      c2w: array of shape [3, 4]. Camera-to-world transformation matrix.
      ndc: bool. If True, represent ray origin, direction in NDC coordinates.
      near: float or array of shape [batch_size]. Nearest distance for a ray.
      far: float or array of shape [batch_size]. Farthest distance for a ray.
      use_viewdirs: bool. If True, use viewing direction of a point in space in model.
      c2w_staticcam: array of shape [3, 4]. If not None, use this transformation matrix for 
       camera while using other c2w argument for viewing directions.
    Returns:
      rgb_map: [batch_size, 3]. Predicted RGB values for rays.
      disp_map: [batch_size]. Disparity map. Inverse of depth.
      acc_map: [batch_size]. Accumulated opacity (alpha) along a ray.
      extras: dict with everything returned by render_rays().
    """
    if c2w is not None:
        # special case to render full image
        rays_o, rays_d = get_rays(H, W, K, c2w)
        rays_radii = get_rays_radii(H, W, K, c2w)
        batch_radii = torch.reshape(rays_radii, [-1, 1]).float()
    else:
        # use provided ray batch
        rays_o, rays_d = rays

    if use_viewdirs:
        # provide ray directions as input
        viewdirs = rays_d
        viewdirs = viewdirs / torch.norm(viewdirs, dim=-1, keepdim=True)
        viewdirs = torch.reshape(viewdirs, [-1,3]).float()

    sh = rays_d.shape # [..., 3]
    if ndc:
        # for forward facing scenes
        rays_o, rays_d = ndc_rays(H, W, K[0][0], 1., rays_o, rays_d)

    # Create ray batch
    rays_o = torch.reshape(rays_o, [-1,3]).float()
    rays_d = torch.reshape(rays_d, [-1,3]).float()

    near, far = near * torch.ones_like(rays_d[...,:1]), far * torch.ones_like(rays_d[...,:1])
    rays = torch.cat([rays_o, rays_d, near, far], -1)
    if use_viewdirs:
        rays = torch.cat([rays, viewdirs], -1)

    # Render and reshape
    all_ret = batchify_rays(rays, batch_radii, chunk, **kwargs)
    for k in all_ret:
        k_sh = list(sh[:-1]) + list(all_ret[k].shape[1:])
        all_ret[k] = torch.reshape(all_ret[k], k_sh)

    k_extract = ['rgb_map', 'disp_map', 'acc_map']
    ret_list = [all_ret[k] for k in k_extract]
    ret_dict = {k : all_ret[k] for k in all_ret if k not in k_extract}
    return ret_list + [ret_dict]


def render_path(render_poses, hwf, K, chunk, render_kwargs, gt_imgs=None, savedir=None, render_factor=0):

    H, W, focal = hwf

    if render_factor!=0:
        # Render downsampled for speed
        H = H//render_factor
        W = W//render_factor
        focal = focal/render_factor

    rgbs = []
    disps = []

    SSIM = []
    PSNR = []
    LPIPS = []
    
    t = time.time()
    for i, c2w in enumerate(tqdm(render_poses)):
        print(i, time.time() - t)
        t = time.time()
        rgb, disp, acc, _ = render(H, W, K, chunk=chunk, c2w=c2w[:3,:4], **render_kwargs)
        rgb = rgb.cpu().numpy()
        disp = disp.cpu().numpy()
        rgbs.append(rgb)
        disps.append(disp)
        if i==0:
            print(rgb.shape, disp.shape)

        if savedir is not None:
            img2mse = lambda x, y : np.mean((x - y) ** 2)
            mse2psnr = lambda x : -10. * np.log(x) / np.log(10.)
            
            GT = gt_imgs[i].cpu().numpy()
            
            print(GT.shape, rgb.shape)
            mse = img2mse(rgb, GT)
            psnr = mse2psnr(mse)
            ssim = ssim_metric(rgb, GT, multichannel=True)
        
            loss_fn = LPI.LPIPS(net='vgg')
            lpips = loss_fn(torch.from_numpy(np.transpose(rgb, (2, 0, 1))).to(device), torch.from_numpy(np.transpose(GT, (2, 0, 1))).to(device))

            tqdm.write(f"[TEST] Iter: {i} SSIM: {ssim}  PSNR: {psnr}  LPIPS: {lpips.item()}")
            PSNR.append(psnr)
            SSIM.append(ssim)
            LPIPS.append(lpips.cpu().numpy())
            rgb8 = to8b(rgbs[-1])
            filename = os.path.join(savedir, '{:03d}.png'.format(i))
            gt_filename = os.path.join(savedir, '{:03d}_gt.png'.format(i))
            imageio.imwrite(filename, rgb8)
            imageio.imwrite(gt_filename, to8b(GT))

    if gt_imgs is not None:
        print('mean PSNR: ', np.mean(PSNR))
        print('mean SSIM: ', np.mean(SSIM))
        print('mean LPIPS: ', np.mean(LPIPS))
    
    rgbs = np.stack(rgbs, 0)
    disps = np.stack(disps, 0)

    return rgbs, disps


def create_nerf(args):
    """Instantiate NeRF's MLP model.
    """
    embed_fn, input_ch = get_embedder(args.multires, args.i_embed)

    input_ch_views = 0
    embeddirs_fn = None
    if args.use_viewdirs:
        embeddirs_fn, input_ch_views = get_embedder(args.multires_views, args.i_embed)
    output_ch = 5 if args.N_importance > 0 else 4
    skips = [4]
    model = NeRF(D=args.netdepth, W=args.netwidth,
                 input_ch=input_ch, output_ch=output_ch, skips=skips,
                 input_ch_views=input_ch_views, use_viewdirs=args.use_viewdirs).to(device)
    grad_vars = list(model.parameters())

    model_fine = None
    if args.N_importance > 0:
        model_fine = NeRF(D=args.netdepth_fine, W=args.netwidth_fine,
                          input_ch=input_ch, output_ch=output_ch, skips=skips,
                          input_ch_views=input_ch_views, use_viewdirs=args.use_viewdirs).to(device)
        grad_vars += list(model_fine.parameters())

    network_query_fn = lambda inputs, viewdirs, network_fn : run_network(inputs, viewdirs, network_fn,
                                                                embed_fn=embed_fn,
                                                                embeddirs_fn=embeddirs_fn,
                                                                netchunk=args.netchunk)

    # Create optimizer
    optimizer = torch.optim.Adam(params=grad_vars, lr=args.lrate, betas=(0.9, 0.999))

    start = 0
    basedir = args.basedir
    expname = args.expname

    ##########################

    # Load checkpoints
    if args.ft_path is not None and args.ft_path!='None':
        ckpts = [args.ft_path]
    else:
        ckpts = [os.path.join(basedir, expname, f) for f in sorted(os.listdir(os.path.join(basedir, expname))) if 'tar' in f]

    print('Found ckpts', ckpts)
    if len(ckpts) > 0 and not args.no_reload:
        ckpt_path = ckpts[-1]
        print('Reloading from', ckpt_path)
        ckpt = torch.load(ckpt_path)

        start = ckpt['global_step']
        optimizer.load_state_dict(ckpt['optimizer_state_dict'])

        # Load model
        model.load_state_dict(ckpt['network_fn_state_dict'])
        if model_fine is not None:
            model_fine.load_state_dict(ckpt['network_fine_state_dict'])

    ##########################

    render_kwargs_train = {
        'network_query_fn' : network_query_fn,
        'perturb' : args.perturb,
        'N_importance' : args.N_importance,
        'network_fine' : model_fine,
        'N_samples' : args.N_samples,
        'network_fn' : model,
        'use_viewdirs' : args.use_viewdirs,
        'white_bkgd' : args.white_bkgd,
        'raw_noise_std' : args.raw_noise_std,
    }

    # NDC only good for LLFF-style forward facing data
    if args.dataset_type != 'llff' or args.no_ndc:
        print('Not ndc!')
        render_kwargs_train['ndc'] = False
        render_kwargs_train['lindisp'] = args.lindisp

    render_kwargs_test = {k : render_kwargs_train[k] for k in render_kwargs_train}
    render_kwargs_test['perturb'] = False
    render_kwargs_test['raw_noise_std'] = 0.

    return render_kwargs_train, render_kwargs_test, start, grad_vars, optimizer

def render_rays(ray_batch,
                ray_radii,
                network_fn,
                network_query_fn,
                N_samples,
                retraw=False,
                lindisp=False,
                perturb=0.,
                N_importance=0,
                network_fine=None,
                white_bkgd=False,
                raw_noise_std=0.,
                verbose=False,
                pytest=False):
    """Volumetric rendering.
    Args:
      ray_batch: array of shape [batch_size, ...]. All information necessary
        for sampling along a ray, including: ray origin, ray direction, min
        dist, max dist, and unit-magnitude viewing direction.
      network_fn: function. Model for predicting RGB and density at each point
        in space.
      network_query_fn: function used for passing queries to network_fn.
      N_samples: int. Number of different times to sample along each ray.
      retraw: bool. If True, include model's raw, unprocessed predictions.
      lindisp: bool. If True, sample linearly in inverse depth rather than in depth.
      perturb: float, 0 or 1. If non-zero, each ray is sampled at stratified
        random points in time.
      N_importance: int. Number of additional times to sample along each ray.
        These samples are only passed to network_fine.
      network_fine: "fine" network with same spec as network_fn.
      white_bkgd: bool. If True, assume a white background.
      raw_noise_std: ...
      verbose: bool. If True, print more debugging info.
    Returns:
      rgb_map: [num_rays, 3]. Estimated RGB color of a ray. Comes from fine model.
      disp_map: [num_rays]. Disparity map. 1 / depth.
      acc_map: [num_rays]. Accumulated opacity along each ray. Comes from fine model.
      raw: [num_rays, num_samples, 4]. Raw predictions from model.
      rgb0: See rgb_map. Output for coarse model.
      disp0: See disp_map. Output for coarse model.
      acc0: See acc_map. Output for coarse model.
      z_std: [num_rays]. Standard deviation of distances along ray for each
        sample.
    """
    N_rays = ray_batch.shape[0]
    rays_o, rays_d = ray_batch[:,0:3], ray_batch[:,3:6] # [N_rays, 3] each
    viewdirs = ray_batch[:,-3:] if ray_batch.shape[-1] > 8 else None
    bounds = torch.reshape(ray_batch[...,6:8], [-1,1,2])
    near, far = bounds[...,0], bounds[...,1] # [-1,1]

    z_vals, (means, covs) = sample_along_rays(rays_o, rays_d, radii=ray_radii, num_samples=N_samples, \
        near=near, far=far, lindisp=False, ray_shape="cone", randomized=False)

    pts = torch.stack([means, covs], dim=0)
    raw = network_query_fn(pts, viewdirs, network_fn)
    rgb_map, disp_map, acc_map, weights, alpha = volumetric_rendering(rgb=raw[..., :3], density=raw[..., 3, None], t_vals=z_vals, dirs=rays_d, white_bkgd=white_bkgd)

    if N_importance > 0:

        rgb_map_0, disp_map_0, acc_map_0 = rgb_map, disp_map, acc_map

        z_vals, (means, covs) = resample_along_rays(rays_o, rays_d, radii=ray_radii, t_vals=z_vals, weights=weights, randomized=False, \
                        stop_grad=True, ray_shape="cone", resample_padding=0.01)
        pts = torch.stack([means, covs], dim=0)

        run_fn = network_fn if network_fine is None else network_fine
        raw = network_query_fn(pts, viewdirs, run_fn)

        rgb_map, disp_map, acc_map, weights, alpha = volumetric_rendering(rgb=raw[..., :3], density=raw[..., 3, None], t_vals=z_vals, dirs=rays_d, white_bkgd=white_bkgd)

    ret = {'rgb_map' : rgb_map, 'disp_map' : disp_map, 'acc_map' : acc_map}
    if retraw:
        ret['raw'] = raw
    if N_importance > 0:
        ret['rgb0'] = rgb_map_0
        ret['disp0'] = disp_map_0
        ret['acc0'] = acc_map_0

    for k in ret:
        if (torch.isnan(ret[k]).any() or torch.isinf(ret[k]).any()) and DEBUG:
            print(f"! [Numerical Error] {k} contains nan or inf.")

    return ret




def train():

    parser = config_parser()
    args = parser.parse_args()

    # Load data
    K = None
    if args.dataset_type == 'llff':
        images, poses, bds, render_poses, i_test = load_llff_data(args.datadir, args.factor,
                                                                  recenter=True, bd_factor=.75,
                                                                  spherify=args.spherify)
        hwf = poses[0,:3,-1]
        poses = poses[:,:3,:4]
        print('Loaded llff', images.shape, render_poses.shape, hwf, args.datadir)
        if not isinstance(i_test, list):
            i_test = [i_test]

        if args.llffhold > 0:
            print('Auto LLFF holdout,', args.llffhold)
            i_test = np.arange(images.shape[0])[::args.llffhold]

        i_val = i_test
        i_train = np.array([i for i in np.arange(int(images.shape[0])) if
                        (i not in i_test and i not in i_val)])

        print('DEFINING BOUNDS')
        if args.no_ndc:
            near = np.ndarray.min(bds) * .9
            far = np.ndarray.max(bds) * 1.
            
        else:
            near = 0.
            far = 1.
        print('NEAR FAR', near, far)

    elif args.dataset_type == 'blender':
        images, poses, render_poses, hwf, i_split = load_blender_data(args.datadir, args.half_res, args.testskip)
        print('Loaded blender', images.shape, render_poses.shape, hwf, args.datadir)
        i_train, i_val, i_test = i_split

        near = 2.
        far = 6.

        if args.white_bkgd:
            images = images[...,:3]*images[...,-1:] + (1.-images[...,-1:])
        else:
            images = images[...,:3]

    elif args.dataset_type == 'LINEMOD':
        images, poses, render_poses, hwf, K, i_split, near, far = load_LINEMOD_data(args.datadir, args.half_res, args.testskip)
        print(f'Loaded LINEMOD, images shape: {images.shape}, hwf: {hwf}, K: {K}')
        print(f'[CHECK HERE] near: {near}, far: {far}.')
        i_train, i_val, i_test = i_split

        if args.white_bkgd:
            images = images[...,:3]*images[...,-1:] + (1.-images[...,-1:])
        else:
            images = images[...,:3]

    elif args.dataset_type == 'deepvoxels':

        images, poses, render_poses, hwf, i_split = load_dv_data(scene=args.shape,
                                                                 basedir=args.datadir,
                                                                 testskip=args.testskip)

        print('Loaded deepvoxels', images.shape, render_poses.shape, hwf, args.datadir)
        i_train, i_val, i_test = i_split

        hemi_R = np.mean(np.linalg.norm(poses[:,:3,-1], axis=-1))
        near = hemi_R-1.
        far = hemi_R+1.

    else:
        print('Unknown dataset type', args.dataset_type, 'exiting')
        return

    # Cast intrinsics to right types
    H, W, focal = hwf
    H, W = int(H), int(W)
    hwf = [H, W, focal]

    if K is None:
        K = np.array([
            [focal, 0, 0.5*W],
            [0, focal, 0.5*H],
            [0, 0, 1]
        ])

    if args.render_test:
        render_poses = np.array(poses[i_test])

    # Create log dir and copy the config file
    basedir = args.basedir
    expname = args.expname
    os.makedirs(os.path.join(basedir, expname), exist_ok=True)
    f = os.path.join(basedir, expname, 'args.txt')
    with open(f, 'w') as file:
        for arg in sorted(vars(args)):
            attr = getattr(args, arg)
            file.write('{} = {}\n'.format(arg, attr))
    if args.config is not None:
        f = os.path.join(basedir, expname, 'config.txt')
        with open(f, 'w') as file:
            file.write(open(args.config, 'r').read())

    # Create nerf model
    render_kwargs_train, render_kwargs_test, start, grad_vars, optimizer = create_nerf(args)
    global_step = start

    bds_dict = {
        'near' : near,
        'far' : far,
    }
    render_kwargs_train.update(bds_dict)
    render_kwargs_test.update(bds_dict)

    # Move testing data to GPU
    render_poses = torch.Tensor(render_poses).to(device)

    # Short circuit if only rendering out from trained model
    if args.render_only:
        print('RENDER ONLY')
        with torch.no_grad():
            if args.render_test:
                # render_test switches to test poses
                images = images[i_test]
            else:
                # Default is smoother render_poses path
                images = None

            testsavedir = os.path.join(basedir, expname, 'renderonly_{}_{:06d}'.format('test' if args.render_test else 'path', start))
            os.makedirs(testsavedir, exist_ok=True)
            print('test poses shape', render_poses.shape)

            rgbs, _ = render_path(render_poses, hwf, K, args.chunk, render_kwargs_test, gt_imgs=images, savedir=testsavedir, render_factor=args.render_factor)
            print('Done rendering', testsavedir)
            imageio.mimwrite(os.path.join(testsavedir, 'video.mp4'), to8b(rgbs), fps=30, quality=8)

            return

    # Prepare raybatch tensor if batching random rays
    N_rand = args.N_rand
    use_batching = not args.no_batching
    if use_batching:
        # For random ray batching
        print('get rays')
        rays = np.stack([get_rays_np(H, W, K, p) for p in poses[:,:3,:4]], 0) # [N, ro+rd, H, W, 3]
        rays_radii = np.stack([get_rays_radii_np(H, W, K, p) for p in poses[:, :3, :4]], 0)

        print('done, concats')
        rays_rgb = np.concatenate([rays, images[:,None]], 1) # [N, ro+rd+rgb, H, W, 3]
        rays_rgb = np.transpose(rays_rgb, [0,2,3,1,4]) # [N, H, W, ro+rd+rgb, 3]
        rays_rgb = np.stack([rays_rgb[i] for i in i_train], 0) # train images only
        rays_radii = np.stack([rays_radii[i] for i in i_train], 0) # train images only
        
        rays_rgb = np.reshape(rays_rgb, [-1,3,3]) # [(N-1)*H*W, ro+rd+rgb, 3]
        rays_radii = np.reshape(rays_radii, [-1, 1])
        rays_rgb = rays_rgb.astype(np.float32)
        print('shuffle rays')
        
        rand_idx = np.random.permutation(rays_rgb.shape[0])
        rays_rgb = rays_rgb[rand_idx]
        rays_radii = rays_radii[rand_idx]

        print('done')
        i_batch = 0

    # Move training data to GPU
    if use_batching:
        images = torch.Tensor(images).to(device)
    poses = torch.Tensor(poses).to(device)
    if use_batching:
        rays_rgb = torch.Tensor(rays_rgb).to(device)
        rays_radii = torch.Tensor(rays_radii).to(device)


    N_iters = 200000 + 1
    print('Begin')
    print('TRAIN views are', i_train)
    print('TEST views are', i_test)
    print('VAL views are', i_val)

    # Summary writers
    # writer = SummaryWriter(os.path.join(basedir, 'summaries', expname))
    
    import torch.utils.tensorboard as tb
    import shutil
    if not args.no_logger:
        log_path = os.path.join(basedir, expname)
        os.makedirs(log_path, exist_ok=True)
        shutil.rmtree(os.path.join(log_path, 'train'), ignore_errors=True)
        logger = tb.SummaryWriter(os.path.join(log_path, 'train'), flush_secs=1)
    else: 
        logger = None
        
    start = start + 1
    for i in trange(start, N_iters):
        time0 = time.time()

        # Sample random ray batch
        if use_batching:
            # Random over all images
            batch = rays_rgb[i_batch:i_batch+N_rand] # [B, 2+1, 3*?]
            batch_radii = rays_radii[i_batch:i_batch+N_rand]
            batch = torch.transpose(batch, 0, 1)
            batch_rays, target_s = batch[:2], batch[2]

            i_batch += N_rand
            if i_batch >= rays_rgb.shape[0]:
                print("Shuffle data after an epoch!")
                rand_idx = torch.randperm(rays_rgb.shape[0])
                rays_rgb = rays_rgb[rand_idx]
                i_batch = 0

        #####  Core optimization loop  #####
        rgb, disp, acc, extras = render(H, W, K, chunk=args.chunk, rays=batch_rays, batch_radii=batch_radii,
                                                verbose=i < 10, retraw=True,
                                                **render_kwargs_train)

        optimizer.zero_grad()
        img_loss = img2mse(rgb, target_s)
        trans = extras['raw'][...,-1]
        loss = img_loss
        psnr = mse2psnr(img_loss)

        if 'rgb0' in extras:
            img_loss0 = img2mse(extras['rgb0'], target_s)
            loss = loss + 0.1 * img_loss0
            psnr0 = mse2psnr(img_loss0)

        loss.backward()
        optimizer.step()

        if logger is not None:
            logger.add_scalar('train/loss', float(loss.detach().cpu().numpy()), global_step=i)
            logger.add_scalar('train/coarse_psnr', float(psnr0), global_step=i)
            logger.add_scalar('train/fine_psnr', float(psnr), global_step=i)
            logger.add_scalar('train/lr', float(optimizer.state_dict()['param_groups'][0]['lr']), global_step=i)


        # NOTE: IMPORTANT!
        ###   update learning rate   ###
        decay_rate = 0.1
        decay_steps = args.lrate_decay * 1000
        new_lrate = args.lrate * (decay_rate ** (global_step / decay_steps))
        for param_group in optimizer.param_groups:
            param_group['lr'] = new_lrate
        ################################

        dt = time.time()-time0
        #####           end            #####

        # Rest is logging
        if i%args.i_weights==0:
            path = os.path.join(basedir, expname, '{:06d}.tar'.format(i))
            torch.save({
                'global_step': global_step,
                'network_fn_state_dict': render_kwargs_train['network_fn'].state_dict(),
                'network_fine_state_dict': render_kwargs_train['network_fine'].state_dict(),
                'optimizer_state_dict': optimizer.state_dict(),
            }, path)
            print('Saved checkpoints at', path)

        if i%args.i_video==0 and i > 0:
            # Turn on testing mode
            with torch.no_grad():
                rgbs, disps = render_path(render_poses[:5], hwf, K, args.chunk, render_kwargs_test)
            print('Done, saving', rgbs.shape, disps.shape)
            moviebase = os.path.join(basedir, expname, '{}_spiral_{:06d}_'.format(expname, i))
            imageio.mimwrite(moviebase + 'rgb.mp4', to8b(rgbs), fps=30, quality=8)
            imageio.mimwrite(moviebase + 'disp.mp4', to8b(disps / np.max(disps)), fps=30, quality=8)

        if i%args.i_testset==0 and i > 0:
            testsavedir = os.path.join(basedir, expname, 'testset_{:06d}'.format(i))
            os.makedirs(testsavedir, exist_ok=True)
            print('test poses shape', poses[i_test].shape)
            with torch.no_grad():
                render_path(torch.Tensor(poses[i_test]).to(device), hwf, K, args.chunk, render_kwargs_test, gt_imgs=images[i_test], savedir=testsavedir)
            print('Saved test set')


    
        if i%args.i_print==0:
            tqdm.write(f"[TRAIN] Iter: {i} Loss: {loss.item()}  PSNR: {psnr.item()}")

        global_step += 1


if __name__=='__main__':
    torch.set_default_tensor_type('torch.cuda.FloatTensor')

    train()