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 torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm, trange

import matplotlib.pyplot as plt

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 skimage.metrics import structural_similarity as SSIM
import lpips

loss_fn_alex = lpips.LPIPS(net='alex')
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):
        return torch.cat([fn(inputs[i:i+chunk]) for i in range(0, inputs.shape[0], chunk)], 0)
    return ret


def run_network(inputs, viewdirs, fn, netchunk=1024*64):
    """Prepares inputs and applies network 'fn'.
    """
    inputs_flat = torch.reshape(inputs, [-1, inputs.shape[-1]])

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

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


def batchify_rays(rays_flat, 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], **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, focal, chunk=1024*32, rays=None, c2w=None, ndc=True,
                  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, focal, c2w)
    else:
        # use provided ray batch
        rays_o, rays_d = rays
        ##print('here2')

    if use_viewdirs:
        # provide ray directions as input
        viewdirs = rays_d
        if c2w_staticcam is not None:
            # special case to visualize effect of viewdirs
            rays_o, rays_d = get_rays(H, W, focal, c2w_staticcam)
        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, focal, 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, 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', 'depth_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, 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 = []

    for i, c2w in enumerate(tqdm(render_poses, ncols=60)):
        rgb, disp, acc, _ = render(H, W, focal, chunk=chunk, c2w=c2w[:3,:4], **render_kwargs)
        rgbs.append(rgb.cpu().numpy())
        disps.append(disp.cpu().numpy())

        if savedir is not None:
            rgb8 = to8b(rgbs[-1])
            filename = os.path.join(savedir, '{:03d}.png'.format(i))
            imageio.imwrite(filename, rgb8)


    rgbs = np.stack(rgbs, 0)
    disps = np.stack(disps, 0)

    return rgbs, disps


def create_nerf(args):
    """Instantiate NeRF's MLP model.
    """
    skips = [4]
    if args.i_embed == 0:
        print('Positional embedding: Yes.')
    else:
        print('Positional embedding: No.')
        
    if args.use_viewdirs:
        print('View dependency: Yes.')
        model = NeRF(embed=(args.i_embed == 0), L_x=args.multires, L_d=args.multires_views).to(device)
    else:
        print('View dependency: No.')
        model = NeRF(embed=(args.i_embed == 0), L_x=args.multires, L_d=args.multires_views, inc_d=0).to(device)
    grad_vars = list(model.parameters())

    model_fine = None
    if args.N_importance > 0:
        print('Hierarchical sampling: Yes.')
        if args.use_viewdirs:
            model_fine = NeRF(embed=(args.i_embed == 0), L_x=args.multires, L_d=args.multires_views).to(device)
        else:
            model_fine = NeRF(embed=(args.i_embed == 0), L_x=args.multires, L_d=args.multires_views, inc_d=0).to(device)
        
        grad_vars += list(model_fine.parameters())
    else:
        print('Hierarchical sampling: No.')
    print('N_c, N_f: (%d, %d)'%(args.N_samples, args.N_importance))
        
    if args.i_embed == 0 and args.use_viewdirs and args.N_importance > 0:
        print('Complete model: Yes.')
    else:
        print('Complete model: No (ablation study).')

    network_query_fn = lambda inputs, viewdirs, network_fn : run_network(inputs, viewdirs, network_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:
            if 'network_fine_state_dict' not in ckpt:
                print('Error! The fine model is not recorded. Use the coarse model.')
                model_fine.load_state_dict(ckpt['network_fn_state_dict'])
            else:
                print('Use the fine model as well.')
                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,
                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]

    # batch (R, 2=o+d, 3)
    N_samples = 64
    ts = torch.linspace(0.0, 1.0, N_samples+1, device=ray_batch.device) # (S+1)

    # jittering
    deltas = torch.rand(N_rays, len(ts), device=ray_batch.device) / N_samples
    ts = ts.expand(N_rays, -1)
    ts = ts + deltas # (R, S+1)
    ts = near + ts * (far - near)

    # position sampling for stochastic quadrature
    o, d = rays_o, rays_d # (R, 3)
    x = o[...,None] + ts[:,None,:] * d[...,None] # (R, 3, S+1)
    if viewdirs is not None:
        viewdirs = viewdirs[...,None].expand(viewdirs.shape[0], viewdirs.shape[1], x.shape[-1]) # (R, 3, S+1)

    # evaluating the neural radiance field
    # x: (B*S, 3+3)
    x = x.permute(0,2,1).reshape(-1, x.shape[1])
    if viewdirs is not None:
        viewdirs = viewdirs.permute(0,2,1).reshape(-1, viewdirs.shape[1])
        ipt = torch.cat([x, viewdirs], -1) # (R*(S+1), 3 + 3)
    else:
        ipt = x
        
    out = network_fn(ipt) # (R*(S+1), 3 + 1)
    color = torch.sigmoid(out[:,:3]).reshape(N_rays, -1, 3).permute(0,2,1)
    sigma = out[:,3:4]
    if raw_noise_std > 0.:
        noise = torch.randn(sigma.shape) * raw_noise_std
        sigma += noise
    alpha = F.relu(sigma).reshape(N_rays, -1, 1).permute(0,2,1)

    # rendering
    deltas = ts[:,1:] - ts[:,:-1] # (R, S)
    deltas = deltas * torch.norm(rays_d[...,None,:], dim=-1)
    weights = deltas * alpha[:,0,:-1] # (R, S)
    weights = torch.cumsum(weights, dim=1)
    weights = torch.cat([torch.zeros(ts.shape[0],1,device=weights.device), weights], 1) # (R, S+1)
    weights = torch.exp(-weights)
    weights = weights[:,:-1] - weights[:,1:] # (R, S)
    out = torch.sum(weights[:,None,:] * color[...,:-1], -1) # (R, 3)

    # white background
    if white_bkgd:
        out += (1 - torch.sum(weights, -1)[...,None])
    rgb_map = out
    
    depth_map = torch.sum(weights * ts[...,:-1], -1)
    disp_map = 1 / torch.max(1e-10 * torch.ones_like(depth_map), depth_map / torch.sum(weights, -1))
    acc_map = torch.sum(weights, -1)

    if N_importance > 0:

        rgb_map_0, depth_map_0, acc_map_0 = rgb_map, depth_map, acc_map

        z_vals_mid = .5 * (ts[...,1:] + ts[...,:-1]) # (R, S)
        z_samples = sample_pdf(z_vals_mid, weights[...,1:], N_importance)
        z_samples = z_samples.detach() # (R, N_importance)

        ts, _ = torch.sort(torch.cat([ts, z_samples], -1), -1) # (R, S+1+N_importance)
        
        x = o[...,None] + ts[:,None,:] * d[...,None] # (R, 3, S+1)
        viewdirs = ray_batch[:,-3:] if ray_batch.shape[-1] > 8 else None
        if viewdirs is not None:
            viewdirs = viewdirs[...,None].expand(viewdirs.shape[0], viewdirs.shape[1], x.shape[-1])

        # evaluating the neural radiance field
        # x: (B*S, 3+3)
        x = x.permute(0,2,1).reshape(-1, x.shape[1])
        if viewdirs is not None:
            viewdirs = viewdirs.permute(0,2,1).reshape(-1, viewdirs.shape[1])
            ipt = torch.cat([x, viewdirs], -1) # (R*(S+1), 3 + 3)
        else:
            ipt = x
        out = network_fine(ipt) # (R*(S+1), 3 + 1)
        color = torch.sigmoid(out[:,:3]).reshape(N_rays, -1, 3).permute(0,2,1)
        sigma = out[:,3:4]
        if raw_noise_std > 0.:
            noise = torch.randn(sigma.shape) * raw_noise_std
            sigma += noise
        alpha = F.relu(sigma).reshape(N_rays, -1, 1).permute(0,2,1)

        # rendering
        deltas = ts[:,1:] - ts[:,:-1] # (R, S)
        deltas = deltas * torch.norm(rays_d[...,None,:], dim=-1)
        weights = deltas * alpha[:,0,:-1]
        weights = torch.cumsum(weights, dim=1)
        weights = torch.cat([torch.zeros(ts.shape[0],1,device=weights.device), weights], 1) # (R, S+1)
        weights = torch.exp(-weights)
        weights = weights[:,:-1] - weights[:,1:] # (R, S)
        out = torch.sum(weights[:,None,:] * color[...,:-1], -1) # (R, 3)

        # white background
        if white_bkgd:
            out += (1 - torch.sum(weights, -1)[...,None])
        rgb_map = out
        
        depth_map = torch.sum(weights * ts[...,:-1], -1)
        disp_map = 1 / torch.max(1e-10 * torch.ones_like(depth_map), depth_map / torch.sum(weights, -1))
        acc_map = torch.sum(weights, -1)

    ret = {'rgb_map' : rgb_map, 'depth_map' : depth_map, 'acc_map' : acc_map}
    
    if N_importance > 0:
        ret['rgb0'] = rgb_map_0
        ret['depth0'] = depth_map_0
        ret['acc0'] = acc_map_0
        ret['z_std'] = torch.std(z_samples, dim=-1, unbiased=False)  # [N_rays]

    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 config_parser():

    import configargparse
    parser = configargparse.ArgumentParser()
    parser.add_argument('--config', is_config_file=True, 
                        help='config file path')
    parser.add_argument("--expname", type=str, 
                        help='experiment name')
    parser.add_argument("--basedir", type=str, default='./logs/', 
                        help='where to store ckpts and logs')
    parser.add_argument("--datadir", type=str, default='./data/llff/fern', 
                        help='input data directory')

    # training options
    parser.add_argument("--netdepth", type=int, default=8, 
                        help='layers in network')
    parser.add_argument("--netwidth", type=int, default=256, 
                        help='channels per layer')
    parser.add_argument("--netdepth_fine", type=int, default=8, 
                        help='layers in fine network')
    parser.add_argument("--netwidth_fine", type=int, default=256, 
                        help='channels per layer in fine network')
    parser.add_argument("--N_rand", type=int, default=32*32*4, 
                        help='batch size (number of random rays per gradient step)')
    parser.add_argument("--lrate", type=float, default=5e-4, 
                        help='learning rate')
    parser.add_argument("--lrate_decay", type=int, default=250, 
                        help='exponential learning rate decay (in 1000 steps)')
    parser.add_argument("--chunk", type=int, default=1024*32, 
                        help='number of rays processed in parallel, decrease if running out of memory')
    parser.add_argument("--netchunk", type=int, default=1024*64, 
                        help='number of pts sent through network in parallel, decrease if running out of memory')
    parser.add_argument("--no_batching", action='store_true', 
                        help='only take random rays from 1 image at a time')
    parser.add_argument("--no_reload", action='store_true', 
                        help='do not reload weights from saved ckpt')
    parser.add_argument("--ft_path", type=str, default=None, 
                        help='specific weights npy file to reload for coarse network')

    # rendering options
    parser.add_argument("--N_samples", type=int, default=64, 
                        help='number of coarse samples per ray')
    parser.add_argument("--N_importance", type=int, default=0,
                        help='number of additional fine samples per ray')
    parser.add_argument("--perturb", type=float, default=1.,
                        help='set to 0. for no jitter, 1. for jitter')
    parser.add_argument("--use_viewdirs", action='store_true', 
                        help='use full 5D input instead of 3D')
    parser.add_argument("--i_embed", type=int, default=0, 
                        help='set 0 for default positional encoding, -1 for none')
    parser.add_argument("--multires", type=int, default=10, 
                        help='log2 of max freq for positional encoding (3D location)')
    parser.add_argument("--multires_views", type=int, default=4, 
                        help='log2 of max freq for positional encoding (2D direction)')
    parser.add_argument("--raw_noise_std", type=float, default=0., 
                        help='std dev of noise added to regularize sigma_a output, 1e0 recommended')

    parser.add_argument("--render_only", action='store_true', 
                        help='do not optimize, reload weights and render out render_poses path')
    parser.add_argument("--render_test", action='store_true', 
                        help='render the test set instead of render_poses path')
    parser.add_argument("--render_factor", type=int, default=0, 
                        help='downsampling factor to speed up rendering, set 4 or 8 for fast preview')

    # training options
    parser.add_argument("--precrop_iters", type=int, default=0,
                        help='number of steps to train on central crops')
    parser.add_argument("--precrop_frac", type=float,
                        default=.5, help='fraction of img taken for central crops') 

    # dataset options
    parser.add_argument("--dataset_type", type=str, default='llff', 
                        help='options: llff / blender / deepvoxels')
    parser.add_argument("--testskip", type=int, default=8, 
                        help='will load 1/N images from test/val sets, useful for large datasets like deepvoxels')

    ## deepvoxels flags
    parser.add_argument("--shape", type=str, default='greek', 
                        help='options : armchair / cube / greek / vase')

    ## blender flags
    parser.add_argument("--white_bkgd", action='store_true', 
                        help='set to render synthetic data on a white bkgd (always use for dvoxels)')
    parser.add_argument("--half_res", action='store_true', 
                        help='load blender synthetic data at 400x400 instead of 800x800')

    ## llff flags
    parser.add_argument("--factor", type=int, default=8, 
                        help='downsample factor for LLFF images')
    parser.add_argument("--no_ndc", action='store_true', 
                        help='do not use normalized device coordinates (set for non-forward facing scenes)')
    parser.add_argument("--lindisp", action='store_true', 
                        help='sampling linearly in disparity rather than depth')
    parser.add_argument("--spherify", action='store_true', 
                        help='set for spherical 360 scenes')
    parser.add_argument("--llffhold", type=int, default=8, 
                        help='will take every 1/N images as LLFF test set, paper uses 8')

    # logging/saving options
    parser.add_argument("--i_print",   type=int, default=500, 
                        help='frequency of console printout and metric loggin')
    parser.add_argument("--i_img",     type=int, default=500, 
                        help='frequency of tensorboard image logging')
    parser.add_argument("--i_weights", type=int, default=25000, 
                        help='frequency of weight ckpt saving')
    parser.add_argument("--i_testset", type=int, default=100000, 
                        help='frequency of testset saving')
    parser.add_argument("--i_video",   type=int, default=25000, 
                        help='frequency of render_poses video saving')

    return parser


def print_error(images, i_test, args, start=199999):
    basedir = args.basedir
    expname = args.expname
    images = images[i_test]
    print('image size: ', images.shape)

    testsavedir = os.path.join(basedir, expname, 'renderonly_{}_{:06d}'.format('test' if args.render_test else 'path', start))
    os.makedirs(testsavedir, exist_ok=True)
    
    rgbs = np.zeros((200, 800, 800, 3))
    for idx in trange(200):
        rgbs[idx] = imageio.imread(os.path.join(testsavedir, '0'*(3-len(str(idx)))+str(idx)+'.png'))
    rgbs = rgbs.astype(float) / 255.0

    psnr = -10.0 * np.log(((rgbs - images) ** 2).mean()) / np.log(np.array([10.]))
    ssim = np.array([SSIM(rgbs[i], images[i], channel_axis=2, multichannel=True) for i in range(len(rgbs))]).mean()
    lpips = loss_fn_alex(torch.from_numpy(rgbs[:len(rgbs)//2]).float().permute(0,3,1,2), torch.from_numpy(images[:len(rgbs)//2]).float().permute(0,3,1,2))
    lpips += loss_fn_alex(torch.from_numpy(rgbs[len(rgbs)//2:]).float().permute(0,3,1,2), torch.from_numpy(images[len(rgbs)//2:]).float().permute(0,3,1,2))
    lpips /= 2.0
    
    print('%%%%%%%%%%%%%%%%')
    print(args.expname)
    print('PSNR: ', psnr.item())
    print('SSIM: ', ssim)
    print('LPIPS: ', lpips.mean().item())
    print('%%%%%%%%%%%%%%%%')

    print('Done rendering', testsavedir)
    imageio.mimwrite(os.path.join(testsavedir, 'video.mp4'), to8b(rgbs), fps=30, quality=8)


def train():

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

    # Load data
    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 == '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 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())
    
    #print_error(images, i_test, args, start=199999)
    #return

    # 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]
                print('image size: ', images.shape)
            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)
            print('render factor', args.render_factor)

            rgbs, _ = render_path(render_poses, hwf, args.chunk, render_kwargs_test, gt_imgs=images, savedir=testsavedir, render_factor=args.render_factor)
            
            psnr = -10.0 * np.log(((rgbs - images) ** 2).mean()) / np.log(np.array([10.]))
            ssim = np.array([SSIM(rgbs[i], images[i], channel_axis=2, multichannel=True) for i in range(len(rgbs))]).mean()
            lpips = loss_fn_alex(torch.from_numpy(rgbs).float().permute(0,3,1,2), torch.from_numpy(images).float().permute(0,3,1,2))
            
            print('%%%%%%%%%%%%%%%%')
            print(args.expname)
            print('PSNR: ', psnr.item())
            print('SSIM: ', ssim)
            print('LPIPS: ', lpips.mean().item())
            print('%%%%%%%%%%%%%%%%')
            
            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, focal, p) for p in poses[:,:3,:4]], 0) # [N, ro+rd, H, W, 3]
        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_rgb = np.reshape(rays_rgb, [-1,3,3]) # [(N-1)*H*W, ro+rd+rgb, 3]
        rays_rgb = rays_rgb.astype(np.float32)
        print('shuffle rays')
        np.random.shuffle(rays_rgb)

        print('done')
        i_batch = 0

    # Move training data to GPU
    images = torch.Tensor(images).to(device)
    poses = torch.Tensor(poses).to(device)
    if use_batching:
        rays_rgb = torch.Tensor(rays_rgb).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)
    
    if args.precrop_iters > 0:
        dH = int(H//2 * args.precrop_frac)
        dW = int(W//2 * args.precrop_frac)
        print(f"[Config] Center cropping of size {2*dH} x {2*dW} is enabled until iter {args.precrop_iters}")
    
    start = start + 1
    for i in trange(start, N_iters, ncols=60):
        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 = 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

        else:
            # Random from one image
            img_i = np.random.choice(i_train)
            target = images[img_i]
            pose = poses[img_i, :3,:4]

            if N_rand is not None:
                rays_o, rays_d = get_rays(H, W, focal, torch.Tensor(pose))  # (H, W, 3), (H, W, 3)

                if i < args.precrop_iters:
                    dH = int(H//2 * args.precrop_frac)
                    dW = int(W//2 * args.precrop_frac)
                    coords = torch.stack(
                        torch.meshgrid(
                            torch.linspace(H//2 - dH, H//2 + dH - 1, 2*dH), 
                            torch.linspace(W//2 - dW, W//2 + dW - 1, 2*dW)
                        ), -1)
                else:
                    coords = torch.stack(torch.meshgrid(torch.linspace(0, H-1, H), torch.linspace(0, W-1, W)), -1)  # (H, W, 2)

                coords = torch.reshape(coords, [-1,2])  # (H * W, 2)
                select_inds = np.random.choice(coords.shape[0], size=[N_rand], replace=False)  # (N_rand,)
                select_coords = coords[select_inds].long()  # (N_rand, 2)
                rays_o = rays_o[select_coords[:, 0], select_coords[:, 1]]  # (N_rand, 3)
                rays_d = rays_d[select_coords[:, 0], select_coords[:, 1]]  # (N_rand, 3)
                batch_rays = torch.stack([rays_o, rays_d], 0)
                target_s = target[select_coords[:, 0], select_coords[:, 1]]  # (N_rand, 3)

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

        optimizer.zero_grad()
        img_loss = img2mse(rgb, target_s)
        loss = img_loss
        psnr = mse2psnr(img_loss)

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

        loss.backward()
        optimizer.step()

        # 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

        # Rest is logging
        if i%args.i_print==0:
            tqdm.write(f"[TRAIN] Iter: {i} Loss: {img_loss.item()}  PSNR: {psnr.item()} (Coarse)")
            if 'rgb0' in extras:
                tqdm.write(f"[TRAIN] Iter: {i} Loss: {img_loss0.item()}  PSNR: {psnr0.item()} (Fine)")
            
        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, hwf, 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 + 'depth.mp4', to8b(disps / np.max(disps)), fps=30, quality=8)
            
        global_step += 1


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