train.py

import os
import numpy as np
import random

import time

from absl import app
import gin
from internal import configs
from internal import datasets
from internal import image
from internal import models
from internal import train_utils
from internal import utils
from internal import vis
from internal import checkpoints
import torch
import accelerate
import tensorboardX
from tqdm import tqdm
from torch.utils._pytree import tree_map

configs.define_common_flags()

TIME_PRECISION = 1000  # Internally represent integer times in milliseconds.


def main(unused_argv):
    config = configs.load_config()

    config.exp_path = os.path.join("exp", config.exp_name)
    utils.makedirs(config.exp_path)
    with utils.open_file(os.path.join(config.exp_path, 'config.gin'), 'w') as f:
        f.write(gin.config_str())

    # accelerator for DDP
    accelerator = accelerate.Accelerator()
    config.world_size = accelerator.num_processes
    config.local_rank = accelerator.local_process_index
    if config.batch_size % accelerator.num_processes != 0:
        config.batch_size -= config.batch_size % accelerator.num_processes != 0
        accelerator.print('turn batch size to', config.batch_size)
    # Shift random seed by local_process_index to shuffle data loaded by different hosts.
    utils.seed_everything(config.seed + accelerator.local_process_index)

    # setup model and optimizer
    model = models.Model(config=config)
    optimizer, lr_fn = train_utils.create_optimizer(config, model)
    init_step = checkpoints.restore_checkpoint(config.exp_path, model, optimizer) + 1
    # model = torch.compile(model)  # not work yet

    # load dataset
    dataset = datasets.load_dataset('train', config.data_dir, config)
    config.factor = 2  # for efficiency downsample 8x for test image while training
    test_dataset = datasets.load_dataset('test', config.data_dir, config)
    dataloader = torch.utils.data.DataLoader(np.arange(len(dataset)),
                                             num_workers=8,
                                             shuffle=True,
                                             batch_size=1,
                                             collate_fn=dataset.collate_fn,
                                             persistent_workers=True,
                                             )
    test_dataloader = torch.utils.data.DataLoader(np.arange(len(test_dataset)),
                                                  num_workers=4,
                                                  shuffle=False,
                                                  batch_size=1,
                                                  persistent_workers=True,
                                                  collate_fn=test_dataset.collate_fn,
                                                  )
    if config.rawnerf_mode:
        postprocess_fn = test_dataset.metadata['postprocess_fn']
    else:
        postprocess_fn = lambda z, _=None: z

    # use accelerate to prepare.
    # no need to prepare dataloader because data in each process are loaded differently
    model, optimizer = accelerator.prepare(model, optimizer)
    model.eval()
    module = accelerator.unwrap_model(model)
    dataiter = iter(dataloader)
    test_dataiter = iter(test_dataloader)

    num_params = train_utils.tree_len(list(model.parameters()))
    accelerator.print(f'Number of parameters being optimized: {num_params}')

    if (dataset.size > module.num_glo_embeddings and module.num_glo_features > 0):
        raise ValueError(f'Number of glo embeddings {module.num_glo_embeddings} '
                         f'must be at least equal to number of train images '
                         f'{dataset.size}')

    # metric handler
    metric_harness = image.MetricHarness()

    # tensorboard
    if accelerator.is_local_main_process:
        summary_writer = tensorboardX.SummaryWriter(config.exp_path)
        # function to convert image for tensorboard
        tb_process_fn = lambda x: x.transpose(2, 0, 1) if len(x.shape) == 3 else x[None]

        if config.rawnerf_mode:
            for name, data in zip(['train', 'test'], [dataset, test_dataset]):
                # Log shutter speed metadata in TensorBoard for debug purposes.
                for key in ['exposure_idx', 'exposure_values', 'unique_shutters']:
                    summary_writer.add_text(f'{name}_{key}', str(data.metadata[key]), 0)
    accelerator.print("Begin training...")
    total_time = 0
    total_steps = 0
    reset_stats = True
    if config.early_exit_steps is not None:
        num_steps = config.early_exit_steps
    else:
        num_steps = config.max_steps

    if accelerator.is_local_main_process:
        tbar = tqdm(range(init_step, num_steps + 1))
    else:
        tbar = range(init_step, num_steps + 1)
    for step in tbar:
        try:
            batch = next(dataiter)
        except StopIteration:
            dataiter = iter(dataloader)
            batch = next(dataiter)
        batch = accelerate.utils.send_to_device(batch, accelerator.device)
        if reset_stats and accelerator.is_local_main_process:
            stats_buffer = []
            train_start_time = time.time()
            reset_stats = False

        # use lr_fn to control learning rate
        learning_rate = lr_fn(step)
        for param_group in optimizer.param_groups:
            param_group['lr'] = learning_rate

        # fraction of training period
        train_frac = np.clip((step - 1) / (config.max_steps - 1), 0, 1)

        # Indicates whether we need to compute output normal or depth maps in 2D.
        compute_extras = (config.compute_disp_metrics or config.compute_normal_metrics)
        optimizer.zero_grad()
        with accelerator.autocast():
            renderings, ray_history = model(
                True,
                batch,
                train_frac=train_frac,
                compute_extras=compute_extras,
                zero_glo=False)

        losses = {}

        # supervised by data
        data_loss, stats = train_utils.compute_data_loss(batch, renderings, config)
        losses['data'] = data_loss

        # interlevel loss in MipNeRF360
        if config.interlevel_loss_mult > 0 and not module.single_mlp:
            losses['interlevel'] = train_utils.interlevel_loss(ray_history, config)

        if config.distortion_loss_mult > 0:
            losses['distortion'] = train_utils.distortion_loss(ray_history, config)

        # orientation loss in RefNeRF
        if (config.orientation_coarse_loss_mult > 0 or
                config.orientation_loss_mult > 0):
            losses['orientation'] = train_utils.orientation_loss(batch, module, ray_history,
                                                                 config)

        # normal supervision loss in RefNeRF
        if (config.predicted_normal_coarse_loss_mult > 0 or
                config.predicted_normal_loss_mult > 0):
            losses['predicted_normals'] = train_utils.predicted_normal_loss(
                module, ray_history, config)
        loss = sum(losses.values())
        stats['loss'] = loss.item()
        stats['losses'] = tree_map(lambda x: x.item(), losses)

        # accelerator automatically handle the scale
        accelerator.backward(loss)
        # clip gradient by max/norm/nan
        train_utils.clip_gradients(model, accelerator, config)
        optimizer.step()

        stats['psnrs'] = image.mse_to_psnr(stats['mses'])
        stats['psnr'] = stats['psnrs'][-1]

        # Log training summaries. This is put behind a host_id check because in
        # multi-host evaluation, all hosts need to run inference even though we
        # only use host 0 to record results.
        if accelerator.is_local_main_process:
            stats_buffer.append(stats)
            if step == init_step or step % config.print_every == 0:
                elapsed_time = time.time() - train_start_time
                steps_per_sec = config.print_every / elapsed_time
                rays_per_sec = config.batch_size * steps_per_sec

                # A robust approximation of total training time, in case of pre-emption.
                total_time += int(round(TIME_PRECISION * elapsed_time))
                total_steps += config.print_every
                approx_total_time = int(round(step * total_time / total_steps))

                # Transpose and stack stats_buffer along axis 0.

                fs = [utils.flatten_dict(s, sep='/') for s in stats_buffer]
                stats_stacked = {k: np.stack([f[k] for f in fs]) for k in fs[0].keys()}

                # Split every statistic that isn't a vector into a set of statistics.
                stats_split = {}
                for k, v in stats_stacked.items():
                    if v.ndim not in [1, 2] and v.shape[0] != len(stats_buffer):
                        raise ValueError('statistics must be of size [n], or [n, k].')
                    if v.ndim == 1:
                        stats_split[k] = v
                    elif v.ndim == 2:
                        for i, vi in enumerate(tuple(v.T)):
                            stats_split[f'{k}/{i}'] = vi

                # Summarize the entire histogram of each statistic.
                for k, v in stats_split.items():
                    summary_writer.add_histogram('train_' + k, v, step)

                # Take the mean and max of each statistic since the last summary.
                avg_stats = {k: np.mean(v) for k, v in stats_split.items()}
                max_stats = {k: np.max(v) for k, v in stats_split.items()}

                summ_fn = lambda s, v: summary_writer.add_scalar(s, v, step)  # pylint:disable=cell-var-from-loop

                # Summarize the mean and max of each statistic.
                for k, v in avg_stats.items():
                    summ_fn(f'train_avg_{k}', v)
                for k, v in max_stats.items():
                    summ_fn(f'train_max_{k}', v)

                summ_fn('train_num_params', num_params)
                summ_fn('train_learning_rate', learning_rate)
                summ_fn('train_steps_per_sec', steps_per_sec)
                summ_fn('train_rays_per_sec', rays_per_sec)

                summary_writer.add_scalar('train_avg_psnr_timed', avg_stats['psnr'],
                                          total_time // TIME_PRECISION)
                summary_writer.add_scalar('train_avg_psnr_timed_approx', avg_stats['psnr'],
                                          approx_total_time // TIME_PRECISION)

                if dataset.metadata is not None and module.learned_exposure_scaling:
                    scalings = module.exposure_scaling_offsets.weight
                    num_shutter_speeds = dataset.metadata['unique_shutters'].shape[0]
                    for i_s in range(num_shutter_speeds):
                        for j_s, value in enumerate(scalings[i_s]):
                            summary_name = f'exposure/scaling_{i_s}_{j_s}'
                            summary_writer.add_scalar(summary_name, value, step)

                precision = int(np.ceil(np.log10(config.max_steps))) + 1
                avg_loss = avg_stats['loss']
                avg_psnr = avg_stats['psnr']
                str_losses = {  # Grab each "losses_{x}" field and print it as "x[:4]".
                    k[7:11]: (f'{v:0.5f}' if v >= 1e-4 and v < 10 else f'{v:0.1e}')
                    for k, v in avg_stats.items()
                    if k.startswith('losses/')
                }
                tbar.write(f'{step:{precision}d}' + f'/{config.max_steps:d}: ' +
                           f'loss={avg_loss:0.5f}, ' + f'psnr={avg_psnr:6.3f}, ' +
                           f'lr={learning_rate:0.2e} | ' +
                           ', '.join([f'{k}={s}' for k, s in str_losses.items()]) +
                           f', {rays_per_sec:0.0f} r/s')

                # Reset everything we are tracking between summarizations.
                reset_stats = True

            if step > 0 and step % config.checkpoint_every == 0:
                checkpoints.save_checkpoint(
                    config.exp_path,
                    {
                        'state_dict': accelerator.unwrap_model(model).state_dict(),
                        'optimizer': optimizer.state_dict(),
                    },
                    int(step), keep=1)

        # Test-set evaluation.
        if step == 1 or step % config.train_render_every == 0:
            # We reuse the same random number generator from the optimization step
            # here on purpose so that the visualization matches what happened in
            # training.
            eval_start_time = time.time()
            try:
                test_batch = next(test_dataiter)
            except StopIteration:
                test_dataiter = iter(test_dataloader)
                test_batch = next(test_dataiter)
            test_batch = accelerate.utils.send_to_device(test_batch, accelerator.device)

            # render a single image with all distributed processes
            rendering = models.render_image(
                lambda rand, x: model(rand,
                                      x,
                                      train_frac=train_frac,
                                      compute_extras=True),
                accelerator,
                test_batch, False, config)

            # move to numpy
            rendering = tree_map(lambda x: x.detach().cpu().numpy(), rendering)
            test_batch = tree_map(lambda x: x.detach().cpu().numpy() if x is not None else None, test_batch)
            # Log eval summaries on host 0.
            if accelerator.is_local_main_process:
                eval_time = time.time() - eval_start_time
                num_rays = np.prod(test_batch['directions'].shape[:-1])
                rays_per_sec = num_rays / eval_time
                summary_writer.add_scalar('test_rays_per_sec', rays_per_sec, step)
                print(f'Eval {step}: {eval_time:0.3f}s., {rays_per_sec:0.0f} rays/sec')

                metric_start_time = time.time()
                metric = metric_harness(
                    postprocess_fn(rendering['rgb']), postprocess_fn(test_batch['rgb']))
                print(f'Metrics computed in {(time.time() - metric_start_time):0.3f}s')
                for name, val in metric.items():
                    if not np.isnan(val):
                        print(f'{name} = {val:.4f}')
                        summary_writer.add_scalar('train_metrics/' + name, val, step)

                if config.vis_decimate > 1:
                    d = config.vis_decimate
                    decimate_fn = lambda x, d=d: None if x is None else x[::d, ::d]
                else:
                    decimate_fn = lambda x: x
                rendering = tree_map(decimate_fn, rendering)
                test_batch = tree_map(decimate_fn, test_batch)
                vis_start_time = time.time()
                vis_suite = vis.visualize_suite(rendering, test_batch)
                print(f'Visualized in {(time.time() - vis_start_time):0.3f}s')
                if config.rawnerf_mode:
                    # Unprocess raw output.
                    vis_suite['color_raw'] = rendering['rgb']
                    # Autoexposed colors.
                    vis_suite['color_auto'] = postprocess_fn(rendering['rgb'], None)
                    summary_writer.add_image('test_true_auto',
                                             tb_process_fn(postprocess_fn(test_batch['rgb'], None)), step)
                    # Exposure sweep colors.
                    exposures = test_dataset.metadata['exposure_levels']
                    for p, x in list(exposures.items()):
                        vis_suite[f'color/{p}'] = postprocess_fn(rendering['rgb'], x)
                        summary_writer.add_image(f'test_true_color/{p}',
                                                 tb_process_fn(postprocess_fn(test_batch['rgb'], x)), step)
                summary_writer.add_image('test_true_color', tb_process_fn(test_batch['rgb']), step)
                if config.compute_normal_metrics:
                    summary_writer.add_image('test_true_normals',
                                             tb_process_fn(test_batch['normals']) / 2. + 0.5, step)
                for k, v in vis_suite.items():
                    summary_writer.add_image('test_output_' + k, tb_process_fn(v), step)

    if accelerator.is_local_main_process and config.max_steps % config.checkpoint_every != 0:
        checkpoints.save_checkpoint(
            config.exp_path,
            {
                "state_dict": accelerator.unwrap_model(model).state_dict(),
                "optimizer": optimizer.state_dict(),
            },
            int(config.max_steps), keep=1)


if __name__ == '__main__':
    with gin.config_scope('train'):
        app.run(main)