run.py

# Copyright (c) 2018-present, Facebook, Inc.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
#

import numpy as np

from common.arguments import parse_args
import torch
import random

import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import os
import sys
import errno

from common.camera import *
from common.model import *
from common.loss import *
from common.generators import ChunkedGenerator, Evaluate_Generator
from time import time
from common.utils import deterministic_random

args = parse_args()
print(args)

seed = 4321
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)

try:
    # Create checkpoint directory if it does not exist
    os.makedirs(args.checkpoint)
except OSError as e:
    if e.errno != errno.EEXIST:
        raise RuntimeError('Unable to create checkpoint directory:', args.checkpoint)

print('Loading dataset...')
dataset_path = 'data/data_3d_' + args.dataset + '.npz'
if args.dataset == 'h36m':
    from common.h36m_dataset import Human36mDataset

    dataset = Human36mDataset(dataset_path)
elif args.dataset.startswith('humaneva'):
    from common.humaneva_dataset import HumanEvaDataset

    dataset = HumanEvaDataset(dataset_path)
else:
    raise KeyError('Invalid dataset')

print('Preparing data...')

# Restore 3d positions in anim
for subject in dataset.subjects():
    for action in dataset[subject].keys():
        anim = dataset[subject][action]

        if 'positions' in anim:
            positions_3d = []
            for cam in anim['cameras']:
                pos_3d = world_to_camera(anim['positions'], R=cam['orientation'], t=cam['translation'])
                pos_3d[:, 1:] -= pos_3d[:, :1]  # Remove global offset, but keep trajectory in first position
                positions_3d.append(pos_3d)
            anim['positions_3d'] = positions_3d

# Restore 2d positions in keypoints

print('Loading 2D detections...')
keypoints = np.load('data/data_2d_' + args.dataset + '_' + args.keypoints + '.npz', allow_pickle=True)
keypoints_metadata = keypoints['metadata'].item()
keypoints_symmetry = keypoints_metadata['keypoints_symmetry']
kps_left, kps_right = list(keypoints_symmetry[0]), list(keypoints_symmetry[1])
joints_left, joints_right = list(dataset.skeleton().joints_left()), list(dataset.skeleton().joints_right())
keypoints = keypoints['positions_2d'].item()

for subject in dataset.subjects():
    assert subject in keypoints, 'Subject {} is missing from the 2D detections dataset'.format(subject)
    for action in dataset[subject].keys():
        assert action in keypoints[subject], 'Action {} of subject {} is missing from the 2D detections dataset'.format(
            action, subject)
        if 'positions_3d' not in dataset[subject][action]:
            continue

        for cam_idx in range(len(keypoints[subject][action])):

            # We check for >= instead of == because some videos in H3.6M contain extra frames
            mocap_length = dataset[subject][action]['positions_3d'][cam_idx].shape[0]
            assert keypoints[subject][action][cam_idx].shape[0] >= mocap_length

            if keypoints[subject][action][cam_idx].shape[0] > mocap_length:
                # Shorten sequence
                keypoints[subject][action][cam_idx] = keypoints[subject][action][cam_idx][:mocap_length]

        assert len(keypoints[subject][action]) == len(dataset[subject][action]['positions_3d'])

for subject in keypoints.keys():
    for action in keypoints[subject]:
        for cam_idx, kps in enumerate(keypoints[subject][action]):
            # Normalize camera frame
            cam = dataset.cameras()[subject][cam_idx]
            kps[..., :2] = normalize_screen_coordinates(kps[..., :2], w=cam['res_w'], h=cam['res_h'])
            keypoints[subject][action][cam_idx] = kps

subjects_train = args.subjects_train.split(',')
if not args.render:
    subjects_test = args.subjects_test.split(',')
else:
    subjects_test = [args.viz_subject]


def fetch(subjects, action_filter=None, subset=1, parse_3d_poses=True):
    out_poses_3d = []
    out_poses_2d = []
    out_camera_params = []
    for subject in subjects:
        for action in keypoints[subject].keys():
            if action_filter is not None:
                found = False
                for a in action_filter:
                    if action.startswith(a):
                        found = True
                        break
                if not found:
                    continue

            poses_2d = keypoints[subject][action]
            for i in range(len(poses_2d)):  # Iterate across cameras
                out_poses_2d.append(poses_2d[i])

            if subject in dataset.cameras():
                cams = dataset.cameras()[subject]
                assert len(cams) == len(poses_2d), 'Camera count mismatch'
                for cam in cams:
                    if 'intrinsic' in cam:
                        out_camera_params.append(cam['intrinsic'])

            if parse_3d_poses and 'positions_3d' in dataset[subject][action]:
                poses_3d = dataset[subject][action]['positions_3d']
                assert len(poses_3d) == len(poses_2d), 'Camera count mismatch'
                for i in range(len(poses_3d)):  # Iterate across cameras
                    out_poses_3d.append(poses_3d[i])

    if len(out_camera_params) == 0:
        out_camera_params = None
    if len(out_poses_3d) == 0:
        out_poses_3d = None

    stride = args.downsample
    if subset < 1:
        for i in range(len(out_poses_2d)):
            n_frames = int(round(len(out_poses_2d[i]) // stride * subset) * stride)
            start = deterministic_random(0, len(out_poses_2d[i]) - n_frames + 1, str(len(out_poses_2d[i])))
            out_poses_2d[i] = out_poses_2d[i][start:start + n_frames:stride]
            if out_poses_3d is not None:
                out_poses_3d[i] = out_poses_3d[i][start:start + n_frames:stride]
    elif stride > 1:
        # Downsample as requested
        for i in range(len(out_poses_2d)):
            out_poses_2d[i] = out_poses_2d[i][::stride]
            if out_poses_3d is not None:
                out_poses_3d[i] = out_poses_3d[i][::stride]

    return out_camera_params, out_poses_3d, out_poses_2d


action_filter = None if args.actions == '*' else args.actions.split(',')
if action_filter is not None:
    print('Selected actions:', action_filter)

cameras_valid, poses_valid, poses_valid_2d = fetch(subjects_test, action_filter)

filter_widths = [int(x) for x in args.architecture.split(',')]
if not args.disable_optimizations and not args.dense and args.stride == 1:
    # Use optimized model for single-frame predictions
    model_pos_train = RIEModel(poses_valid_2d[0].shape[-2], poses_valid_2d[0].shape[-1],
                               dataset.skeleton().num_joints(),
                               filter_widths=filter_widths, causal=args.causal, dropout=args.dropout,
                               channels=args.channels, latten_features=args.latent_features_dim,
                               dense=args.dense, is_train=True, Optimize1f=True, stage=args.stage)
else:
    # When incompatible settings are detected (stride > 1, dense filters, or disabled optimization) fall back to normal model
    model_pos_train = RIEModel(poses_valid_2d[0].shape[-2], poses_valid_2d[0].shape[-1],
                               dataset.skeleton().num_joints(),
                               filter_widths=filter_widths, causal=args.causal, dropout=args.dropout,
                               channels=args.channels,
                               latten_features=args.latent_features_dim, dense=args.dense, is_train=True,
                               stage=args.stage)

model_pos = RIEModel(poses_valid_2d[0].shape[-2], poses_valid_2d[0].shape[-1], dataset.skeleton().num_joints(),
                     filter_widths=filter_widths, causal=args.causal, dropout=args.dropout, channels=args.channels,
                     latten_features=args.latent_features_dim, dense=args.dense, is_train=False, Optimize1f=True,
                     stage=args.stage)

receptive_field = model_pos.receptive_field()
print('INFO: Receptive field: {} frames'.format(receptive_field))
pad = (receptive_field - 1) // 2  # Padding on each side
if args.causal:
    print('INFO: Using causal convolutions')
    causal_shift = pad
else:
    causal_shift = 0

model_params = 0
for parameter in model_pos.parameters():
    model_params += parameter.numel()
print('INFO: Trainable parameter count:', model_params)

if torch.cuda.is_available():
    model_pos = model_pos.cuda()
    model_pos_train = model_pos_train.cuda()

if args.pretrain:
    pretrain_filename = os.path.join(args.checkpoint, args.pretrain)
    print('Loading pretrain model', pretrain_filename)
    checkpoint_p = torch.load(pretrain_filename, map_location=lambda storage, loc: storage)
    pretrain_dict = checkpoint_p['model_pos']

    temp = pretrain_dict.items()

    model_dict = model_pos_train.state_dict()

    state_dict = {k: v for k, v in pretrain_dict.items() if k in model_dict.keys()}

    state_dict = {k: v for i, (k, v) in enumerate(state_dict.items()) if i < 317}

    model_dict.update(state_dict)
    model_pos_train.load_state_dict(model_dict)

    cnt = 0
    for name, value in model_pos_train.named_parameters():

        if cnt < 167:
            value.requires_grad = False
        cnt = cnt + 1

if args.resume or args.evaluate or args.finetune:
    filename = ""
    if args.resume != "":
        filename = args.resume
    elif args.evaluate != "":
        filename = args.evaluate
    else:
        filename = args.finetune
    chk_filename = os.path.join(args.checkpoint, filename)
    print('Loading checkpoint', chk_filename)
    checkpoint = torch.load(chk_filename, map_location=lambda storage, loc: storage)
    print('This model was trained for {} epochs'.format(checkpoint['epoch']))
    model_pos_train.load_state_dict(checkpoint['model_pos'])
    model_pos.load_state_dict(checkpoint['model_pos'])

test_generator = ChunkedGenerator(args.batch_size // args.stride, cameras_valid, poses_valid, poses_valid_2d,
                                  args.stride,
                                  pad=pad, causal_shift=causal_shift, augment=args.test_time_augmentation,
                                  shuffle=False,
                                  kps_left=kps_left, kps_right=kps_right, joints_left=joints_left,
                                  joints_right=joints_right)
print('INFO: Testing on {} frames'.format(test_generator.num_frames()))

if not args.evaluate:
    cameras_train, poses_train, poses_train_2d = fetch(subjects_train, action_filter, subset=args.subset)

    lr = args.learning_rate

    optimizer = optim.AdamW(filter(lambda p: p.requires_grad, model_pos_train.parameters()), lr=lr, amsgrad=True)

    lr_decay = args.lr_decay

    losses_3d_train = []
    losses_3d_train_eval = []
    losses_3d_valid = []

    epoch = 0
    initial_momentum = 0.1
    final_momentum = 0.001

    train_generator = ChunkedGenerator(args.batch_size // args.stride, cameras_train, poses_train, poses_train_2d,
                                       args.stride,
                                       pad=pad, causal_shift=causal_shift, shuffle=True, augment=args.data_augmentation,
                                       kps_left=kps_left, kps_right=kps_right, joints_left=joints_left,
                                       joints_right=joints_right)
    train_generator_eval = ChunkedGenerator(args.batch_size // args.stride, cameras_train, poses_train, poses_train_2d,
                                            args.stride,
                                            pad=pad, causal_shift=causal_shift, shuffle=False,
                                            augment=args.data_augmentation,
                                            kps_left=kps_left, kps_right=kps_right, joints_left=joints_left,
                                            joints_right=joints_right)
    print('INFO: Training on {} frames'.format(train_generator_eval.num_frames()))

    if args.resume:
        epoch = checkpoint['epoch']
        if 'optimizer' in checkpoint and checkpoint['optimizer'] is not None:
            optimizer.load_state_dict(checkpoint['optimizer'])
            train_generator.set_random_state(checkpoint['random_state'])
        else:
            print('WARNING: this checkpoint does not contain an optimizer state. The optimizer will be reinitialized.')

        lr = checkpoint['lr']

    print('** Note: reported losses are averaged over all frames and test-time augmentation is not used here.')
    print('** The final evaluation will be carried out after the last training epoch.')

    # Pos model only
    while epoch < args.epochs:
        start_time = time()
        epoch_loss_3d_train = 0
        epoch_loss_traj_train = 0
        epoch_loss_2d_train_unlabeled = 0
        batch_ptr = 0
        N = 0
        N_semi = 0
        batch_cnt = 0
        batch_num = train_generator.get_num_batches()
        model_pos_train.train()

        # Regular supervised scenario
        for _, batch_3d, batch_2d in train_generator.next_epoch():
            inputs_3d = torch.from_numpy(batch_3d.astype('float32'))
            inputs_2d = torch.from_numpy(batch_2d.astype('float32'))
            if torch.cuda.is_available():
                inputs_3d = inputs_3d.cuda()
                inputs_2d = inputs_2d.cuda()
            inputs_3d[:, :, 0] = 0

            if batch_cnt % 100 == 0:
                print("Batch: %d/%d..." % (batch_cnt, batch_num))
            batch_cnt = batch_cnt + 1

            optimizer.zero_grad()

            # Predict 3D poses
            predicted_3d_pos = model_pos_train(inputs_2d)
            loss_3d_pos = mpjpe(predicted_3d_pos, inputs_3d)
            epoch_loss_3d_train += inputs_3d.shape[0] * inputs_3d.shape[1] * loss_3d_pos.item()
            N += inputs_3d.shape[0] * inputs_3d.shape[1]

            loss_total = loss_3d_pos
            loss_total.backward()

            optimizer.step()

        losses_3d_train.append(epoch_loss_3d_train / N)

        # End-of-epoch evaluation
        with torch.no_grad():
            model_pos.load_state_dict(model_pos_train.state_dict())
            model_pos.eval()

            epoch_loss_3d_valid = 0
            epoch_loss_traj_valid = 0
            epoch_loss_2d_valid = 0
            N = 0
            eval_time = time()
            batch_ptr = 0

            if not args.no_eval:
                # Evaluate on test set
                for cam, batch, batch_2d in test_generator.next_epoch():
                    inputs_3d = torch.from_numpy(batch.astype('float32'))
                    inputs_2d = torch.from_numpy(batch_2d.astype('float32'))
                    if torch.cuda.is_available():
                        inputs_3d = inputs_3d.cuda()
                        inputs_2d = inputs_2d.cuda()
                    inputs_traj = inputs_3d[:, :, :1].clone()
                    inputs_3d[:, :, 0] = 0

                    # Predict 3D poses
                    predicted_3d_pos = model_pos(inputs_2d)
                    loss_3d_pos = mpjpe(predicted_3d_pos, inputs_3d)
                    epoch_loss_3d_valid += inputs_3d.shape[0] * inputs_3d.shape[1] * loss_3d_pos.item()
                    N += inputs_3d.shape[0] * inputs_3d.shape[1]

                losses_3d_valid.append(epoch_loss_3d_valid / N)

                # Evaluate on training set, this time in evaluation mode
                epoch_loss_3d_train_eval = 0
                epoch_loss_traj_train_eval = 0
                epoch_loss_2d_train_labeled_eval = 0
                N = 0
                for cam, batch, batch_2d in train_generator_eval.next_epoch():
                    if batch_2d.shape[1] == 0:
                        # This can only happen when downsampling the dataset
                        continue

                    inputs_3d = torch.from_numpy(batch.astype('float32'))
                    inputs_2d = torch.from_numpy(batch_2d.astype('float32'))
                    if torch.cuda.is_available():
                        inputs_3d = inputs_3d.cuda()
                        inputs_2d = inputs_2d.cuda()
                    inputs_traj = inputs_3d[:, :, :1].clone()
                    inputs_3d[:, :, 0] = 0

                    # Compute 3D poses
                    predicted_3d_pos = model_pos(inputs_2d)
                    loss_3d_pos = mpjpe(predicted_3d_pos, inputs_3d)
                    epoch_loss_3d_train_eval += inputs_3d.shape[0] * inputs_3d.shape[1] * loss_3d_pos.item()
                    N += inputs_3d.shape[0] * inputs_3d.shape[1]

                losses_3d_train_eval.append(epoch_loss_3d_train_eval / N)

        elapsed = (time() - start_time) / 60
        eval_elapsed = (time() - eval_time) / 60

        if args.no_eval:
            print('[%d] time %.2f lr %f 3d_train %f' % (
                epoch + 1,
                elapsed,
                lr,
                losses_3d_train[-1] * 1000))
        else:

            print('[%d] time %.2f eval_time %.2f lr %f 3d_train %f 3d_eval %f 3d_valid %f' % (
                epoch + 1,
                elapsed,
                eval_elapsed,
                lr,
                losses_3d_train[-1] * 1000,
                losses_3d_train_eval[-1] * 1000,
                losses_3d_valid[-1] * 1000))

            log_path = os.path.join(args.checkpoint, 'training_log.txt')
            f = open(log_path, mode='a')
            f.write('[%d] time %.2f eval_time %.2f lr %f 3d_train %f 3d_eval %f 3d_valid %f\n' % (
                epoch + 1,
                elapsed,
                eval_elapsed,
                lr,
                losses_3d_train[-1] * 1000,
                losses_3d_train_eval[-1] * 1000,
                losses_3d_valid[-1] * 1000))
            f.close()

        # Decay learning rate exponentially
        lr *= lr_decay
        for param_group in optimizer.param_groups:
            param_group['lr'] *= lr_decay
        epoch += 1

        # Decay BatchNorm momentum
        momentum = initial_momentum * np.exp(-epoch / args.epochs * np.log(initial_momentum / final_momentum))
        model_pos_train.set_bn_momentum(momentum)

        # Save checkpoint if necessary
        if epoch % args.checkpoint_frequency == 0:
            chk_path = os.path.join(args.checkpoint, 'stage_' + str(args.stage) + '_epoch_{}.bin'.format(epoch))
            print('Saving checkpoint to', chk_path)

            torch.save({
                'epoch': epoch,
                'lr': lr,
                'random_state': train_generator.random_state(),
                'optimizer': optimizer.state_dict(),
                'model_pos': model_pos_train.state_dict(),
            }, chk_path)

        # Save training curves after every epoch, as .png images (if requested)
        if args.export_training_curves and epoch > 3:
            if 'matplotlib' not in sys.modules:
                import matplotlib

                matplotlib.use('Agg')
                import matplotlib.pyplot as plt

            plt.figure()
            epoch_x = np.arange(3, len(losses_3d_train)) + 1
            plt.plot(epoch_x, losses_3d_train[3:], '--', color='C0')
            plt.plot(epoch_x, losses_3d_train_eval[3:], color='C0')
            plt.plot(epoch_x, losses_3d_valid[3:], color='C1')
            plt.legend(['3d train', '3d train (eval)', '3d valid (eval)'])
            plt.ylabel('MPJPE (m)')
            plt.xlabel('Epoch')
            plt.xlim((3, epoch))
            plt.savefig(os.path.join(args.checkpoint, 'loss_3d.png'))

            plt.close('all')


# Evaluate
def evaluate(test_generator, action=None, return_predictions=False):
    epoch_loss_3d_pos = 0
    epoch_loss_3d_pos_procrustes = 0

    with torch.no_grad():
        model_pos.eval()
        N = 0
        cnt = 0
        #frame_num = test_generator.num_frames()
        #output = torch.zeros([frame_num, 17, 3], dtype=torch.float32)
        output = []

        if args.test_time_augmentation:
            for _, batch, batch_2d, batch_2d_flip in test_generator.next_epoch():
                inputs_2d = torch.from_numpy(batch_2d.astype('float32'))
                inputs_2d_flip = torch.from_numpy(batch_2d_flip.astype('float32'))
                if torch.cuda.is_available():
                    inputs_2d = inputs_2d.cuda()
                    inputs_2d_flip = inputs_2d_flip.cuda()

                # Positional model
                predicted_3d_pos = model_pos(inputs_2d)
                predicted_3d_pos_flip = model_pos(inputs_2d_flip)
                predicted_3d_pos_flip[:, :, :, 0] *= -1
                predicted_3d_pos_flip[:, :, joints_left + joints_right] = predicted_3d_pos_flip[:, :,
                                                                          joints_right + joints_left]

                predicted_3d_pos = torch.mean(torch.cat((predicted_3d_pos, predicted_3d_pos_flip), dim=1), dim=1,
                                              keepdim=True)

                if return_predictions:
                    if cnt == 0:
                        output = predicted_3d_pos.squeeze().cpu()
                    else:
                        output = np.concatenate((output, predicted_3d_pos.squeeze().cpu()), axis=0)
                    cnt = cnt + 1
                    continue

                inputs_3d = torch.from_numpy(batch.astype('float32'))
                if torch.cuda.is_available():
                    inputs_3d = inputs_3d.cuda()
                inputs_3d[:, :, 0] = 0

                error = mpjpe(predicted_3d_pos, inputs_3d)

                epoch_loss_3d_pos += inputs_3d.shape[0] * inputs_3d.shape[1] * error.item()
                N += inputs_3d.shape[0] * inputs_3d.shape[1]

                inputs = inputs_3d.cpu().numpy().reshape(-1, inputs_3d.shape[-2], inputs_3d.shape[-1])
                predicted_3d_pos = predicted_3d_pos.cpu().numpy().reshape(-1, inputs_3d.shape[-2], inputs_3d.shape[-1])

                epoch_loss_3d_pos_procrustes += inputs_3d.shape[0] * inputs_3d.shape[1] * p_mpjpe(predicted_3d_pos,
                                                                                                  inputs)
        else:
            for _, batch, batch_2d in test_generator.next_epoch():
                inputs_2d = torch.from_numpy(batch_2d.astype('float32'))
                if torch.cuda.is_available():
                    inputs_2d = inputs_2d.cuda()

                # Positional model
                predicted_3d_pos = model_pos(inputs_2d)

                if return_predictions:
                    return predicted_3d_pos.squeeze().cpu().numpy()

                inputs_3d = torch.from_numpy(batch.astype('float32'))
                if torch.cuda.is_available():
                    inputs_3d = inputs_3d.cuda()
                inputs_3d[:, :, 0] = 0

                error = mpjpe(predicted_3d_pos, inputs_3d)

                epoch_loss_3d_pos += inputs_3d.shape[0] * inputs_3d.shape[1] * error.item()
                N += inputs_3d.shape[0] * inputs_3d.shape[1]

                inputs = inputs_3d.cpu().numpy().reshape(-1, inputs_3d.shape[-2], inputs_3d.shape[-1])
                predicted_3d_pos = predicted_3d_pos.cpu().numpy().reshape(-1, inputs_3d.shape[-2], inputs_3d.shape[-1])

                epoch_loss_3d_pos_procrustes += inputs_3d.shape[0] * inputs_3d.shape[1] * p_mpjpe(predicted_3d_pos,
                                                                                                  inputs)

        if return_predictions:
            return output

    log_path = os.path.join(args.checkpoint, 'evaluating_log.txt')
    f = open(log_path, mode='a')
    if action is None:
        print('----------')
        f.write('----------\n')
    else:
        print('----' + action + '----\n')
        f.write('----' + action + '----\n')
    e1 = (epoch_loss_3d_pos / N) * 1000
    e2 = (epoch_loss_3d_pos_procrustes / N) * 1000
    print('Test time augmentation:', test_generator.augment_enabled())
    print('Protocol #1 Error (MPJPE):', e1, 'mm')
    print('Protocol #2 Error (P-MPJPE):', e2, 'mm')
    print('----------')

    f.write('Test time augmentation:' + str(test_generator.augment_enabled()) + '\n')
    f.write('Protocol #1 Error (MPJPE):' + str(e1) + 'mm\n')
    f.write('Protocol #2 Error (P-MPJPE):' + str(e2) + 'mm\n')
    f.write('----------\n')
    f.close()

    return e1, e2


if args.render:

    print('Rendering...')

    input_keypoints = keypoints[args.viz_subject][args.viz_action][args.viz_camera].copy()
    ground_truth = None
    if args.viz_subject in dataset.subjects() and args.viz_action in dataset[args.viz_subject]:
        if 'positions_3d' in dataset[args.viz_subject][args.viz_action]:
            ground_truth = dataset[args.viz_subject][args.viz_action]['positions_3d'][args.viz_camera].copy()
    if ground_truth is None:
        print('INFO: this action is unlabeled. Ground truth will not be rendered.')

    gen = Evaluate_Generator(args.batch_size, None, None, [input_keypoints], args.stride,
                             pad=pad, causal_shift=causal_shift, augment=args.test_time_augmentation,
                             shuffle=False,
                             kps_left=kps_left, kps_right=kps_right, joints_left=joints_left,
                             joints_right=joints_right)
    prediction = evaluate(gen, return_predictions=True)

    if args.viz_export is not None:
        print('Exporting joint positions to', args.viz_export)
        # Predictions are in camera space
        np.save(args.viz_export, prediction)

    if args.viz_output is not None:
        if ground_truth is not None:
            # Reapply trajectory
            trajectory = ground_truth[:, :1]
            ground_truth[:, 1:] += trajectory
            prediction += trajectory

        # Invert camera transformation
        cam = dataset.cameras()[args.viz_subject][args.viz_camera]
        if ground_truth is not None:
            prediction = camera_to_world(prediction, R=cam['orientation'], t=cam['translation'])
            ground_truth = camera_to_world(ground_truth, R=cam['orientation'], t=cam['translation'])
        else:
            # If the ground truth is not available, take the camera extrinsic params from a random subject.
            # They are almost the same, and anyway, we only need this for visualization purposes.
            for subject in dataset.cameras():
                if 'orientation' in dataset.cameras()[subject][args.viz_camera]:
                    rot = dataset.cameras()[subject][args.viz_camera]['orientation']
                    break
            prediction = camera_to_world(prediction, R=rot, t=0)
            # We don't have the trajectory, but at least we can rebase the height
            prediction[:, :, 2] -= np.min(prediction[:, :, 2])

        anim_output = {'Reconstruction': prediction}
        if ground_truth is not None and not args.viz_no_ground_truth:
            anim_output['Ground truth'] = ground_truth

        input_keypoints = image_coordinates(input_keypoints[..., :2], w=cam['res_w'], h=cam['res_h'])

        from common.visualization import render_animation

        render_animation(input_keypoints, keypoints_metadata, anim_output,
                         dataset.skeleton(), dataset.fps(), args.viz_bitrate, cam['azimuth'], args.viz_output,
                         limit=args.viz_limit, downsample=args.viz_downsample, size=args.viz_size,
                         input_video_path=args.viz_video, viewport=(cam['res_w'], cam['res_h']),
                         input_video_skip=args.viz_skip, viz_action=args.viz_action, viz_subject=args.viz_subject)


else:
    print('Evaluating...')
    all_actions = {}
    all_actions_by_subject = {}
    for subject in subjects_test:
        if subject not in all_actions_by_subject:
            all_actions_by_subject[subject] = {}

        for action in dataset[subject].keys():
            action_name = action.split(' ')[0]
            if action_name not in all_actions:
                all_actions[action_name] = []
            if action_name not in all_actions_by_subject[subject]:
                all_actions_by_subject[subject][action_name] = []
            all_actions[action_name].append((subject, action))
            all_actions_by_subject[subject][action_name].append((subject, action))


    def fetch_actions(actions):
        out_poses_3d = []
        out_poses_2d = []

        for subject, action in actions:
            poses_2d = keypoints[subject][action]
            for i in range(len(poses_2d)):  # Iterate across cameras
                out_poses_2d.append(poses_2d[i])

            poses_3d = dataset[subject][action]['positions_3d']
            assert len(poses_3d) == len(poses_2d), 'Camera count mismatch'
            for i in range(len(poses_3d)):  # Iterate across cameras
                out_poses_3d.append(poses_3d[i])

        stride = args.downsample
        if stride > 1:
            # Downsample as requested
            for i in range(len(out_poses_2d)):
                out_poses_2d[i] = out_poses_2d[i][::stride]
                if out_poses_3d is not None:
                    out_poses_3d[i] = out_poses_3d[i][::stride]

        return out_poses_3d, out_poses_2d


    def run_evaluation(actions, action_filter=None):
        errors_p1 = []
        errors_p2 = []

        for action_key in actions.keys():
            if action_filter is not None:
                found = False
                for a in action_filter:
                    if action_key.startswith(a):
                        found = True
                        break
                if not found:
                    continue

            poses_act, poses_2d_act = fetch_actions(actions[action_key])
            gen = Evaluate_Generator(args.batch_size, None, poses_act, poses_2d_act, args.stride,
                                     pad=pad, causal_shift=causal_shift, augment=args.test_time_augmentation,
                                     shuffle=False,
                                     kps_left=kps_left, kps_right=kps_right, joints_left=joints_left,
                                     joints_right=joints_right)
            e1, e2 = evaluate(gen, action_key)
            errors_p1.append(e1)
            errors_p2.append(e2)

        print('Protocol #1   (MPJPE) action-wise average:', round(np.mean(errors_p1), 1), 'mm')
        print('Protocol #2 (P-MPJPE) action-wise average:', round(np.mean(errors_p2), 1), 'mm')

        log_path = os.path.join(args.checkpoint, 'evaluating_log.txt')
        f = open(log_path, mode='a')
        f.write('Protocol #1   (MPJPE) action-wise average:' + str(round(np.mean(errors_p1), 1)) + 'mm\n')
        f.write('Protocol #2 (P-MPJPE) action-wise average:' + str(round(np.mean(errors_p2), 1)) + 'mm\n')
        f.close()


    if not args.by_subject:
        run_evaluation(all_actions, action_filter)
    else:
        for subject in all_actions_by_subject.keys():
            print('Evaluating on subject', subject)
            run_evaluation(all_actions_by_subject[subject], action_filter)
            print('')