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 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, UnchunkedGenerator
from time import time
from common.utils import deterministic_random

args = parse_args()
print(args)

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)
elif args.dataset.startswith('custom'):
    from common.custom_dataset import CustomDataset
    dataset = CustomDataset('data/data_2d_' + args.dataset + '_' + args.keypoints + '.npz')
else:
    raise KeyError('Invalid dataset')

print('Preparing data...')
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

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

semi_supervised = len(subjects_semi) > 0
if semi_supervised and not dataset.supports_semi_supervised():
    raise RuntimeError('Semi-supervised training is not implemented for this dataset')
            
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 = TemporalModelOptimized1f(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)
else:
    # When incompatible settings are detected (stride > 1, dense filters, or disabled optimization) fall back to normal model
    model_pos_train = TemporalModel(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,
                                dense=args.dense)
    
model_pos = TemporalModel(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,
                            dense=args.dense)

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.resume or args.evaluate:
    chk_filename = os.path.join(args.checkpoint, args.resume if args.resume else args.evaluate)
    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 = UnchunkedGenerator(cameras_valid, poses_valid, poses_valid_2d,
                                    pad=pad, causal_shift=causal_shift, augment=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
    if semi_supervised:
        cameras_semi, _, poses_semi_2d = fetch(subjects_semi, action_filter, parse_3d_poses=False)
        
        if not args.disable_optimizations and not args.dense and args.stride == 1:
            # Use optimized model for single-frame predictions
            model_traj_train = TemporalModelOptimized1f(poses_valid_2d[0].shape[-2], poses_valid_2d[0].shape[-1], 1,
                    filter_widths=filter_widths, causal=args.causal, dropout=args.dropout, channels=args.channels)
        else:
            # When incompatible settings are detected (stride > 1, dense filters, or disabled optimization) fall back to normal model
            model_traj_train = TemporalModel(poses_valid_2d[0].shape[-2], poses_valid_2d[0].shape[-1], 1,
                    filter_widths=filter_widths, causal=args.causal, dropout=args.dropout, channels=args.channels,
                    dense=args.dense)
        
        model_traj = TemporalModel(poses_valid_2d[0].shape[-2], poses_valid_2d[0].shape[-1], 1,
                            filter_widths=filter_widths, causal=args.causal, dropout=args.dropout, channels=args.channels,
                            dense=args.dense)
        if torch.cuda.is_available():
            model_traj = model_traj.cuda()
            model_traj_train = model_traj_train.cuda()
        optimizer = optim.Adam(list(model_pos_train.parameters()) + list(model_traj_train.parameters()),
                               lr=lr, amsgrad=True)
        
        losses_2d_train_unlabeled = []
        losses_2d_train_labeled_eval = []
        losses_2d_train_unlabeled_eval = []
        losses_2d_valid = []

        losses_traj_train = []
        losses_traj_train_eval = []
        losses_traj_valid = []
    else:
        optimizer = optim.Adam(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 = UnchunkedGenerator(cameras_train, poses_train, poses_train_2d,
                                              pad=pad, causal_shift=causal_shift, augment=False)
    print('INFO: Training on {} frames'.format(train_generator_eval.num_frames()))
    if semi_supervised:
        semi_generator = ChunkedGenerator(args.batch_size//args.stride, cameras_semi, None, poses_semi_2d, args.stride,
                                          pad=pad, causal_shift=causal_shift, shuffle=True,
                                          random_seed=4321, augment=args.data_augmentation,
                                          kps_left=kps_left, kps_right=kps_right, joints_left=joints_left, joints_right=joints_right,
                                          endless=True)
        semi_generator_eval = UnchunkedGenerator(cameras_semi, None, poses_semi_2d,
                                                 pad=pad, causal_shift=causal_shift, augment=False)
        print('INFO: Semi-supervision on {} frames'.format(semi_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']
        if semi_supervised:
            model_traj_train.load_state_dict(checkpoint['model_traj'])
            model_traj.load_state_dict(checkpoint['model_traj'])
            semi_generator.set_random_state(checkpoint['random_state_semi'])
            
    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
        N = 0
        N_semi = 0
        model_pos_train.train()
        if semi_supervised:
            # Semi-supervised scenario
            model_traj_train.train()
            for (_, batch_3d, batch_2d), (cam_semi, _, batch_2d_semi) in \
                zip(train_generator.next_epoch(), semi_generator.next_epoch()):
                
                # Fall back to supervised training for the first epoch (to avoid instability)
                skip = epoch < args.warmup
                
                cam_semi = torch.from_numpy(cam_semi.astype('float32'))
                inputs_3d = torch.from_numpy(batch_3d.astype('float32'))
                if torch.cuda.is_available():
                    cam_semi = cam_semi.cuda()
                    inputs_3d = inputs_3d.cuda()
                    
                inputs_traj = inputs_3d[:, :, :1].clone()
                inputs_3d[:, :, 0] = 0
                
                # Split point between labeled and unlabeled samples in the batch
                split_idx = inputs_3d.shape[0]

                inputs_2d = torch.from_numpy(batch_2d.astype('float32'))
                inputs_2d_semi = torch.from_numpy(batch_2d_semi.astype('float32'))
                if torch.cuda.is_available():
                    inputs_2d = inputs_2d.cuda()
                    inputs_2d_semi = inputs_2d_semi.cuda()
                inputs_2d_cat =  torch.cat((inputs_2d, inputs_2d_semi), dim=0) if not skip else inputs_2d

                optimizer.zero_grad()

                # Compute 3D poses
                predicted_3d_pos_cat = model_pos_train(inputs_2d_cat)

                loss_3d_pos = mpjpe(predicted_3d_pos_cat[:split_idx], 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

                # Compute global trajectory
                predicted_traj_cat = model_traj_train(inputs_2d_cat)
                w = 1 / inputs_traj[:, :, :, 2] # Weight inversely proportional to depth
                loss_traj = weighted_mpjpe(predicted_traj_cat[:split_idx], inputs_traj, w)
                epoch_loss_traj_train += inputs_3d.shape[0]*inputs_3d.shape[1] * loss_traj.item()
                assert inputs_traj.shape[0]*inputs_traj.shape[1] == inputs_3d.shape[0]*inputs_3d.shape[1]
                loss_total += loss_traj

                if not skip:
                    # Semi-supervised loss for unlabeled samples
                    predicted_semi = predicted_3d_pos_cat[split_idx:]
                    if pad > 0:
                        target_semi = inputs_2d_semi[:, pad:-pad, :, :2].contiguous()
                    else:
                        target_semi = inputs_2d_semi[:, :, :, :2].contiguous()
                        
                    projection_func = project_to_2d_linear if args.linear_projection else project_to_2d
                    reconstruction_semi = projection_func(predicted_semi + predicted_traj_cat[split_idx:], cam_semi)

                    loss_reconstruction = mpjpe(reconstruction_semi, target_semi) # On 2D poses
                    epoch_loss_2d_train_unlabeled += predicted_semi.shape[0]*predicted_semi.shape[1] * loss_reconstruction.item()
                    if not args.no_proj:
                        loss_total += loss_reconstruction
                    
                    # Bone length term to enforce kinematic constraints
                    if args.bone_length_term:
                        dists = predicted_3d_pos_cat[:, :, 1:] - predicted_3d_pos_cat[:, :, dataset.skeleton().parents()[1:]]
                        bone_lengths = torch.mean(torch.norm(dists, dim=3), dim=1)
                        penalty = torch.mean(torch.abs(torch.mean(bone_lengths[:split_idx], dim=0) \
                                                     - torch.mean(bone_lengths[split_idx:], dim=0)))
                        loss_total += penalty
                        
                    
                    N_semi += predicted_semi.shape[0]*predicted_semi.shape[1]
                else:
                    N_semi += 1 # To avoid division by zero

                loss_total.backward()

                optimizer.step()
            losses_traj_train.append(epoch_loss_traj_train / N)
            losses_2d_train_unlabeled.append(epoch_loss_2d_train_unlabeled / N_semi)
        else:
            # 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

                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()
            if semi_supervised:
                model_traj.load_state_dict(model_traj_train.state_dict())
                model_traj.eval()

            epoch_loss_3d_valid = 0
            epoch_loss_traj_valid = 0
            epoch_loss_2d_valid = 0
            N = 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]

                    if semi_supervised:
                        cam = torch.from_numpy(cam.astype('float32'))
                        if torch.cuda.is_available():
                            cam = cam.cuda()

                        predicted_traj = model_traj(inputs_2d)
                        loss_traj = mpjpe(predicted_traj, inputs_traj)
                        epoch_loss_traj_valid += inputs_traj.shape[0]*inputs_traj.shape[1] * loss_traj.item()
                        assert inputs_traj.shape[0]*inputs_traj.shape[1] == inputs_3d.shape[0]*inputs_3d.shape[1]

                        if pad > 0:
                            target = inputs_2d[:, pad:-pad, :, :2].contiguous()
                        else:
                            target = inputs_2d[:, :, :, :2].contiguous()
                        reconstruction = project_to_2d(predicted_3d_pos + predicted_traj, cam)
                        loss_reconstruction = mpjpe(reconstruction, target) # On 2D poses
                        epoch_loss_2d_valid += reconstruction.shape[0]*reconstruction.shape[1] * loss_reconstruction.item()
                        assert reconstruction.shape[0]*reconstruction.shape[1] == inputs_3d.shape[0]*inputs_3d.shape[1]

                losses_3d_valid.append(epoch_loss_3d_valid / N)
                if semi_supervised:
                    losses_traj_valid.append(epoch_loss_traj_valid / N)
                    losses_2d_valid.append(epoch_loss_2d_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]

                    if semi_supervised:
                        cam = torch.from_numpy(cam.astype('float32'))
                        if torch.cuda.is_available():
                            cam = cam.cuda()
                        predicted_traj = model_traj(inputs_2d)
                        loss_traj = mpjpe(predicted_traj, inputs_traj)
                        epoch_loss_traj_train_eval += inputs_traj.shape[0]*inputs_traj.shape[1] * loss_traj.item()
                        assert inputs_traj.shape[0]*inputs_traj.shape[1] == inputs_3d.shape[0]*inputs_3d.shape[1]

                        if pad > 0:
                            target = inputs_2d[:, pad:-pad, :, :2].contiguous()
                        else:
                            target = inputs_2d[:, :, :, :2].contiguous()
                        reconstruction = project_to_2d(predicted_3d_pos + predicted_traj, cam)
                        loss_reconstruction = mpjpe(reconstruction, target)
                        epoch_loss_2d_train_labeled_eval += reconstruction.shape[0]*reconstruction.shape[1] * loss_reconstruction.item()
                        assert reconstruction.shape[0]*reconstruction.shape[1] == inputs_3d.shape[0]*inputs_3d.shape[1]

                losses_3d_train_eval.append(epoch_loss_3d_train_eval / N)
                if semi_supervised:
                    losses_traj_train_eval.append(epoch_loss_traj_train_eval / N)
                    losses_2d_train_labeled_eval.append(epoch_loss_2d_train_labeled_eval / N)

                # Evaluate 2D loss on unlabeled training set (in evaluation mode)
                epoch_loss_2d_train_unlabeled_eval = 0
                N_semi = 0
                if semi_supervised:
                    for cam, _, batch_2d in semi_generator_eval.next_epoch():
                        cam = torch.from_numpy(cam.astype('float32'))
                        inputs_2d_semi = torch.from_numpy(batch_2d.astype('float32'))
                        if torch.cuda.is_available():
                            cam = cam.cuda()
                            inputs_2d_semi = inputs_2d_semi.cuda()

                        predicted_3d_pos_semi = model_pos(inputs_2d_semi)
                        predicted_traj_semi = model_traj(inputs_2d_semi)
                        if pad > 0:
                            target_semi = inputs_2d_semi[:, pad:-pad, :, :2].contiguous()
                        else:
                            target_semi = inputs_2d_semi[:, :, :, :2].contiguous()
                        reconstruction_semi = project_to_2d(predicted_3d_pos_semi + predicted_traj_semi, cam)
                        loss_reconstruction_semi = mpjpe(reconstruction_semi, target_semi)

                        epoch_loss_2d_train_unlabeled_eval += reconstruction_semi.shape[0]*reconstruction_semi.shape[1] \
                                                              * loss_reconstruction_semi.item()
                        N_semi += reconstruction_semi.shape[0]*reconstruction_semi.shape[1]
                    losses_2d_train_unlabeled_eval.append(epoch_loss_2d_train_unlabeled_eval / N_semi)

        elapsed = (time() - start_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:
            if semi_supervised:
                print('[%d] time %.2f lr %f 3d_train %f 3d_eval %f traj_eval %f 3d_valid %f '
                      'traj_valid %f 2d_train_sup %f 2d_train_unsup %f 2d_valid %f' % (
                        epoch + 1,
                        elapsed,
                        lr,
                        losses_3d_train[-1] * 1000,
                        losses_3d_train_eval[-1] * 1000,
                        losses_traj_train_eval[-1] * 1000,
                        losses_3d_valid[-1] * 1000,
                        losses_traj_valid[-1] * 1000,
                        losses_2d_train_labeled_eval[-1],
                        losses_2d_train_unlabeled_eval[-1],
                        losses_2d_valid[-1]))
            else:
                print('[%d] time %.2f lr %f 3d_train %f 3d_eval %f 3d_valid %f' % (
                        epoch + 1,
                        elapsed,
                        lr,
                        losses_3d_train[-1] * 1000,
                        losses_3d_train_eval[-1] * 1000,
                        losses_3d_valid[-1]  *1000))
        
        # 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)
        if semi_supervised:
            model_traj_train.set_bn_momentum(momentum)
            
        # Save checkpoint if necessary
        if epoch % args.checkpoint_frequency == 0:
            chk_path = os.path.join(args.checkpoint, '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(),
                'model_traj': model_traj_train.state_dict() if semi_supervised else None,
                'random_state_semi': semi_generator.random_state() if semi_supervised else None,
            }, 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'))

            if semi_supervised:
                plt.figure()
                plt.plot(epoch_x, losses_traj_train[3:], '--', color='C0')
                plt.plot(epoch_x, losses_traj_train_eval[3:], color='C0')
                plt.plot(epoch_x, losses_traj_valid[3:], color='C1')
                plt.legend(['traj. train', 'traj. train (eval)', 'traj. valid (eval)'])
                plt.ylabel('Mean distance (m)')
                plt.xlabel('Epoch')
                plt.xlim((3, epoch))
                plt.savefig(os.path.join(args.checkpoint, 'loss_traj.png'))

                plt.figure()
                plt.plot(epoch_x, losses_2d_train_labeled_eval[3:], color='C0')
                plt.plot(epoch_x, losses_2d_train_unlabeled[3:], '--', color='C1')
                plt.plot(epoch_x, losses_2d_train_unlabeled_eval[3:], color='C1')
                plt.plot(epoch_x, losses_2d_valid[3:], color='C2')
                plt.legend(['2d train labeled (eval)', '2d train unlabeled', '2d train unlabeled (eval)', '2d valid (eval)'])
                plt.ylabel('MPJPE (2D)')
                plt.xlabel('Epoch')
                plt.xlim((3, epoch))
                plt.savefig(os.path.join(args.checkpoint, 'loss_2d.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
    epoch_loss_3d_pos_scale = 0
    epoch_loss_3d_vel = 0
    with torch.no_grad():
        model_pos.eval()
        N = 0
        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)

            # Test-time augmentation (if enabled)
            if test_generator.augment_enabled():
                # Undo flipping and take average with non-flipped version
                predicted_3d_pos[1, :, :, 0] *= -1
                predicted_3d_pos[1, :, joints_left + joints_right] = predicted_3d_pos[1, :, joints_right + joints_left]
                predicted_3d_pos = torch.mean(predicted_3d_pos, dim=0, keepdim=True)
                
            if return_predictions:
                return predicted_3d_pos.squeeze(0).cpu().numpy()
                
            inputs_3d = torch.from_numpy(batch.astype('float32'))
            if torch.cuda.is_available():
                inputs_3d = inputs_3d.cuda()
            inputs_3d[:, :, 0] = 0    
            if test_generator.augment_enabled():
                inputs_3d = inputs_3d[:1]

            error = mpjpe(predicted_3d_pos, inputs_3d)
            epoch_loss_3d_pos_scale += inputs_3d.shape[0]*inputs_3d.shape[1] * n_mpjpe(predicted_3d_pos, inputs_3d).item()

            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)

            # Compute velocity error
            epoch_loss_3d_vel += inputs_3d.shape[0]*inputs_3d.shape[1] * mean_velocity_error(predicted_3d_pos, inputs)
            
    if action is None:
        print('----------')
    else:
        print('----'+action+'----')
    e1 = (epoch_loss_3d_pos / N)*1000
    e2 = (epoch_loss_3d_pos_procrustes / N)*1000
    e3 = (epoch_loss_3d_pos_scale / N)*1000
    ev = (epoch_loss_3d_vel / 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('Protocol #3 Error (N-MPJPE):', e3, 'mm')
    print('Velocity Error (MPJVE):', ev, 'mm')
    print('----------')

    return e1, e2, e3, ev


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 = UnchunkedGenerator(None, None, [input_keypoints],
                             pad=pad, causal_shift=causal_shift, augment=args.test_time_augmentation,
                             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)
    
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 = []
        errors_p3 = []
        errors_vel = []

        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 = UnchunkedGenerator(None, poses_act, poses_2d_act,
                                     pad=pad, causal_shift=causal_shift, augment=args.test_time_augmentation,
                                     kps_left=kps_left, kps_right=kps_right, joints_left=joints_left, joints_right=joints_right)
            e1, e2, e3, ev = evaluate(gen, action_key)
            errors_p1.append(e1)
            errors_p2.append(e2)
            errors_p3.append(e3)
            errors_vel.append(ev)

        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')
        print('Protocol #3 (N-MPJPE) action-wise average:', round(np.mean(errors_p3), 1), 'mm')
        print('Velocity      (MPJVE) action-wise average:', round(np.mean(errors_vel), 2), 'mm')

    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('')