def _test_step(self, image, kp3d, return_kps=False):
tf.keras.backend.set_learning_phase(0)
if len(tf.shape(image)) is not 4:
image = tf.expand_dims(image, 0)
kp3d = tf.expand_dims(kp3d, 0)
result = self.generator(image, training=False)
# only use last computed theta (from accumulated iterative feedback loop)
_, _, kp3d_pred, _, _, _ = result[-1]
factor = tf.constant(1000, tf.float32)
kp3d, kp3d_predict = kp3d * factor, kp3d_pred * factor # convert back from m -> mm
kp3d_predict = kp3d_predict[:, :self.config.NUM_KP3D, :]
real_kp3d = batch_align_by_pelvis(kp3d)
predict_kp3d = batch_align_by_pelvis(kp3d_predict)
kp3d_mpjpe = tf.norm(real_kp3d - predict_kp3d, axis=2)
aligned_kp3d = batch_compute_similarity_transform(
real_kp3d, predict_kp3d)
kp3d_mpjpe_aligned = tf.norm(real_kp3d - aligned_kp3d, axis=2)
if return_kps:
return kp3d_mpjpe, kp3d_mpjpe_aligned, predict_kp3d, real_kp3d
return kp3d_mpjpe, kp3d_mpjpe_aligned, None, None
def _val_step(self, images, kp2d, kp3d, has3d):
tf.keras.backend.set_learning_phase(0)
result = self.generator(images, training=False)
# only use last computed theta (from accumulated iterative feedback loop)
_, kp2d_pred, kp3d_pred, _, _, _ = result[-1]
vis = kp2d[:, :, 2]
kp2d_norm = tf.norm(
kp2d_pred[:, :self.config.NUM_KP2D, :] - kp2d[:, :, :2],
axis=2) * vis
kp2d_mpjpe = tf.reduce_sum(kp2d_norm) / tf.reduce_sum(vis)
self.kp2d_mpjpe_log(kp2d_mpjpe)
if self.config.USE_3D:
# check if at least one 3d sample available
if tf.reduce_sum(has3d) > 0:
kp3d_real = tf.boolean_mask(kp3d, has3d)
kp3d_predict = tf.boolean_mask(kp3d_pred, has3d)
kp3d_predict = kp3d_predict[:, :self.config.NUM_KP3D, :]
kp3d_real = batch_align_by_pelvis(kp3d_real)
kp3d_predict = batch_align_by_pelvis(kp3d_predict)
kp3d_mpjpe = tf.norm(kp3d_predict - kp3d_real, axis=2)
kp3d_mpjpe = tf.reduce_mean(kp3d_mpjpe)
aligned_kp3d = batch_compute_similarity_transform(
kp3d_real, kp3d_predict)
kp3d_mpjpe_aligned = tf.norm(aligned_kp3d - kp3d_real, axis=2)
kp3d_mpjpe_aligned = tf.reduce_mean(kp3d_mpjpe_aligned)
self.kp3d_mpjpe_log.update_state(kp3d_mpjpe)
self.kp3d_mpjpe_aligned_log.update_state(kp3d_mpjpe_aligned)
def test_batch_align_by_pelvis(self):
joints_3d = tf.ones((self.config.BATCH_SIZE, self.config.NUM_KP3D, 3))
output = batch_align_by_pelvis(joints_3d)
expected = tf.zeros((self.config.BATCH_SIZE, self.config.NUM_KP3D, 3))
self.assertAllCloseAccordingToType(expected, output)
self.assertEqual((self.config.BATCH_SIZE, self.config.NUM_KP3D, 3),
output.shape)
def _train_step(self, images, kp2d, kp3d, has3d, theta):
tf.keras.backend.set_learning_phase(1)
batch_size = images.shape[0]
with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape:
generator_outputs = self.generator(images, training=True)
# only use last computed theta (from iterative feedback loop)
_, kp2d_pred, kp3d_pred, pose_pred, shape_pred, _ = generator_outputs[
-1]
vis = tf.expand_dims(kp2d[:, :, 2], -1)
kp2d_loss = v1_loss.absolute_difference(kp2d[:, :, :2],
kp2d_pred,
weights=vis)
kp2d_loss = kp2d_loss * self.config.GENERATOR_2D_LOSS_WEIGHT
if self.config.USE_3D:
has3d = tf.expand_dims(has3d, -1)
kp3d_real = batch_align_by_pelvis(kp3d)
kp3d_pred = batch_align_by_pelvis(
kp3d_pred[:, :self.config.NUM_KP3D, :])
kp3d_real = tf.reshape(kp3d_real, [batch_size, -1])
kp3d_pred = tf.reshape(kp3d_pred, [batch_size, -1])
kp3d_loss = v1_loss.mean_squared_error(
kp3d_real, kp3d_pred, weights=has3d) * 0.5
kp3d_loss = kp3d_loss * self.config.GENERATOR_3D_LOSS_WEIGHT
"""Calculating pose and shape loss basically makes no sense
due to missing paired 3d and mosh ground truth data.
The original implementation has paired data for Human 3.6 M dataset
which was not published due to licence conflict.
Nevertheless with SMPLify paired data can be generated
(see http://smplify.is.tue.mpg.de/ for more information)
"""
pose_pred = tf.reshape(pose_pred, [batch_size, -1])
shape_pred = tf.reshape(shape_pred, [batch_size, -1])
pose_shape_pred = tf.concat([pose_pred, shape_pred], 1)
# fake ground truth
has_smpl = tf.zeros(batch_size,
tf.float32) # do not include loss
has_smpl = tf.expand_dims(has_smpl, -1)
pose_shape_real = tf.zeros(pose_shape_pred.shape)
ps_loss = v1_loss.mean_squared_error(
pose_shape_real, pose_shape_pred, weights=has_smpl) * 0.5
ps_loss = ps_loss * self.config.GENERATOR_3D_LOSS_WEIGHT
# use all poses and shapes from iterative feedback loop
fake_disc_input = self.accumulate_fake_disc_input(
generator_outputs)
fake_disc_output = self.discriminator(fake_disc_input,
training=True)
real_disc_input = self.accumulate_real_disc_input(theta)
real_disc_output = self.discriminator(real_disc_input,
training=True)
gen_disc_loss = tf.reduce_mean(
tf.reduce_sum((fake_disc_output - 1)**2, axis=1))
gen_disc_loss = gen_disc_loss * self.config.DISCRIMINATOR_LOSS_WEIGHT
generator_loss = tf.reduce_sum([kp2d_loss, gen_disc_loss])
if self.config.USE_3D:
generator_loss = tf.reduce_sum(
[generator_loss, kp3d_loss, ps_loss])
disc_real_loss = tf.reduce_mean(
tf.reduce_sum((real_disc_output - 1)**2, axis=1))
disc_fake_loss = tf.reduce_mean(
tf.reduce_sum(fake_disc_output**2, axis=1))
discriminator_loss = tf.reduce_sum(
[disc_real_loss, disc_fake_loss])
discriminator_loss = discriminator_loss * self.config.DISCRIMINATOR_LOSS_WEIGHT
generator_grads = gen_tape.gradient(generator_loss,
self.generator.trainable_variables)
discriminator_grads = disc_tape.gradient(
discriminator_loss, self.discriminator.trainable_variables)
self.generator_opt.apply_gradients(
zip(generator_grads, self.generator.trainable_variables))
self.discriminator_opt.apply_gradients(
zip(discriminator_grads, self.discriminator.trainable_variables))
self.generator_loss_log.update_state(generator_loss)
self.kp2d_loss_log.update_state(kp2d_loss)
self.gen_disc_loss_log.update_state(gen_disc_loss)
if self.config.USE_3D:
self.kp3d_loss_log.update_state(kp3d_loss)
self.pose_shape_loss_log.update_state(ps_loss)
self.discriminator_loss_log.update_state(discriminator_loss)
self.disc_real_loss_log.update_state(disc_real_loss)
self.disc_fake_loss_log.update_state(disc_fake_loss)
