def val(train_loader,
model,
input_n=20,
output_n=10,
is_cuda=False,
dim_used=[]):
# t_l = utils.AccumLoss()
t_e = utils.AccumLoss()
t_3d = utils.AccumLoss()
model.eval()
st = time.time()
bar = Bar('>>>', fill='>', max=len(train_loader))
for i, (inputs, targets, all_seq) in enumerate(train_loader):
bt = time.time()
if is_cuda:
inputs = Variable(inputs.cuda()).float()
# targets = Variable(targets.cuda(async=True)).float()
all_seq = Variable(all_seq.cuda()).float()
outputs = model(inputs)
n = outputs.shape[0]
outputs = outputs.view(n, -1)
# targets = targets.view(n, -1)
# loss = loss_funcs.sen_loss(outputs, all_seq, dim_used)
n, _, _ = all_seq.data.shape
m_err = loss_funcs.mpjpe_error(outputs, all_seq, input_n, dim_used,
input_n + output_n)
e_err = loss_funcs.euler_error(outputs, all_seq, input_n, dim_used,
input_n + output_n)
# t_l.update(loss.cpu().data.numpy()[0] * n, n)
t_e.update(e_err.cpu().data.numpy() * n, n)
t_3d.update(m_err.cpu().data.numpy() * n, n)
bar.suffix = '{}/{}|batch time {:.4f}s|total time{:.2f}s'.format(
i + 1, len(train_loader),
time.time() - bt,
time.time() - st)
bar.next()
bar.finish()
return t_e.avg, t_3d.avg
def train(self, train_loader, dataset='h3.6m', input_n=20, dct_n=20, lr_now=None, cartesian=False,
lambda_=0.01, max_norm=True, dim_used=[]):
t_l = utils.AccumLoss()
t_l_joint = utils.AccumLoss()
t_l_vlb = utils.AccumLoss()
t_l_latent = utils.AccumLoss()
t_e = utils.AccumLoss()
t_3d = utils.AccumLoss()
self.model.train()
st = time.time()
bar = Bar('>>>', fill='>', max=len(train_loader))
for i, (inputs, targets, all_seq) in enumerate(train_loader):
# skip the last batch if only have one sample for batch_norm layers
batch_size = inputs.shape[0]
if batch_size == 1:
continue
bt = time.time()
if self.is_cuda:
inputs = Variable(inputs.cuda()).float()
targets = Variable(targets.cuda(non_blocking=True)).float()
all_seq = Variable(all_seq.cuda(non_blocking=True)).float()
outputs, reconstructions, log_var, z = self.model(inputs.float())
KL = self.model.KL
n = outputs.shape[0]
outputs = outputs.view(n, -1)
loss, joint_loss, vlb, latent_loss = loss_funcs.sen_loss(outputs, all_seq, dim_used, dct_n, inputs,
cartesian, lambda_, KL, reconstructions, log_var)
# Print losses for epoch
ret_log = np.array([i, loss.cpu().data.numpy(), joint_loss.cpu().data.numpy(), vlb.cpu().data.numpy(),
latent_loss.cpu().data.numpy()])
df = pd.DataFrame(np.expand_dims(ret_log, axis=0))
if i == 0:
head = ['iteration', 'loss', 'joint_loss', 'vlb', 'latent_loss']
df.to_csv('losses.csv', header=head, index=False)
with open('losses.csv', 'a') as f:
df.to_csv(f, header=False, index=False)
# calculate loss and backward
self.optimizer.zero_grad()
loss.backward()
if max_norm:
nn.utils.clip_grad_norm(self.model.parameters(), max_norm=1)
self.optimizer.step()
n, _, _ = all_seq.data.shape
if dataset == 'h3.6m':
# 3d error
m_err = loss_funcs.mpjpe_error(outputs, all_seq, input_n, dim_used, dct_n)
# angle space error
e_err = loss_funcs.euler_error(outputs, all_seq, input_n, dim_used, dct_n)
elif dataset == 'cmu_mocap':
m_err = loss_funcs.mpjpe_error_cmu(outputs, all_seq, input_n, dim_used=dim_used, dct_n=dct_n)
e_err = loss_funcs.euler_error(outputs, all_seq, input_n, dim_used=dim_used, dct_n=dct_n)
elif dataset == 'cmu_mocap_3d':
m_err = loss
e_err = loss
# update the training loss
t_l.update(loss.cpu().data.numpy() * n, n)
t_l_joint.update(joint_loss.cpu().data.numpy() * n, n)
t_l_vlb.update(vlb.cpu().data.numpy() * n, n)
t_l_latent.update(latent_loss.cpu().data.numpy() * n, n)
t_e.update(e_err.cpu().data.numpy() * n, n)
t_3d.update(m_err.cpu().data.numpy() * n, n)
bar.suffix = '{}/{}|batch time {:.4f}s|total time{:.2f}s'.format(i + 1, len(train_loader), time.time() - bt,
time.time() - st)
bar.next()
bar.finish()
print("\nJoint loss: ", t_l_joint.avg)
print("vlb: ", t_l_vlb.avg)
print("Latent loss: ", t_l_latent.avg)
print("loss: ", t_l.avg)
return lr_now, t_l.avg, t_l_joint.avg, t_l_vlb.avg, t_l_latent.avg, t_e.avg, t_3d.avg
def train(train_loader,
model,
optimizer,
input_n=20,
dct_n=20,
lr_now=None,
max_norm=True,
is_cuda=False,
dim_used=[]):
t_l = utils.AccumLoss()
t_e = utils.AccumLoss()
t_3d = utils.AccumLoss()
model.train()
st = time.time()
bar = Bar('>>>', fill='>', max=len(train_loader))
for i, (inputs, targets, all_seq) in enumerate(train_loader):
# skip the last batch if only have one sample for batch_norm layers
batch_size = inputs.shape[0]
if batch_size == 1:
continue
bt = time.time()
if is_cuda:
inputs = Variable(inputs.cuda()).float()
# targets = Variable(targets.cuda(async=True)).float()
all_seq = Variable(all_seq.cuda(async=True)).float()
outputs = model(inputs)
n = outputs.shape[0]
outputs = outputs.view(n, -1)
# targets = targets.view(n, -1)
loss = loss_funcs.sen_loss(outputs, all_seq, dim_used, dct_n)
# calculate loss and backward
optimizer.zero_grad()
loss.backward()
if max_norm:
nn.utils.clip_grad_norm(model.parameters(), max_norm=1)
optimizer.step()
n, _, _ = all_seq.data.shape
# 3d error
m_err = loss_funcs.mpjpe_error(outputs, all_seq, input_n, dim_used,
dct_n)
# angle space error
e_err = loss_funcs.euler_error(outputs, all_seq, input_n, dim_used,
dct_n)
# update the training loss
t_l.update(loss.cpu().data.numpy()[0] * n, n)
t_e.update(e_err.cpu().data.numpy()[0] * n, n)
t_3d.update(m_err.cpu().data.numpy()[0] * n, n)
bar.suffix = '{}/{}|batch time {:.4f}s|total time{:.2f}s'.format(
i + 1, len(train_loader),
time.time() - bt,
time.time() - st)
bar.next()
bar.finish()
return lr_now, t_l.avg, t_e.avg, t_3d.avg