101 lines
3.4 KiB
Python
101 lines
3.4 KiB
Python
import torch
|
|
import torch.nn.functional as F
|
|
import torch.optim as optim
|
|
import sys
|
|
import os
|
|
|
|
from tqdm import tqdm
|
|
sys.path.append(os.getcwd())
|
|
|
|
|
|
class Early_Stop:
|
|
def __init__(self, eps = -1e-3, stop_threshold = 10) -> None:
|
|
self.min_loss=float('inf')
|
|
self.eps=eps
|
|
self.stop_threshold=stop_threshold
|
|
self.satis_num=0
|
|
|
|
def update(self, loss):
|
|
delta = (loss - self.min_loss) / self.min_loss
|
|
if float(loss) < self.min_loss:
|
|
self.min_loss = float(loss)
|
|
update_res=True
|
|
else:
|
|
update_res=False
|
|
if delta >= self.eps:
|
|
self.satis_num += 1
|
|
else:
|
|
self.satis_num = max(0,self.satis_num-1)
|
|
return update_res, self.satis_num >= self.stop_threshold
|
|
|
|
|
|
def init(smpl_layer, target, device, cfg):
|
|
params={}
|
|
params["pose_params"] = torch.rand(target.shape[0], 72) * 0.0
|
|
params["shape_params"] = torch.rand(target.shape[0], 10) * 0.03
|
|
params["scale"] = torch.ones([1])
|
|
|
|
smpl_layer = smpl_layer.to(device)
|
|
params["pose_params"] = params["pose_params"].to(device)
|
|
params["shape_params"] = params["shape_params"].to(device)
|
|
target = target.to(device)
|
|
params["scale"] = params["scale"].to(device)
|
|
|
|
params["pose_params"].requires_grad = True
|
|
params["shape_params"].requires_grad = True
|
|
params["scale"].requires_grad = False
|
|
|
|
optimizer = optim.Adam([params["pose_params"], params["shape_params"]],
|
|
lr=cfg.TRAIN.LEARNING_RATE)
|
|
|
|
index={}
|
|
smpl_index=[]
|
|
dataset_index=[]
|
|
for tp in cfg.DATASET.DATA_MAP:
|
|
smpl_index.append(tp[0])
|
|
dataset_index.append(tp[1])
|
|
|
|
index["smpl_index"]=torch.tensor(smpl_index).to(device)
|
|
index["dataset_index"]=torch.tensor(dataset_index).to(device)
|
|
|
|
return smpl_layer, params,target, optimizer, index
|
|
|
|
|
|
def train(smpl_layer, target,
|
|
logger, writer, device,
|
|
args, cfg):
|
|
res = []
|
|
smpl_layer, params,target, optimizer, index = \
|
|
init(smpl_layer, target, device, cfg)
|
|
pose_params = params["pose_params"]
|
|
shape_params = params["shape_params"]
|
|
scale = params["scale"]
|
|
|
|
early_stop = Early_Stop()
|
|
for epoch in tqdm(range(cfg.TRAIN.MAX_EPOCH)):
|
|
# for epoch in range(cfg.TRAIN.MAX_EPOCH):
|
|
verts, Jtr = smpl_layer(pose_params, th_betas=shape_params)
|
|
loss = F.smooth_l1_loss(Jtr.index_select(1, index["smpl_index"]) * 100,
|
|
target.index_select(1, index["dataset_index"]) * 100)
|
|
optimizer.zero_grad()
|
|
loss.backward()
|
|
optimizer.step()
|
|
|
|
update_res, stop = early_stop.update(float(loss))
|
|
if update_res:
|
|
res = [pose_params, shape_params, verts, Jtr]
|
|
if stop:
|
|
logger.info("Early stop at epoch {} !".format(epoch))
|
|
break
|
|
|
|
if epoch % cfg.TRAIN.WRITE == 0:
|
|
# logger.info("Epoch {}, lossPerBatch={:.6f}, EarlyStopSatis: {}".format(
|
|
# epoch, float(loss), early_stop.satis_num))
|
|
writer.add_scalar('loss', float(loss), epoch)
|
|
writer.add_scalar('learning_rate', float(
|
|
optimizer.state_dict()['param_groups'][0]['lr']), epoch)
|
|
|
|
|
|
logger.info('Train ended, min_loss = {:.9f}'.format(float(early_stop.min_loss)))
|
|
return res
|