From 93855b2faf8b795d0058c217ee980d435f23227d Mon Sep 17 00:00:00 2001
From: natanielruiz <nataniel777@hotmail.com>
Date: 星期四, 14 九月 2017 08:54:14 +0800
Subject: [PATCH] Training on AFLW with different yaw loss multipliers
---
code/train.py | 24 ++++++++++++++++++++----
1 files changed, 20 insertions(+), 4 deletions(-)
diff --git a/code/train.py b/code/train.py
index 6e1ae5b..03d5cf5 100644
--- a/code/train.py
+++ b/code/train.py
@@ -48,6 +48,7 @@
default=0.001, type=float)
parser.add_argument('--iter_ref', dest='iter_ref', help='Number of iterative refinement passes.',
default=1, type=int)
+ parser.add_argument('--dataset', dest='dataset', help='Dataset type.', default='Pose_300W_LP', type=str)
args = parser.parse_args()
return args
@@ -124,8 +125,19 @@
transforms.RandomCrop(224), transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
- pose_dataset = datasets.Pose_300W_LP(args.data_dir, args.filename_list,
- transformations)
+ if args.dataset == 'Pose_300W_LP':
+ pose_dataset = datasets.Pose_300W_LP(args.data_dir, args.filename_list, transformations)
+ elif args.dataset == 'AFLW2000':
+ pose_dataset = datasets.AFLW2000(args.data_dir, args.filename_list, transformations)
+ elif args.dataset == 'BIWI':
+ pose_dataset = datasets.BIWI(args.data_dir, args.filename_list, transformations)
+ elif args.dataset == 'AFLW':
+ pose_dataset = datasets.AFLW(args.data_dir, args.filename_list, transformations)
+ elif args.dataset == 'AFW':
+ pose_dataset = datasets.AFW(args.data_dir, args.filename_list, transformations)
+ else:
+ print 'Error: not a valid dataset name'
+ sys.exit()
train_loader = torch.utils.data.DataLoader(dataset=pose_dataset,
batch_size=batch_size,
shuffle=True,
@@ -239,10 +251,14 @@
loss_pitch += alpha * loss_reg_pitch
loss_roll += alpha * loss_reg_roll
+ loss_yaw *= 0.35
+
# Finetuning loss
loss_seq = [loss_yaw, loss_pitch, loss_roll]
- for idx in xrange(args.iter_ref+1):
- loss_angles = reg_criterion(angles[idx], label_angles.float())
+ for idx in xrange(1,len(angles)):
+ label_angles_residuals = label_angles.float() - angles[0]
+ label_angles_residuals = label_angles_residuals.detach()
+ loss_angles = reg_criterion(angles[idx], label_angles_residuals)
loss_seq.append(loss_angles)
grad_seq = [torch.Tensor(1).cuda(gpu) for _ in range(len(loss_seq))]
--
Gitblit v1.8.0