From af51d0ecb51ad4d6c8ed086855bd3c411ebc4aa0 Mon Sep 17 00:00:00 2001
From: natanielruiz <nruiz9@gatech.edu>
Date: 星期一, 30 十月 2017 06:29:51 +0800
Subject: [PATCH] Fixed stuff
---
code/train_alexnet.py | 22 ++++++++++++++--------
1 files changed, 14 insertions(+), 8 deletions(-)
diff --git a/code/train_alexnet.py b/code/train_alexnet.py
index 9254ee7..51cf43b 100644
--- a/code/train_alexnet.py
+++ b/code/train_alexnet.py
@@ -129,6 +129,9 @@
# Regression loss coefficient
alpha = args.alpha
+ idx_tensor = [idx for idx in xrange(66)]
+ idx_tensor = Variable(torch.FloatTensor(idx_tensor)).cuda(gpu)
+
optimizer = torch.optim.Adam([{'params': get_ignored_params(model), 'lr': 0},
{'params': get_non_ignored_params(model), 'lr': args.lr},
{'params': get_fc_params(model), 'lr': args.lr * 5}],
@@ -150,17 +153,21 @@
label_roll_cont = Variable(cont_labels[:,2]).cuda(gpu)
# Forward pass
- yaw, pitch, roll, angles = model(images)
+ pre_yaw, pre_pitch, pre_roll = model(images)
# Cross entropy loss
- loss_yaw = criterion(yaw, label_yaw)
- loss_pitch = criterion(pitch, label_pitch)
- loss_roll = criterion(roll, label_roll)
+ loss_yaw = criterion(pre_yaw, label_yaw)
+ loss_pitch = criterion(pre_pitch, label_pitch)
+ loss_roll = criterion(pre_roll, label_roll)
# MSE loss
- yaw_predicted = angles[:,0]
- pitch_predicted = angles[:,1]
- roll_predicted = angles[:,2]
+ yaw_predicted = softmax(pre_yaw)
+ pitch_predicted = softmax(pre_pitch)
+ roll_predicted = softmax(pre_roll)
+
+ yaw_predicted = torch.sum(yaw_predicted * idx_tensor, 1) * 3 - 99
+ pitch_predicted = torch.sum(pitch_predicted * idx_tensor, 1) * 3 - 99
+ roll_predicted = torch.sum(roll_predicted * idx_tensor, 1) * 3 - 99
loss_reg_yaw = reg_criterion(yaw_predicted, label_yaw_cont)
loss_reg_pitch = reg_criterion(pitch_predicted, label_pitch_cont)
@@ -173,7 +180,6 @@
loss_seq = [loss_yaw, loss_pitch, loss_roll]
grad_seq = [torch.Tensor(1).cuda(gpu) for _ in range(len(loss_seq))]
- optimizer.zero_grad()
torch.autograd.backward(loss_seq, grad_seq)
optimizer.step()
--
Gitblit v1.8.0