From 54818d253649ff588ed0054d10dabb2a3a170309 Mon Sep 17 00:00:00 2001
From: natanielruiz <nataniel777@hotmail.com>
Date: 星期四, 10 八月 2017 04:08:12 +0800
Subject: [PATCH] Doing pretty well now with resnet50 and adam with low learning rate. Also fixed test script to use large batches.
---
code/train_resnet_bins.py | 47 ++++++++++++++++++++++++++++++++++++++++++++---
1 files changed, 44 insertions(+), 3 deletions(-)
diff --git a/code/train_resnet_bins.py b/code/train_resnet_bins.py
index dab3800..f98bbc3 100644
--- a/code/train_resnet_bins.py
+++ b/code/train_resnet_bins.py
@@ -6,6 +6,7 @@
from torchvision import transforms
import torchvision
import torch.backends.cudnn as cudnn
+import torch.nn.functional as F
import cv2
import matplotlib.pyplot as plt
@@ -113,11 +114,19 @@
model.cuda(gpu)
criterion = nn.CrossEntropyLoss()
+ reg_criterion = nn.MSELoss()
+ # Regression loss coefficient
+ alpha = 0.01
+ lsm = nn.Softmax()
+
+ idx_tensor = [idx for idx in xrange(66)]
+ idx_tensor = torch.FloatTensor(idx_tensor).cuda(gpu)
+
optimizer = torch.optim.Adam([{'params': get_ignored_params(model), 'lr': args.lr},
{'params': get_non_ignored_params(model), 'lr': args.lr * 10}],
lr = args.lr)
# optimizer = torch.optim.SGD([{'params': get_ignored_params(model), 'lr': args.lr},
- # {'params': get_non_ignored_params(model), 'lr': args.lr}],
+ # {'params': get_non_ignored_params(model), 'lr': args.lr}],
# lr = args.lr, momentum=0.9)
# optimizer = torch.optim.RMSprop([{'params': get_ignored_params(model), 'lr': args.lr},
# {'params': get_non_ignored_params(model), 'lr': args.lr * 10}],
@@ -134,24 +143,56 @@
optimizer.zero_grad()
yaw, pitch, roll = model(images)
+
loss_yaw = criterion(yaw, label_yaw)
loss_pitch = criterion(pitch, label_pitch)
loss_roll = criterion(roll, label_roll)
+ # loss_seq = [loss_yaw, loss_pitch, loss_roll]
+ # grad_seq = [torch.Tensor(1).cuda(gpu) for _ in range(len(loss_seq))]
+ # torch.autograd.backward(loss_seq, grad_seq)
+ # optimizer.step()
+
+ # MSE loss
+ yaw_predicted = F.softmax(yaw)
+ pitch_predicted = F.softmax(pitch)
+ roll_predicted = F.softmax(roll)
+
+ yaw_predicted = torch.sum(yaw_predicted.data * idx_tensor, 1)
+ pitch_predicted = torch.sum(pitch_predicted.data * idx_tensor, 1)
+ roll_predicted = torch.sum(roll_predicted.data * idx_tensor, 1)
+
+ loss_reg_yaw = reg_criterion(yaw_predicted, label_yaw.float())
+ loss_reg_pitch = reg_criterion(pitch_predicted, label_pitch.float())
+ loss_reg_roll = reg_criterion(roll_predicted, label_roll.float())
+
+ # print yaw_predicted[0], label_yaw.data[0]
+
+ loss_yaw += alpha * loss_reg_yaw
+ loss_pitch += alpha * loss_reg_pitch
+ loss_roll += alpha * loss_reg_roll
+
loss_seq = [loss_yaw, loss_pitch, loss_roll]
grad_seq = [torch.Tensor(1).cuda(gpu) for _ in range(len(loss_seq))]
+ model.zero_grad()
torch.autograd.backward(loss_seq, grad_seq)
optimizer.step()
+
+ # print ('Epoch [%d/%d], Iter [%d/%d] Losses: Yaw %.4f, Pitch %.4f, Roll %.4f'
+ # %(epoch+1, num_epochs, i+1, len(pose_dataset)//batch_size, loss_yaw.data[0], loss_pitch.data[0], loss_roll.data[0]))
if (i+1) % 100 == 0:
print ('Epoch [%d/%d], Iter [%d/%d] Losses: Yaw %.4f, Pitch %.4f, Roll %.4f'
%(epoch+1, num_epochs, i+1, len(pose_dataset)//batch_size, loss_yaw.data[0], loss_pitch.data[0], loss_roll.data[0]))
+ # if epoch == 0:
+ # torch.save(model.state_dict(),
+ # 'output/snapshots/resnet18_sgd_iter_'+ str(i+1) + '.pkl')
# Save models at numbered epochs.
if epoch % 1 == 0 and epoch < num_epochs - 1:
print 'Taking snapshot...'
torch.save(model.state_dict(),
- 'output/snapshots/resnet18_cr_epoch_'+ str(epoch+1) + '.pkl')
+ 'output/snapshots/resnet18_sgd_epoch_'+ str(epoch+1) + '.pkl')
# Save the final Trained Model
- torch.save(model.state_dict(), 'output/snapshots/resnet18_cr_epoch_' + str(epoch+1) + '.pkl')
+ torch.save(model.state_dict(), 'output/snapshots/resnet18_sgd_epoch_' + str(epoch+1) + '.pkl')
--
Gitblit v1.8.0