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/test_on_video.py | 4 code/train_resnet_bins_comb.py | 198 +++++++++++++++++++ practice/.ipynb_checkpoints/smoothing_ypr-checkpoint.ipynb | 42 ++- code/test_resnet_bins.py | 34 ++- code/train_resnet_bins.py | 47 ++++ code/train_resnet_bins_comb_dup.py | 198 +++++++++++++++++++ practice/smoothing_ypr.ipynb | 42 ++- 7 files changed, 512 insertions(+), 53 deletions(-) diff --git a/code/test_on_video.py b/code/test_on_video.py index 247c2db..20dfaac 100644 --- a/code/test_on_video.py +++ b/code/test_on_video.py @@ -48,9 +48,9 @@ # ResNet101 with 3 outputs. # model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 23, 3], 66) # ResNet50 - # model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 6, 3], 66) + model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 6, 3], 66) # ResNet18 - model = hopenet.Hopenet(torchvision.models.resnet.BasicBlock, [2, 2, 2, 2], 66) + # model = hopenet.Hopenet(torchvision.models.resnet.BasicBlock, [2, 2, 2, 2], 66) print 'Loading snapshot.' # Load snapshot diff --git a/code/test_resnet_bins.py b/code/test_resnet_bins.py index 00d2109..699c9c9 100644 --- a/code/test_resnet_bins.py +++ b/code/test_resnet_bins.py @@ -42,16 +42,15 @@ args = parse_args() cudnn.enabled = True - batch_size = 1 gpu = args.gpu_id snapshot_path = os.path.join('output/snapshots', args.snapshot + '.pkl') # ResNet101 with 3 outputs. # model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 23, 3], 66) # ResNet50 - # model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 6, 3], 66) + model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 6, 3], 66) # ResNet18 - model = hopenet.Hopenet(torchvision.models.resnet.BasicBlock, [2, 2, 2, 2], 66) + # model = hopenet.Hopenet(torchvision.models.resnet.BasicBlock, [2, 2, 2, 2], 66) print 'Loading snapshot.' # Load snapshot @@ -66,7 +65,7 @@ pose_dataset = datasets.AFLW2000_binned(args.data_dir, args.filename_list, transformations) test_loader = torch.utils.data.DataLoader(dataset=pose_dataset, - batch_size=batch_size, + batch_size=args.batch_size, num_workers=2) model.cuda(gpu) @@ -88,12 +87,14 @@ pitch_error = .0 roll_error = .0 + l1loss = torch.nn.L1Loss(size_average=False) + for i, (images, labels, name) in enumerate(test_loader): images = Variable(images).cuda(gpu) total += labels.size(0) - label_yaw = labels[:,0] - label_pitch = labels[:,1] - label_roll = labels[:,2] + label_yaw = labels[:,0].float() + label_pitch = labels[:,1].float() + label_roll = labels[:,2].float() yaw, pitch, roll = model(images) @@ -107,14 +108,18 @@ roll_predicted = F.softmax(roll) # Continuous predictions - yaw_predicted = torch.sum(yaw_predicted.data[0] * idx_tensor) - pitch_predicted = torch.sum(pitch_predicted.data[0] * idx_tensor) - roll_predicted = torch.sum(roll_predicted.data[0] * idx_tensor) + 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) + + yaw_predicted = yaw_predicted.cpu() + pitch_predicted = pitch_predicted.cpu() + roll_predicted = roll_predicted.cpu() # Mean absolute error - yaw_error += abs(yaw_predicted - label_yaw[0]) * 3 - pitch_error += abs(pitch_predicted - label_pitch[0]) * 3 - roll_error += abs(roll_predicted - label_roll[0]) * 3 + yaw_error += torch.sum(torch.abs(yaw_predicted - label_yaw) * 3) + pitch_error += torch.sum(torch.abs(pitch_predicted - label_pitch) * 3) + roll_error += torch.sum(torch.abs(roll_predicted - label_roll) * 3) # Binned Accuracy # for er in xrange(n_margins): @@ -125,13 +130,14 @@ # print label_yaw[0], yaw_bpred[0,0] # Save images with pose cube. + # TODO: fix for larger batch size if args.save_viz: name = name[0] cv2_img = cv2.imread(os.path.join(args.data_dir, name + '.jpg')) #print os.path.join('output/images', name + '.jpg') #print label_yaw[0] * 3 - 99, label_pitch[0] * 3 - 99, label_roll[0] * 3 - 99 #print yaw_predicted * 3 - 99, pitch_predicted * 3 - 99, roll_predicted * 3 - 99 - utils.plot_pose_cube(cv2_img, yaw_predicted * 3 - 99, pitch_predicted * 3 - 99, roll_predicted * 3 - 99) + utils.plot_pose_cube(cv2_img, yaw_predicted[0] * 3 - 99, pitch_predicted[0] * 3 - 99, roll_predicted[0] * 3 - 99) cv2.imwrite(os.path.join('output/images', name + '.jpg'), cv2_img) print('Test error in degrees of the model on the ' + str(total) + 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') diff --git a/code/train_resnet_bins_comb.py b/code/train_resnet_bins_comb.py new file mode 100644 index 0000000..eb23590 --- /dev/null +++ b/code/train_resnet_bins_comb.py @@ -0,0 +1,198 @@ +import numpy as np +import torch +import torch.nn as nn +from torch.autograd import Variable +from torch.utils.data import DataLoader +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 +import sys +import os +import argparse + +import datasets +import hopenet +import torch.utils.model_zoo as model_zoo + +model_urls = { + 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', + 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', + 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', + 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', + 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', +} + +def parse_args(): + """Parse input arguments.""" + parser = argparse.ArgumentParser(description='Head pose estimation using the Hopenet network.') + parser.add_argument('--gpu', dest='gpu_id', help='GPU device id to use [0]', + default=0, type=int) + parser.add_argument('--num_epochs', dest='num_epochs', help='Maximum number of training epochs.', + default=5, type=int) + parser.add_argument('--batch_size', dest='batch_size', help='Batch size.', + default=16, type=int) + parser.add_argument('--lr', dest='lr', help='Base learning rate.', + default=0.001, type=float) + parser.add_argument('--data_dir', dest='data_dir', help='Directory path for data.', + default='', type=str) + parser.add_argument('--filename_list', dest='filename_list', help='Path to text file containing relative paths for every example.', + default='', type=str) + + args = parser.parse_args() + + return args + +def get_ignored_params(model): + # Generator function that yields ignored params. + b = [] + b.append(model.conv1) + b.append(model.bn1) + b.append(model.layer1) + b.append(model.layer2) + b.append(model.layer3) + b.append(model.layer4) + for i in range(len(b)): + for j in b[i].modules(): + for k in j.parameters(): + yield k + +def get_non_ignored_params(model): + # Generator function that yields params that will be optimized. + b = [] + b.append(model.fc_yaw) + b.append(model.fc_pitch) + b.append(model.fc_roll) + for i in range(len(b)): + for j in b[i].modules(): + for k in j.parameters(): + yield k + +def load_filtered_state_dict(model, snapshot): + # By user apaszke from discuss.pytorch.org + model_dict = model.state_dict() + # 1. filter out unnecessary keys + snapshot = {k: v for k, v in snapshot.items() if k in model_dict} + # 2. overwrite entries in the existing state dict + model_dict.update(snapshot) + # 3. load the new state dict + model.load_state_dict(model_dict) + +if __name__ == '__main__': + args = parse_args() + + cudnn.enabled = True + num_epochs = args.num_epochs + batch_size = args.batch_size + gpu = args.gpu_id + + if not os.path.exists('output/snapshots'): + os.makedirs('output/snapshots') + + # ResNet101 with 3 outputs + # model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 23, 3], 66) + # ResNet50 + model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 6, 3], 66) + # ResNet18 + # model = hopenet.Hopenet(torchvision.models.resnet.BasicBlock, [2, 2, 2, 2], 66) + load_filtered_state_dict(model, model_zoo.load_url(model_urls['resnet50'])) + + print 'Loading data.' + + transformations = transforms.Compose([transforms.Scale(224),transforms.RandomCrop(224), + transforms.ToTensor()]) + + pose_dataset = datasets.Pose_300W_LP_binned(args.data_dir, args.filename_list, + transformations) + train_loader = torch.utils.data.DataLoader(dataset=pose_dataset, + batch_size=batch_size, + shuffle=True, + num_workers=2) + + model.cuda(gpu) + criterion = nn.CrossEntropyLoss() + reg_criterion = nn.MSELoss() + # Regression loss coefficient + alpha = 0.1 + 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}], + # lr = args.lr, momentum=0.9, weight_decay=5e-4) + # optimizer = torch.optim.RMSprop([{'params': get_ignored_params(model), 'lr': args.lr}, + # {'params': get_non_ignored_params(model), 'lr': args.lr * 10}], + # lr = args.lr) + + print 'Ready to train network.' + + for epoch in range(num_epochs): + for i, (images, labels, name) in enumerate(train_loader): + images = Variable(images).cuda(gpu) + label_yaw = Variable(labels[:,0]).cuda(gpu) + label_pitch = Variable(labels[:,1]).cuda(gpu) + label_roll = Variable(labels[:,2]).cuda(gpu) + + 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/resnet50_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/resnet50_epoch_'+ str(epoch+1) + '.pkl') + + # Save the final Trained Model + torch.save(model.state_dict(), 'output/snapshots/resnet50_epoch_' + str(epoch+1) + '.pkl') diff --git a/code/train_resnet_bins_comb_dup.py b/code/train_resnet_bins_comb_dup.py new file mode 100644 index 0000000..b435b89 --- /dev/null +++ b/code/train_resnet_bins_comb_dup.py @@ -0,0 +1,198 @@ +import numpy as np +import torch +import torch.nn as nn +from torch.autograd import Variable +from torch.utils.data import DataLoader +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 +import sys +import os +import argparse + +import datasets +import hopenet +import torch.utils.model_zoo as model_zoo + +model_urls = { + 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', + 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', + 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', + 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', + 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', +} + +def parse_args(): + """Parse input arguments.""" + parser = argparse.ArgumentParser(description='Head pose estimation using the Hopenet network.') + parser.add_argument('--gpu', dest='gpu_id', help='GPU device id to use [0]', + default=0, type=int) + parser.add_argument('--num_epochs', dest='num_epochs', help='Maximum number of training epochs.', + default=5, type=int) + parser.add_argument('--batch_size', dest='batch_size', help='Batch size.', + default=16, type=int) + parser.add_argument('--lr', dest='lr', help='Base learning rate.', + default=0.001, type=float) + parser.add_argument('--data_dir', dest='data_dir', help='Directory path for data.', + default='', type=str) + parser.add_argument('--filename_list', dest='filename_list', help='Path to text file containing relative paths for every example.', + default='', type=str) + + args = parser.parse_args() + + return args + +def get_ignored_params(model): + # Generator function that yields ignored params. + b = [] + b.append(model.conv1) + b.append(model.bn1) + b.append(model.layer1) + b.append(model.layer2) + b.append(model.layer3) + b.append(model.layer4) + for i in range(len(b)): + for j in b[i].modules(): + for k in j.parameters(): + yield k + +def get_non_ignored_params(model): + # Generator function that yields params that will be optimized. + b = [] + b.append(model.fc_yaw) + b.append(model.fc_pitch) + b.append(model.fc_roll) + for i in range(len(b)): + for j in b[i].modules(): + for k in j.parameters(): + yield k + +def load_filtered_state_dict(model, snapshot): + # By user apaszke from discuss.pytorch.org + model_dict = model.state_dict() + # 1. filter out unnecessary keys + snapshot = {k: v for k, v in snapshot.items() if k in model_dict} + # 2. overwrite entries in the existing state dict + model_dict.update(snapshot) + # 3. load the new state dict + model.load_state_dict(model_dict) + +if __name__ == '__main__': + args = parse_args() + + cudnn.enabled = True + num_epochs = args.num_epochs + batch_size = args.batch_size + gpu = args.gpu_id + + if not os.path.exists('output/snapshots'): + os.makedirs('output/snapshots') + + # ResNet101 with 3 outputs + # model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 23, 3], 66) + # ResNet50 + model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 6, 3], 66) + # ResNet18 + # model = hopenet.Hopenet(torchvision.models.resnet.BasicBlock, [2, 2, 2, 2], 66) + load_filtered_state_dict(model, model_zoo.load_url(model_urls['resnet50'])) + + print 'Loading data.' + + transformations = transforms.Compose([transforms.Scale(224),transforms.RandomCrop(224), + transforms.ToTensor()]) + + pose_dataset = datasets.Pose_300W_LP_binned(args.data_dir, args.filename_list, + transformations) + train_loader = torch.utils.data.DataLoader(dataset=pose_dataset, + batch_size=batch_size, + shuffle=True, + num_workers=2) + + model.cuda(gpu) + criterion = nn.CrossEntropyLoss() + reg_criterion = nn.MSELoss() + # Regression loss coefficient + alpha = 0.1 + 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}], + # lr = args.lr, momentum=0.9, weight_decay=5e-4) + # optimizer = torch.optim.RMSprop([{'params': get_ignored_params(model), 'lr': args.lr}, + # {'params': get_non_ignored_params(model), 'lr': args.lr * 10}], + # lr = args.lr) + + print 'Ready to train network.' + + for epoch in range(num_epochs): + for i, (images, labels, name) in enumerate(train_loader): + images = Variable(images).cuda(gpu) + label_yaw = Variable(labels[:,0]).cuda(gpu) + label_pitch = Variable(labels[:,1]).cuda(gpu) + label_roll = Variable(labels[:,2]).cuda(gpu) + + 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/resnet50_lowlr_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/resnet50_lowlr_epoch_'+ str(epoch+1) + '.pkl') + + # Save the final Trained Model + torch.save(model.state_dict(), 'output/snapshots/resnet50_lowlr_epoch_' + str(epoch+1) + '.pkl') diff --git a/practice/.ipynb_checkpoints/smoothing_ypr-checkpoint.ipynb b/practice/.ipynb_checkpoints/smoothing_ypr-checkpoint.ipynb index a411c30..8102abf 100644 --- a/practice/.ipynb_checkpoints/smoothing_ypr-checkpoint.ipynb +++ b/practice/.ipynb_checkpoints/smoothing_ypr-checkpoint.ipynb @@ -2,7 +2,7 @@ "cells": [ { "cell_type": "code", - "execution_count": 156, + "execution_count": 187, "metadata": { "collapsed": false }, @@ -17,7 +17,7 @@ }, { "cell_type": "code", - "execution_count": 157, + "execution_count": 188, "metadata": { "collapsed": false }, @@ -26,13 +26,13 @@ "video_path = '../data/video/SGT036_2016_07_25_pivothead_AVI.avi'\n", "bbox_path = '../data/video/annotations/SGT036_childface.txt'\n", "\n", - "annot_path = '../output/video/output-SGT036_resnet18_cr_epoch_1.txt'\n", - "output_string = 'SGT036_resnet18_cr_epoch_1_flat_smoothed'" + "annot_path = '../output/video/output-SGT036_resnet50_lowlr_epoch_20.txt'\n", + "output_string = 'SGT036_resnet50_lowlr_epoch_20_smoothed'" ] }, { "cell_type": "code", - "execution_count": 158, + "execution_count": 189, "metadata": { "collapsed": false }, @@ -41,7 +41,7 @@ "name": "stdout", "output_type": "stream", "text": [ - "[-6.069214 -0.831665 0.53318 ..., -3.836042 -3.868275 -2.377155]\n", + "[ 4.170376 0.790443 -0.178368 ..., -3.437805 0.396835 -1.276176]\n", "(8508,)\n", "(53464,)\n" ] @@ -93,7 +93,7 @@ }, { "cell_type": "code", - "execution_count": 159, + "execution_count": 190, "metadata": { "collapsed": false }, @@ -107,31 +107,39 @@ } ], "source": [ - "window_len = 6\n", + "window_len = 5\n", "pad = window_len / 2\n", "window = 'flat'\n", + "window_2 = 'flat'\n", + "window_len_2 = 7\n", + "pad_2 = window_len_2 / 2\n", "\n", "s = np.r_[y[window_len-1:0:-1],y,y[-2:-window_len-1:-1]]\n", "t = np.r_[p[window_len-1:0:-1],p,p[-2:-window_len-1:-1]]\n", "u = np.r_[r[window_len-1:0:-1],r,r[-2:-window_len-1:-1]]\n", "\n", - "xa = np.r_[x_min_arr[window_len-1:0:-1],x_min_arr,x_min_arr[-2:-window_len-1:-1]]\n", - "xb = np.r_[x_max_arr[window_len-1:0:-1],x_max_arr,x_max_arr[-2:-window_len-1:-1]]\n", - "ya = np.r_[y_min_arr[window_len-1:0:-1],y_min_arr,y_min_arr[-2:-window_len-1:-1]]\n", - "yb = np.r_[y_max_arr[window_len-1:0:-1],y_max_arr,y_max_arr[-2:-window_len-1:-1]]\n", + "xa = np.r_[x_min_arr[window_len_2-1:0:-1],x_min_arr,x_min_arr[-2:-window_len_2-1:-1]]\n", + "xb = np.r_[x_max_arr[window_len_2-1:0:-1],x_max_arr,x_max_arr[-2:-window_len_2-1:-1]]\n", + "ya = np.r_[y_min_arr[window_len_2-1:0:-1],y_min_arr,y_min_arr[-2:-window_len_2-1:-1]]\n", + "yb = np.r_[y_max_arr[window_len_2-1:0:-1],y_max_arr,y_max_arr[-2:-window_len_2-1:-1]]\n", "\n", "if window == 'flat':\n", " w=np.ones(window_len, 'd')\n", "else:\n", " w=eval('np.' + window + '(window_len)')\n", + " \n", + "if window_2 == 'flat':\n", + " w_2=np.ones(window_len_2, 'd')\n", + "else:\n", + " w_2=eval('np.' + window_2 + '(window_len_2)') \n", "\n", "y = np.convolve(w / w.sum(), s, mode='valid')[pad:-pad]\n", "p = np.convolve(w / w.sum(), t, mode='valid')[pad:-pad]\n", "r = np.convolve(w / w.sum(), u, mode='valid')[pad:-pad]\n", - "x_min_arr = np.convolve(w / w.sum(), xa, mode='valid')[pad:-pad]\n", - "x_max_arr = np.convolve(w / w.sum(), xb, mode='valid')[pad:-pad]\n", - "y_min_arr = np.convolve(w / w.sum(), ya, mode='valid')[pad:-pad]\n", - "y_max_arr = np.convolve(w / w.sum(), yb, mode='valid')[pad:-pad]\n", + "x_min_arr = np.convolve(w_2 / w_2.sum(), xa, mode='valid')[pad_2:-pad_2]\n", + "x_max_arr = np.convolve(w_2 / w_2.sum(), xb, mode='valid')[pad_2:-pad_2]\n", + "y_min_arr = np.convolve(w_2 / w_2.sum(), ya, mode='valid')[pad_2:-pad_2]\n", + "y_max_arr = np.convolve(w_2 / w_2.sum(), yb, mode='valid')[pad_2:-pad_2]\n", "\n", "pose_dict = {}\n", "bbox_dict = {}\n", @@ -151,7 +159,7 @@ }, { "cell_type": "code", - "execution_count": 160, + "execution_count": 191, "metadata": { "collapsed": false }, diff --git a/practice/smoothing_ypr.ipynb b/practice/smoothing_ypr.ipynb index a411c30..96dc33f 100644 --- a/practice/smoothing_ypr.ipynb +++ b/practice/smoothing_ypr.ipynb @@ -2,7 +2,7 @@ "cells": [ { "cell_type": "code", - "execution_count": 156, + "execution_count": 197, "metadata": { "collapsed": false }, @@ -17,7 +17,7 @@ }, { "cell_type": "code", - "execution_count": 157, + "execution_count": 198, "metadata": { "collapsed": false }, @@ -26,13 +26,13 @@ "video_path = '../data/video/SGT036_2016_07_25_pivothead_AVI.avi'\n", "bbox_path = '../data/video/annotations/SGT036_childface.txt'\n", "\n", - "annot_path = '../output/video/output-SGT036_resnet18_cr_epoch_1.txt'\n", - "output_string = 'SGT036_resnet18_cr_epoch_1_flat_smoothed'" + "annot_path = '../output/video/output-SGT036_resnet50_lowlr_epoch_20.txt'\n", + "output_string = 'SGT036_resnet50_lowlr_epoch_20_smoothed'" ] }, { "cell_type": "code", - "execution_count": 158, + "execution_count": 199, "metadata": { "collapsed": false }, @@ -41,7 +41,7 @@ "name": "stdout", "output_type": "stream", "text": [ - "[-6.069214 -0.831665 0.53318 ..., -3.836042 -3.868275 -2.377155]\n", + "[ 4.170376 0.790443 -0.178368 ..., -3.437805 0.396835 -1.276176]\n", "(8508,)\n", "(53464,)\n" ] @@ -93,7 +93,7 @@ }, { "cell_type": "code", - "execution_count": 159, + "execution_count": 200, "metadata": { "collapsed": false }, @@ -107,31 +107,39 @@ } ], "source": [ - "window_len = 6\n", + "window_len = 7\n", "pad = window_len / 2\n", "window = 'flat'\n", + "window_2 = 'flat'\n", + "window_len_2 = 7\n", + "pad_2 = window_len_2 / 2\n", "\n", "s = np.r_[y[window_len-1:0:-1],y,y[-2:-window_len-1:-1]]\n", "t = np.r_[p[window_len-1:0:-1],p,p[-2:-window_len-1:-1]]\n", "u = np.r_[r[window_len-1:0:-1],r,r[-2:-window_len-1:-1]]\n", "\n", - "xa = np.r_[x_min_arr[window_len-1:0:-1],x_min_arr,x_min_arr[-2:-window_len-1:-1]]\n", - "xb = np.r_[x_max_arr[window_len-1:0:-1],x_max_arr,x_max_arr[-2:-window_len-1:-1]]\n", - "ya = np.r_[y_min_arr[window_len-1:0:-1],y_min_arr,y_min_arr[-2:-window_len-1:-1]]\n", - "yb = np.r_[y_max_arr[window_len-1:0:-1],y_max_arr,y_max_arr[-2:-window_len-1:-1]]\n", + "xa = np.r_[x_min_arr[window_len_2-1:0:-1],x_min_arr,x_min_arr[-2:-window_len_2-1:-1]]\n", + "xb = np.r_[x_max_arr[window_len_2-1:0:-1],x_max_arr,x_max_arr[-2:-window_len_2-1:-1]]\n", + "ya = np.r_[y_min_arr[window_len_2-1:0:-1],y_min_arr,y_min_arr[-2:-window_len_2-1:-1]]\n", + "yb = np.r_[y_max_arr[window_len_2-1:0:-1],y_max_arr,y_max_arr[-2:-window_len_2-1:-1]]\n", "\n", "if window == 'flat':\n", " w=np.ones(window_len, 'd')\n", "else:\n", " w=eval('np.' + window + '(window_len)')\n", + " \n", + "if window_2 == 'flat':\n", + " w_2=np.ones(window_len_2, 'd')\n", + "else:\n", + " w_2=eval('np.' + window_2 + '(window_len_2)') \n", "\n", "y = np.convolve(w / w.sum(), s, mode='valid')[pad:-pad]\n", "p = np.convolve(w / w.sum(), t, mode='valid')[pad:-pad]\n", "r = np.convolve(w / w.sum(), u, mode='valid')[pad:-pad]\n", - "x_min_arr = np.convolve(w / w.sum(), xa, mode='valid')[pad:-pad]\n", - "x_max_arr = np.convolve(w / w.sum(), xb, mode='valid')[pad:-pad]\n", - "y_min_arr = np.convolve(w / w.sum(), ya, mode='valid')[pad:-pad]\n", - "y_max_arr = np.convolve(w / w.sum(), yb, mode='valid')[pad:-pad]\n", + "x_min_arr = np.convolve(w_2 / w_2.sum(), xa, mode='valid')[pad_2:-pad_2]\n", + "x_max_arr = np.convolve(w_2 / w_2.sum(), xb, mode='valid')[pad_2:-pad_2]\n", + "y_min_arr = np.convolve(w_2 / w_2.sum(), ya, mode='valid')[pad_2:-pad_2]\n", + "y_max_arr = np.convolve(w_2 / w_2.sum(), yb, mode='valid')[pad_2:-pad_2]\n", "\n", "pose_dict = {}\n", "bbox_dict = {}\n", @@ -151,7 +159,7 @@ }, { "cell_type": "code", - "execution_count": 160, + "execution_count": 201, "metadata": { "collapsed": false }, -- Gitblit v1.8.0