def train_valid(model, optimizer, scheduler, epoch, dataloaders, data_size,
start_time):
for phase in ['train', 'valid']:
if args.pure_validation and phase == 'train':
continue
if args.pure_training and phase == 'valid':
continue
labels, distances = [], []
triplet_loss_sum = 0.0
if phase == 'train':
scheduler.step()
model.train()
else:
model.eval()
#for batch_idx in range(0, data_size[phase], 1):
for batch_idx, batch_sample in enumerate(dataloaders[phase]):
#print("batch_idx:", batch_idx)
try:
#batch_sample = dataloaders[phase][batch_idx]
if not 'exception' in batch_sample:
anc_img = batch_sample['anc_img'].to(device)
pos_img = batch_sample['pos_img'].to(device)
neg_img = batch_sample['neg_img'].to(device)
pos_cls = batch_sample['pos_class'].to(device)
neg_cls = batch_sample['neg_class'].to(device)
with torch.set_grad_enabled(phase == 'train'):
# anc_embed, pos_embed and neg_embed are encoding(embedding) of image
anc_embed, pos_embed, neg_embed = model(
anc_img), model(pos_img), model(neg_img)
#for i in anc_embed:
# print(i.item())
if args.num_valid_triplets <= 100:
anc_embed_cpu = anc_embed.cpu()
pos_embed_cpu = pos_embed.cpu()
neg_embed_cpu = neg_embed.cpu()
pos_cls_cpu = pos_cls.cpu()
neg_cls_cpu = neg_cls.cpu()
pd.DataFrame([t.numpy() for t in anc_embed_cpu
]).to_csv("./embeddings.csv",
mode='a',
header=None)
pd.DataFrame([t.numpy() for t in pos_embed_cpu
]).to_csv("./embeddings.csv",
mode='a',
header=None)
pd.DataFrame([t.numpy() for t in neg_embed_cpu
]).to_csv("./embeddings.csv",
mode='a',
header=None)
pd.DataFrame({
'type':
"anc",
'id':
batch_sample['anc_id'],
'class':
pos_cls_cpu,
'train_set':
args.train_csv_name.split('.')[0],
'val_set':
args.valid_csv_name.split('.')[0]
}).to_csv("./embeddings_info.csv",
mode='a',
header=None)
pd.DataFrame({
'type':
"pos",
'id':
batch_sample['pos_id'],
'class':
pos_cls_cpu,
'train_set':
args.train_csv_name.split('.')[0],
'val_set':
args.valid_csv_name.split('.')[0]
}).to_csv("./embeddings_info.csv",
mode='a',
header=None)
pd.DataFrame({
'type':
"neg",
'id':
batch_sample['neg_id'],
'class':
pos_cls_cpu,
'train_set':
args.train_csv_name.split('.')[0],
'val_set':
args.valid_csv_name.split('.')[0]
}).to_csv("./embeddings_info.csv",
mode='a',
header=None)
#print([t.size() for t in anc_embed])
# choose the hard negatives only for "training"
pos_dist = l2_dist.forward(anc_embed, pos_embed)
neg_dist = l2_dist.forward(anc_embed, neg_embed)
all = (neg_dist - pos_dist <
args.margin).cpu().numpy().flatten()
if phase == 'train':
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
else:
hard_triplets = np.where(all >= 0)
anc_hard_embed = anc_embed[hard_triplets].to(device)
pos_hard_embed = pos_embed[hard_triplets].to(device)
neg_hard_embed = neg_embed[hard_triplets].to(device)
anc_hard_img = anc_img[hard_triplets].to(device)
pos_hard_img = pos_img[hard_triplets].to(device)
neg_hard_img = neg_img[hard_triplets].to(device)
pos_hard_cls = pos_cls[hard_triplets].to(device)
neg_hard_cls = neg_cls[hard_triplets].to(device)
anc_img_pred = model.forward_classifier(
anc_hard_img).to(device)
pos_img_pred = model.forward_classifier(
pos_hard_img).to(device)
neg_img_pred = model.forward_classifier(
neg_hard_img).to(device)
triplet_loss = TripletLoss(args.margin).forward(
anc_hard_embed, pos_hard_embed,
neg_hard_embed).to(device)
if phase == 'train':
optimizer.zero_grad()
triplet_loss.backward()
optimizer.step()
dists = l2_dist.forward(anc_embed, pos_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.ones(dists.size(0)))
dists = l2_dist.forward(anc_embed, neg_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.zeros(dists.size(0)))
triplet_loss_sum += triplet_loss.item()
except:
#traceback.print_exc()
print("traceback: ", traceback.format_exc())
print("something went wrong with batch_idx: ", batch_idx,
", batch_sample:", batch_sample, ", neg_img size: ",
batch_sample['neg_img'].shape, ", pos_img size: ",
batch_sample['pos_img'].shape, ", anc_img size: ",
batch_sample['anc_img'].shape)
avg_triplet_loss = triplet_loss_sum / data_size[phase]
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array(
[subdist for dist in distances for subdist in dist])
nrof_pairs = min(len(labels), len(distances))
if nrof_pairs >= 10:
tpr, fpr, accuracy, val, val_std, far = evaluate(distances, labels)
print(' {} set - Triplet Loss = {:.8f}'.format(
phase, avg_triplet_loss))
print(' {} set - Accuracy = {:.8f}'.format(
phase, np.mean(accuracy)))
duration = time.time() - start_time
with open('{}/{}_log_epoch{}.txt'.format(log_dir, phase, epoch),
'w') as f:
f.write(
str(epoch) + '\t' + str(np.mean(accuracy)) + '\t' +
str(avg_triplet_loss) + '\t' + str(duration))
if phase == 'train':
torch.save({
'epoch': epoch,
'state_dict': model.state_dict()
}, '{}/checkpoint_epoch{}.pth'.format(save_dir, epoch))
else:
plot_roc(fpr,
tpr,
figure_name='{}/roc_valid_epoch_{}.png'.format(
log_dir, epoch))
def train_valid(model, optimizer, triploss, scheduler, epoch, dataloaders, data_size):
for phase in ['train', 'valid']:
labels, distances = [], []
triplet_loss_sum = 0.0
if phase == 'train':
scheduler.step()
if scheduler.last_epoch % scheduler.step_size == 0:
print("LR decayed to:", ', '.join(map(str, scheduler.get_lr())))
model.train()
else:
model.eval()
for batch_idx, batch_sample in enumerate(dataloaders[phase]):
anc_img = batch_sample['anc_img'].to(device)
pos_img = batch_sample['pos_img'].to(device)
neg_img = batch_sample['neg_img'].to(device)
# pos_cls = batch_sample['pos_class'].to(device)
# neg_cls = batch_sample['neg_class'].to(device)
with torch.set_grad_enabled(phase == 'train'):
# anc_embed, pos_embed and neg_embed are encoding(embedding) of image
anc_embed, pos_embed, neg_embed = model(anc_img), model(pos_img), model(neg_img)
# choose the semi hard negatives only for "training"
pos_dist = l2_dist.forward(anc_embed, pos_embed)
neg_dist = l2_dist.forward(anc_embed, neg_embed)
all = (neg_dist - pos_dist < args.margin).cpu().numpy().flatten()
if phase == 'train':
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
else:
hard_triplets = np.where(all >= 0)
anc_hard_embed = anc_embed[hard_triplets]
pos_hard_embed = pos_embed[hard_triplets]
neg_hard_embed = neg_embed[hard_triplets]
anc_hard_img = anc_img[hard_triplets]
pos_hard_img = pos_img[hard_triplets]
neg_hard_img = neg_img[hard_triplets]
# pos_hard_cls = pos_cls[hard_triplets]
# neg_hard_cls = neg_cls[hard_triplets]
model.module.forward_classifier(anc_hard_img)
model.module.forward_classifier(pos_hard_img)
model.module.forward_classifier(neg_hard_img)
triplet_loss = triploss.forward(anc_hard_embed, pos_hard_embed, neg_hard_embed)
if phase == 'train':
optimizer.zero_grad()
triplet_loss.backward()
optimizer.step()
distances.append(pos_dist.data.cpu().numpy())
labels.append(np.ones(pos_dist.size(0)))
distances.append(neg_dist.data.cpu().numpy())
labels.append(np.zeros(neg_dist.size(0)))
triplet_loss_sum += triplet_loss.item()
avg_triplet_loss = triplet_loss_sum / data_size[phase]
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array([subdist for dist in distances for subdist in dist])
tpr, fpr, accuracy, val, val_std, far = evaluate(distances, labels)
print(' {} set - Triplet Loss = {:.8f}'.format(phase, avg_triplet_loss))
print(' {} set - Accuracy = {:.8f}'.format(phase, np.mean(accuracy)))
time = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
lr = '_'.join(map(str, scheduler.get_lr()))
layers = '+'.join(args.unfreeze.split(','))
write_csv(f'log/{phase}.csv', [time, epoch, np.mean(accuracy), avg_triplet_loss, layers, args.batch_size, lr])
if phase == 'valid':
save_last_checkpoint({'epoch': epoch,
'state_dict': model.module.state_dict(),
'optimizer_state': optimizer.state_dict(),
'accuracy': np.mean(accuracy),
'loss': avg_triplet_loss
})
save_if_best({'epoch': epoch,
'state_dict': model.module.state_dict(),
'optimizer_state': optimizer.state_dict(),
'accuracy': np.mean(accuracy),
'loss': avg_triplet_loss
}, np.mean(accuracy))
else:
plot_roc(fpr, tpr, figure_name='./log/roc_valid_epoch_{}.png'.format(epoch))
import numpy as np from IPython import embed from eval_metrics import evaluate from eval_metrics import evaluate, plot_roc, plot_acc targets = np.random.randint(1, 10, (100)) distances = np.random.randn(100) tpr, fpr, acc, val, val_std, far = evaluate(distances, targets) embed() plot_roc(fpr, tpr, figure_name='./test.png')
def train_valid(model, optimizer, scheduler, epoch, dataloaders, data_size):
for phase in ['train', 'valid']:
labels, distances = [], []
triplet_loss_sum = 0.0
if phase == 'train':
scheduler.step()
model.train()
else:
model.eval()
for batch_idx, batch_sample in enumerate(dataloaders[phase]):
#break
anc_img = batch_sample['anc_img'].to(device)
pos_img = batch_sample['pos_img'].to(device)
neg_img = batch_sample['neg_img'].to(device)
pos_cls = batch_sample['pos_class'].to(device)
neg_cls = batch_sample['neg_class'].to(device)
with torch.set_grad_enabled(phase == 'train'):
# anc_embed, pos_embed and neg_embed are encoding(embedding) of image
anc_embed, pos_embed, neg_embed = model(anc_img), model(
pos_img), model(neg_img)
# choose the hard negatives only for "training"
pos_dist = l2_dist.forward(anc_embed, pos_embed)
neg_dist = l2_dist.forward(anc_embed, neg_embed)
all = (neg_dist - pos_dist <
args.margin).cpu().numpy().flatten()
if phase == 'train':
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
else:
hard_triplets = np.where(all >= 0)
anc_hard_embed = anc_embed[hard_triplets].to(device)
pos_hard_embed = pos_embed[hard_triplets].to(device)
neg_hard_embed = neg_embed[hard_triplets].to(device)
anc_hard_img = anc_img[hard_triplets].to(device)
pos_hard_img = pos_img[hard_triplets].to(device)
neg_hard_img = neg_img[hard_triplets].to(device)
pos_hard_cls = pos_cls[hard_triplets].to(device)
neg_hard_cls = neg_cls[hard_triplets].to(device)
anc_img_pred = model.forward_classifier(anc_hard_img).to(
device)
pos_img_pred = model.forward_classifier(pos_hard_img).to(
device)
neg_img_pred = model.forward_classifier(neg_hard_img).to(
device)
triplet_loss = TripletLoss(args.margin).forward(
anc_hard_embed, pos_hard_embed, neg_hard_embed).to(device)
if phase == 'train':
optimizer.zero_grad()
triplet_loss.backward()
optimizer.step()
dists = l2_dist.forward(anc_embed, pos_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.ones(dists.size(0)))
dists = l2_dist.forward(anc_embed, neg_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.zeros(dists.size(0)))
triplet_loss_sum += triplet_loss.item()
avg_triplet_loss = triplet_loss_sum / data_size[phase]
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array(
[subdist for dist in distances for subdist in dist])
tpr, fpr, accuracy, val, val_std, far = evaluate(distances, labels)
print(' {} set - Triplet Loss = {:.8f}'.format(
phase, avg_triplet_loss))
print(' {} set - Accuracy = {:.8f}'.format(
phase, np.mean(accuracy)))
with open('./log/{}_log_epoch{}.txt'.format(phase, epoch), 'w') as f:
f.write(
str(epoch) + '\t' + str(np.mean(accuracy)) + '\t' +
str(avg_triplet_loss))
if phase == 'train':
torch.save({
'epoch': epoch,
'state_dict': model.state_dict()
}, './log/checkpoint_epoch{}.pth'.format(epoch))
else:
plot_roc(fpr,
tpr,
figure_name='./log/roc_valid_epoch_{}.png'.format(epoch))
def train_valid(model, optimizer, scheduler, epoch, dataloaders, data_size):
for phase in ['train', 'valid']:
# One step for train or valid
labels, distances = [], []
triplet_loss_sum = 0.0
crossentropy_loss_sum = 0.0
accuracy_sum = 0.0
triplet_loss_sigma = 0.0
crossentropy_loss_sigma = 0.0
accuracy_sigma = 0.0
if phase == 'train':
scheduler.step()
model.train()
else:
model.eval()
for batch_idx, batch_sample in enumerate(dataloaders[phase]):
anc_img = batch_sample['anc_img'].to(device)
pos_img = batch_sample['pos_img'].to(device)
neg_img = batch_sample['neg_img'].to(device)
if (anc_img.shape[0] != cfg.batch_size
or pos_img.shape[0] != cfg.batch_size
or neg_img.shape[0] != cfg.batch_size):
print("Batch Size Not Equal")
continue
pos_cls = batch_sample['pos_class'].to(device)
neg_cls = batch_sample['neg_class'].to(device)
with torch.set_grad_enabled(phase == 'train'):
try:
# anc_embed, pos_embed and neg_embed are encoding(embedding) of image
anc_embed, pos_embed, neg_embed = model(anc_img), model(
pos_img), model(neg_img)
# choose the hard negatives only for "training"
pos_dist = l2_dist.forward(anc_embed, pos_embed)
neg_dist = l2_dist.forward(anc_embed, neg_embed)
all = (neg_dist - pos_dist <
cfg.margin).cpu().numpy().flatten()
if phase == 'train':
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
else:
hard_triplets = np.where(all >= 0)
if len(hard_triplets[0]) == 0:
continue
anc_hard_embed = anc_embed[hard_triplets].to(device)
pos_hard_embed = pos_embed[hard_triplets].to(device)
neg_hard_embed = neg_embed[hard_triplets].to(device)
anc_hard_img = anc_img[hard_triplets].to(device)
pos_hard_img = pos_img[hard_triplets].to(device)
neg_hard_img = neg_img[hard_triplets].to(device)
pos_hard_cls = pos_cls[hard_triplets].to(device)
neg_hard_cls = neg_cls[hard_triplets].to(device)
anc_img_pred = model.forward_classifier(anc_hard_img).to(
device)
pos_img_pred = model.forward_classifier(pos_hard_img).to(
device)
neg_img_pred = model.forward_classifier(neg_hard_img).to(
device)
triplet_loss = TL_loss.forward(anc_hard_embed,
pos_hard_embed,
neg_hard_embed).to(device)
triplet_loss *= cfg.triplet_lambuda
predicted_labels = torch.cat(
[anc_img_pred, pos_img_pred, neg_img_pred])
true_labels = torch.cat(
[pos_hard_cls, pos_hard_cls, neg_hard_cls]).squeeze()
crossentropy_loss = CE_loss(predicted_labels,
true_labels).to(device)
loss = triplet_loss + crossentropy_loss
if phase == 'train':
optimizer.zero_grad()
# triplet_loss.backward()
loss.backward()
optimizer.step()
if phase == 'valid':
pic_array, _ = TestData.get_data()
for i, pic in enumerate(pic_array):
pred = model.forward_classifier(
pic.unsqueeze(0).to(device)).to(device)
pred = torch.argmax(pred, 1).cpu().numpy()
# print(pred)
writer.add_image("Person {}/{}".format(pred[0], i),
pic, epoch)
_, predicted = torch.max(predicted_labels, 1)
correct = (predicted == true_labels).cpu().squeeze().sum(
).numpy() / (len(hard_triplets[0]) * 3)
dists = l2_dist.forward(anc_embed, pos_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.ones(dists.size(0)))
dists = l2_dist.forward(anc_embed, neg_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.zeros(dists.size(0)))
triplet_loss_sum += triplet_loss.item()
crossentropy_loss_sum += crossentropy_loss.item()
accuracy_sum += correct
triplet_loss_sigma += triplet_loss.item()
crossentropy_loss_sigma += crossentropy_loss.item()
accuracy_sigma += correct
if batch_idx % 10 == 0 and batch_idx != 0:
print(
'{} Inter {:4d}/{:4d} - Triplet Loss = {:.5f} - CrossEntropy Loss = {:.5f} - All Loss = {:.5f} - Accuaracy = {:.5f} len:{}'
.format(phase, batch_idx, len(dataloaders[phase]),
triplet_loss_sigma / 10,
crossentropy_loss_sigma / 10,
(triplet_loss_sigma +
crossentropy_loss_sigma) / 10,
accuracy_sigma / 10,
len(hard_triplets[0])))
triplet_loss_sigma = 0
crossentropy_loss_sigma = 0
accuracy_sigma = 0
except Exception as e:
print(e)
pass
avg_triplet_loss = triplet_loss_sum / int(
data_size[phase] / cfg.batch_size)
avg_crossentropy_loss = crossentropy_loss_sum / int(
data_size[phase] / cfg.batch_size)
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array(
[subdist for dist in distances for subdist in dist])
tpr, fpr, accuracy, val, val_std, far = evaluate(distances, labels)
print(' {} set - Triplet Loss = {:.8f}'.format(
phase, avg_triplet_loss))
print(' {} set - CrossEntropy Loss = {:.8f}'.format(
phase, avg_crossentropy_loss))
print(' {} set - All Loss = {:.8f}'.format(
phase, avg_triplet_loss + avg_crossentropy_loss))
print(' {} set - Accuracy = {:.8f}'.format(
phase, np.mean(accuracy)))
# 记录训练loss
writer.add_scalars('Loss/Triplet Loss Group'.format(phase),
{'{} triplet loss'.format(phase): avg_triplet_loss},
epoch)
writer.add_scalars(
'Loss/Crossentropy Loss Group'.format(phase),
{'{} crossentropy loss'.format(phase): avg_crossentropy_loss},
epoch)
writer.add_scalars('Loss/All Loss Group'.format(phase), {
'{} loss'.format(phase):
avg_triplet_loss + avg_crossentropy_loss
}, epoch)
writer.add_scalars('Accuracy_group'.format(phase),
{'{} accuracy'.format(phase): np.mean(accuracy)},
epoch)
# 记录learning rate
writer.add_scalar('learning rate', scheduler.get_lr()[0], epoch)
with open('./log/{}_log_epoch{}.txt'.format(phase, epoch), 'w') as f:
f.write(
str(epoch) + '\t' + str(np.mean(accuracy)) + '\t' +
str(avg_triplet_loss) + '\t' + str(avg_crossentropy_loss) +
'\t' + str(avg_triplet_loss + avg_crossentropy_loss))
if phase == 'train':
torch.save({
'epoch': epoch,
'state_dict': model.state_dict()
}, './log/checkpoint_epoch{}.pth'.format(epoch))
else:
plot_roc(fpr,
tpr,
figure_name='./log/roc_valid_epoch_{}.png'.format(epoch))
def train_model():
step = 0
log_file = open(log_file_path, 'w')
iteration = 0
lr = args.lr
top1 = AverageMeter()
best_model_wts = copy.deepcopy(model.state_dict())
best_acc = 0.0
train_loss = []
val_acc = []
for epoch in range(args.epochs):
# for phase in ['train']:
for phase in ['train', 'val']:
load_t0 = time.time()
if phase == 'train':
model.train()
else:
model.eval()
running_loss = 0.0
running_corrects = 0
# top1=AverageMeter()
if phase == 'train':
for batch_dix, (inputs,
labels) in enumerate(data_loaders[phase]):
#---(batch,3,96,96)
inputs = inputs.to(device)
labels = labels.to(device)
optimizer.zero_grad()
with torch.set_grad_enabled(phase == 'train'):
prediction = model.forward_classifier(inputs)
_, preds = torch.max(F.softmax(prediction), 1)
center_loss, model.centers = model.get_center_loss(
labels, args.alpha)
cross_entropy_loss = criterion(prediction, labels)
loss = args.center_loss_weight * center_loss + cross_entropy_loss
if phase == 'train':
loss.backward()
optimizer.step()
iteration += 1
# top1.update
if epoch in [10, 18]:
step += 1
# lr=adjust_learning_rate(optimizer, args.gamma,step,args.lr
# prec=accuracy(F.softmax(prediction), labels, topk=(1,))
# top1.update(prec[0], inputs.size(0))
load_t1 = time.time()
running_loss += loss.item() * inputs.size(0)
running_corrects += torch.sum(preds == labels.data)
if iteration % 10 == 0 and phase == 'train':
log_file.write('Epoch:' + repr(epoch) +
' || Totel iter ' + repr(iteration) +
' || L: {:.4f} A: {:.4f} || '.format(
loss.item() * inputs.size(0),
torch.sum(preds == labels.data)) +
'Batch time: {:.4f} || '.format(
load_t1 - load_t0) +
'LR: {:.8f}'.format(lr) + '\n')
train_loss.append(running_loss / dataset_sizes[phase])
epoch_loss = running_loss / dataset_sizes[phase]
epoch_acc = running_corrects.double() / dataset_sizes[phase]
print('epoch {} : {} Loss: {:.4f} Acc: {:.4f}'.format(
epoch, phase, epoch_loss, epoch_acc))
else:
targets, distances = [], []
for batch_dix, (data_a, data_p,
labels) in enumerate(data_loaders[phase]):
data_a, data_p = data_a.to(device), data_p.to(device)
labels = labels.to(device)
out_a, out_p = model(data_a), model(data_p)
dists = l2_dist.forward(out_a, out_p)
distances.append(dists.data.cpu().numpy())
targets.append(labels.data.cpu().numpy())
# _, _, plot_acc, _,_,_ = evaluate(dists.data.cpu().numpy().tolist(),labels.data.cpu().tolist())
targets = np.array(
[sublabel for label in targets for sublabel in label])
distances = np.array(
[subdist for dist in distances for subdist in dist])
tpr, fpr, acc, val, val_std, far = evaluate(distances, targets)
acc_max = acc.max()
val_acc.append(acc_max.max())
# embed()
print('epoch {} : {} Acc: {:.4f} '.format(
epoch, phase, acc_max))
log_file.write('epoch {} : {} Acc: {:.4f} '.format(
epoch, phase, acc_max))
plot_roc(
fpr,
tpr,
figure_name='./log/roc_valid_epoch_{}.png'.format(epoch))
if acc_max > best_acc:
best_acc = acc_max
best_model_wts = copy.deepcopy(model.state_dict())
plot_acc(train_loss, val_acc, './loss.png')
torch.save(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'centers': model.centers,
}, './centers_{}.pth'.format(best_acc))
