def validate(self, logger):
batch_time = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to evaluate mode
self.model.eval()
end = time.time()
for i, (images, labels) in enumerate(self.val_loader):
if check_gpu() > 0:
images = images.cuda(async=True)
labels = labels.cuda(async=True)
image_var = torch.autograd.Variable(images)
label_var = torch.autograd.Variable(labels)
# compute y_pred
y_pred = self.model(image_var)
if self.model_type == 'I3D':
y_pred = y_pred[0]
loss = self.criterion(y_pred, label_var)
# measure accuracy and record loss
prec1, prec5 = accuracy(y_pred.data, labels, topk=(1, 5))
losses.update(loss.item(), images.size(0))
acc.update(prec1.item(), images.size(0))
top1.update(prec1.item(), images.size(0))
top5.update(prec5.item(), images.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % self.print_freq == 0:
print('TrainVal: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
i,
len(self.val_loader),
batch_time=batch_time,
loss=losses,
top1=top1,
top5=top5))
print(' * Accuracy {acc.avg:.3f} Loss {loss.avg:.3f}'.format(
acc=acc, loss=losses))
logger.info(' * Accuracy {acc.avg:.3f} Loss {loss.avg:.3f}'.format(
acc=acc, loss=losses))
return losses, acc
def train(train_loader, model, criterion, optimizer, epoch, args):
"""
Train a proposed ResNet for classification
:param train_loader: default data_loader in pytorch
:param model: The proposed model, ResNet for our serup
:param criterion: criterion(loss) for optimization purpose
:param optimizer: to optimize model, adam or sgd is recommended
:param epoch: How many turns to train whole training set around
:param args: arguments for user input
:return:
"""
batch_time = AverageMeter('Time', ':6.3f')
data_time = AverageMeter('Data', ':6.3f')
losses = AverageMeter('Loss', ':.4e')
top1 = AverageMeter('Acc@1', ':6.2f')
top5 = AverageMeter('Acc@5', ':6.2f')
progress = ProgressMeter(len(train_loader),
[batch_time, data_time, losses, top1, top5],
prefix="Epoch: [{}]".format(epoch))
# switch to train mode
model.train()
end = time.time()
for i, (images, target) in enumerate(train_loader):
# measure data loading time
data_time.update(time.time() - end)
if args.gpu is not None:
images = images.cuda(args.gpu, non_blocking=True)
target = target.cuda(args.gpu, non_blocking=True)
# compute output
output = model(images)
loss = criterion(output, target)
# measure accuracy and record loss
acc1, acc5 = accuracy(output, target, topk=(1, 5))
losses.update(loss.item(), images.size(0))
top1.update(acc1[0], images.size(0))
top5.update(acc5[0], images.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % args.print_freq == 0:
progress.display(i)
def init_train(self, con_weight: float = 1.0):
test_img = self.get_test_image()
meter = AverageMeter("Loss")
self.writer.flush()
lr_scheduler = OneCycleLR(self.optimizer_G,
max_lr=0.9999,
steps_per_epoch=len(self.dataloader),
epochs=self.init_train_epoch)
for g in self.optimizer_G.param_groups:
g['lr'] = self.init_lr
for epoch in tqdm(range(self.init_train_epoch)):
meter.reset()
for i, (style, smooth, train) in enumerate(self.dataloader, 0):
# train = transform(test_img).unsqueeze(0)
self.G.zero_grad(set_to_none=self.grad_set_to_none)
train = train.to(self.device)
generator_output = self.G(train)
# content_loss = loss.reconstruction_loss(generator_output, train) * con_weight
content_loss = self.loss.content_loss(generator_output,
train) * con_weight
# content_loss = F.mse_loss(train, generator_output) * con_weight
content_loss.backward()
self.optimizer_G.step()
lr_scheduler.step()
meter.update(content_loss.detach())
self.writer.add_scalar(f"Loss : {self.init_time}",
meter.sum.item(), epoch)
self.write_weights(epoch + 1, write_D=False)
self.eval_image(epoch, f'{self.init_time} reconstructed img',
test_img)
for g in self.optimizer_G.param_groups:
g['lr'] = self.G_lr
def validate(val_loader, model, criterion, args):
"""
Call vaidate() to validate your result. Load validate data accordingly
:param val_loader: default data loader to load validation data
:param model: ResNet by default
:param criterion: the loss to compute
:param args: User defined input
:return:
"""
batch_time = AverageMeter('Time', ':6.3f')
losses = AverageMeter('Loss', ':.4e')
top1 = AverageMeter('Acc@1', ':6.2f')
top5 = AverageMeter('Acc@5', ':6.2f')
progress = ProgressMeter(len(val_loader), [batch_time, losses, top1, top5],
prefix='Test: ')
# switch to evaluate mode
model.eval()
with torch.no_grad():
end = time.time()
for i, (images, target) in enumerate(val_loader):
if args.gpu is not None:
images = images.cuda(args.gpu, non_blocking=True)
target = target.cuda(args.gpu, non_blocking=True)
# compute output
output = model(images)
loss = criterion(output, target)
# measure accuracy and record loss
acc1, acc5 = accuracy(output, target, topk=(1, 5))
losses.update(loss.item(), images.size(0))
top1.update(acc1[0], images.size(0))
top5.update(acc5[0], images.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if (i + 1) % args.print_freq == 0:
progress.display(i)
print(
' Validation finished! Avg stats: Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}'
.format(top1=top1, top5=top5))
return top1.avg, top5.avg
def process(self):
acc = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
losses = AverageMeter()
log_file = os.path.join(self.data_folder, 'test.log')
logger = Logger('test', log_file)
# switch to evaluate mode
self.model.eval()
start_time = time.clock()
print("Begin testing")
for i, (images, labels) in enumerate(self.test_loader):
if check_gpu() > 0:
images = images.cuda(async=True)
labels = labels.cuda(async=True)
image_var = torch.autograd.Variable(images)
label_var = torch.autograd.Variable(labels)
# compute y_pred
y_pred = self.model(image_var)
loss = self.criterion(y_pred, label_var)
# measure accuracy and record loss
prec1, prec5 = accuracy(y_pred.data, labels, topk=(1, 5))
losses.update(loss.item(), images.size(0))
acc.update(prec1.item(), images.size(0))
top1.update(prec1.item(), images.size(0))
top5.update(prec5.item(), images.size(0))
if i % self.print_freq == 0:
print('TestVal: [{0}/{1}]\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
i, len(self.test_loader), loss=losses, top1=top1, top5=top5))
print(
' * Accuracy {acc.avg:.3f} Acc@5 {top5.avg:.3f} Loss {loss.avg:.3f}'.format(acc=acc, top5=top5,
loss=losses))
end_time = time.clock()
print("Total testing time %.2gs" % (end_time - start_time))
logger.info("Total testing time %.2gs" % (end_time - start_time))
logger.info(
' * Accuracy {acc.avg:.3f} Acc@5 {top5.avg:.3f} Loss {loss.avg:.3f}'.format(acc=acc, top5=top5,
loss=losses))
def main(args):
# NTU Dataset
dataset = NTU(
root=args.root,
w=args.width,
h=args.height,
t=args.time,
dataset='train',
train=True,
avi_dir=args.avi_dir,
usual_transform=False,
)
# Pytorch dataloader
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
num_workers=args.workers,
pin_memory=args.cuda,
collate_fn=my_collate)
# Loop
data_time = AverageMeter()
start_data = time.time()
for i, dict_input in enumerate(dataloader):
duration_data = time.time() - start_data
data_time.update(duration_data)
# Get the data
clip, skeleton = dict_input['clip'], dict_input[
'skeleton'] # (B, C, T, 224, 224), (B, T, 2, 25, 2)
# Show
show_one_img(clip[0, :, 0], skeleton[0, 0])
print("{}/{} : {time.val:.3f} ({time.avg:.3f}) sec/batch".format(
i + 1, len(dataloader), time=data_time))
sys.stdout.flush()
start_data = time.time()
def validate(val_loader, model, criterion, epoch, start_time):
timer = TimeMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
model.eval()
eval_start_time = time.time()
for i, (input, target) in enumerate(val_loader):
if args.short_epoch and (i > 10): break
batch_num = i + 1
timer.batch_start()
if args.distributed:
top1acc, top5acc, loss, batch_total = distributed_predict(
input, target, model, criterion)
else:
with torch.no_grad():
output = model(input)
loss = criterion(output, target).data
batch_total = input.size(0)
top1acc, top5acc = accuracy(output.data, target, topk=(1, 5))
# Eval batch done. Logging results
timer.batch_end()
losses.update(to_python_float(loss), to_python_float(batch_total))
top1.update(to_python_float(top1acc), to_python_float(batch_total))
top5.update(to_python_float(top5acc), to_python_float(batch_total))
should_print = (batch_num % args.print_freq
== 0) or (batch_num == len(val_loader))
if args.local_rank == 0 and should_print:
output = (
f'Test: [{epoch}][{batch_num}/{len(val_loader)}]\t'
f'Time {timer.batch_time.val:.3f} ({timer.batch_time.avg:.3f})\t'
f'Loss {losses.val:.4f} ({losses.avg:.4f})\t'
f'Acc@1 {top1.val:.3f} ({top1.avg:.3f})\t'
f'Acc@5 {top5.val:.3f} ({top5.avg:.3f})')
log.verbose(output)
tb.log_eval(top1.avg, top5.avg, time.time() - eval_start_time)
tb.log('epoch', epoch)
return top1.avg, top5.avg
def train(trn_loader, model, criterion, optimizer, scheduler, epoch):
net_meter = NetworkMeter()
timer = TimeMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to train mode
model.train()
for i, (input, target) in enumerate(trn_loader):
if args.short_epoch and (i > 10): break
batch_num = i + 1
timer.batch_start()
scheduler.update_lr(epoch, i + 1, len(trn_loader))
# compute output
output = model(input)
loss = criterion(output, target)
should_print = (batch_num % args.print_freq
== 0) or (batch_num == len(trn_loader))
# compute gradient and do SGD step
if args.fp16:
loss = loss * args.loss_scale
# zero_grad() and converting fp16/fp32 is handled in optimizer
loss.backward()
optimizer.step(wait_for_finish=should_print)
loss = loss / args.loss_scale
else:
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Train batch done. Logging results
timer.batch_end()
if args.local_rank == 0 and should_print:
corr1, corr5 = correct(output.data, target, topk=(1, 5))
reduced_loss, batch_total = to_python_float(
loss.data), to_python_float(input.size(0))
if args.distributed: # Must keep track of global batch size, since not all machines are guaranteed equal batches at the end of an epoch
validate_tensor[0] = batch_total
validate_tensor[1] = reduced_loss
validate_tensor[2] = corr1
validate_tensor[3] = corr5
batch_total, reduced_loss, corr1, corr5 = bps.push_pull(
validate_tensor, average=False, name="validation_tensor")
batch_total = batch_total.cpu().numpy()
reduced_loss = reduced_loss.cpu().numpy()
corr1 = corr1.cpu().numpy()
corr5 = corr5.cpu().numpy()
reduced_loss = reduced_loss / bps.size()
top1acc = to_python_float(corr1) * (100.0 / batch_total)
top5acc = to_python_float(corr5) * (100.0 / batch_total)
losses.update(reduced_loss, batch_total)
top1.update(top1acc, batch_total)
top5.update(top5acc, batch_total)
tb.log_memory()
tb.log_trn_times(timer.batch_time.val, timer.data_time.val,
input.size(0))
tb.log_trn_loss(losses.val, top1.val, top5.val)
recv_gbit, transmit_gbit = net_meter.update_bandwidth()
tb.log("sizes/batch_total", batch_total)
tb.log('net/recv_gbit', recv_gbit)
tb.log('net/transmit_gbit', transmit_gbit)
output = (
f'Epoch: [{epoch}][{batch_num}/{len(trn_loader)}]\t'
f'Time {timer.batch_time.val:.3f} ({timer.batch_time.avg:.3f})\t'
f'Loss {losses.val:.4f} ({losses.avg:.4f})\t'
f'Acc@1 {top1.val:.3f} ({top1.avg:.3f})\t'
f'Acc@5 {top5.val:.3f} ({top5.avg:.3f})\t'
f'Data {timer.data_time.val:.3f} ({timer.data_time.avg:.3f})\t'
f'BW {recv_gbit:.3f} {transmit_gbit:.3f}')
log.verbose(output)
tb.update_step_count(batch_total)
class DataGenerator(object):
def __init__(self,
MovementModule='default',
EnvModule='default',
TextModule='default',
WindowHeight=720,
WindowWidth=1080,
MaxTextNum=15,
DataStoragePath='../../../GeneratedData/DataFraction_1',
camera_anchor_filepath='./camera_anchors/urbancity.txt',
EnvName='',
anchor_freq=10,
max_emissive=5,
FontSize=[8, 16],
use_real_img=0.1,
is_debug=True,
languages=["Latin"],
HighResFactor=2.0,
UnrealProjectName="./",
**kwargs):
self.client = WrappedClient(UnrealCVClient, DataStoragePath,
HighResFactor, UnrealProjectName)
self.DataStoragePath = DataStoragePath
self.UnrealProjectName = UnrealProjectName
self.WindowHeight = WindowHeight
self.WindowWidth = WindowWidth
self.MaxTextNum = MaxTextNum
self.is_debug = is_debug
self.camera_anchor_filepath = camera_anchor_filepath
self.anchor_freq = anchor_freq
self.HighResFactor = HighResFactor
self.RootPath = opa(DataStoragePath)
while os.path.isdir(self.RootPath):
root_path, count = self.RootPath.split('_')
self.RootPath = root_path + '_' + str(int(count) + 1)
print(f"Data will be saved to: {self.RootPath}")
self.LabelPath = osp.join(self.RootPath, 'Label.json')
self.DataLabel = None
self.ImgFolder = osp.join(self.RootPath, 'imgs')
self.LabelFolder = osp.join(self.RootPath, 'labels')
self.WordFolder = osp.join(self.RootPath, 'WordCrops')
self.DataCount = 0
self.isConnected = False
self.SaveFreq = 100
# step 1
self._InitializeDataStorage()
# step 2
if len(EnvName) > 0:
StartEngine(EnvName)
self._ConnectToGame()
# step 3 set resolution & rotation
self.client.setres(self.WindowWidth, self.WindowHeight)
# self.client.setCameraRotation(0, 0, 0)
self.EnvDepth = kwargs.get('EnvDepth', 100)
# load modules
self.Wanderer = CameraSet[MovementModule](
client=self.client,
camera_anchor_filepath=self.camera_anchor_filepath,
anchor_freq=self.anchor_freq)
self.EnvRenderer = EnvSet[EnvModule](client=self.client)
self.TextPlacer = TextPlacement[TextModule](
client=self.client,
MaxTextCount=self.MaxTextNum,
ContentPath=osp.join(self.RootPath, 'WordCrops'),
max_emissive=max_emissive,
FontSize=FontSize,
is_debug=is_debug,
use_real_img=use_real_img,
languages=languages,
HighResFactor=HighResFactor)
# initializer meters
self.camera_meter = AverageMeter()
self.env_meter = AverageMeter()
self.text_meter = AverageMeter()
self.retrieve_label_meter = AverageMeter()
self.save_label_meter = AverageMeter()
self.save_meter = AverageMeter()
self._cleanup()
def __del__(self):
if self.client.isconnected():
self.client.QuitGame()
self.client.disconnect()
self._cleanup()
# os.system('~/cache_light.png')
def _cleanup(self):
os.system(
f'rm ../../../PackagedEnvironment/{self.UnrealProjectName}/Demo/Saved/Screenshots/LinuxNoEditor/*png'
)
#os.system(f'rm ../../../PackagedEnvironment/{self.UnrealProjectName}/Demo/Saved/Logs/*')
def _InitializeDataStorage(self):
os.makedirs(self.ImgFolder, exist_ok=True)
os.makedirs(self.LabelFolder, exist_ok=True)
os.makedirs(self.WordFolder, exist_ok=True)
self.DataCount = 0
self.DataLabel = []
os.system(f'cp vis.py {self.RootPath}/')
def _ConnectToGame(self):
# wait and connect
sleepTime = 1.0
while True:
self.client.connect()
self.isConnected = self.client.isconnected()
if self.isConnected:
break
else:
if sleepTime > 120:
break
time.sleep(sleepTime)
sleepTime *= 2
if not self.isConnected:
print('Failed to connect to UnrealCV server.')
sys.exit(-1)
def _GenerateOneImageInstance(self,
step_count,
force_change_camera_anchor=False):
# step 1: move around
time_stamp = time.time()
if not self.is_debug:
self.Wanderer.step(
height=self.WindowHeight,
width=self.WindowWidth,
step=step_count,
force_change_camera_anchor=force_change_camera_anchor)
time_stamp = self.camera_meter.update(time.time() - time_stamp)
# step 2: render env
self.EnvRenderer.step()
time_stamp = self.env_meter.update(time.time() - time_stamp)
# step 3: place text
self.TextPlacer.PutTextStep()
time_stamp = self.text_meter.update(time.time() - time_stamp)
if self.is_debug:
print('Text Loaded, ready to retrieve data')
# step 4: retrieve data
img_path, Texts, WordBoxes, CharBoxes, TextNum = self.TextPlacer.RetrieveDataStep(
osp.join(self.ImgFolder, f'{self.DataCount}.jpg'))
time_stamp = self.retrieve_label_meter.update(time.time() - time_stamp)
force_change_camera_anchor = TextNum == 0
DataLabel = {
'imgfile': f'imgs/{self.DataCount}.jpg',
'bbox': WordBoxes,
'cbox': CharBoxes,
'text': Texts,
'is_difficult': [0 for _ in range(len(WordBoxes))]
}
json.dump(
DataLabel,
open(osp.join(self.LabelFolder,
str(self.DataCount) + '.json'), 'w'))
time_stamp = self.save_label_meter.update(time.time() - time_stamp)
if self.is_debug:
print('Finished waiting, ready to save img')
self.client.SaveImg(img_path)
time_stamp = self.save_meter.update(time.time() - time_stamp)
self.DataCount += 1
if self.is_debug:
# step 5: visualize
ShowImgAndAnnotation(img_path, Texts, WordBoxes, CharBoxes)
time_stamp = time.time()
return {'force_change_camera_anchor': force_change_camera_anchor}
def StartGeneration(self, IterationNum=10000, sleep_time=0, sleep_freq=1):
status = {'force_change_camera_anchor': False}
for Count in range(IterationNum):
status = self._GenerateOneImageInstance(Count, **status)
if Count % self.anchor_freq == 0:
print(f"{Count} images created. Timing:")
print(f' ----- camera: {self.camera_meter}')
print(f' ----- env: {self.env_meter}')
print(f' ----- text: {self.text_meter}')
print(
f' ----- retrieve label: {self.retrieve_label_meter}'
)
print(f' ----- save label: {self.save_label_meter}')
print(f' ----- retrieve image: {self.save_meter}')
def train(trn_loader, model, criterion, optimizer, scheduler, epoch):
net_meter = NetworkMeter()
timer = TimeMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to train mode
model.train()
for i, (input, target) in enumerate(trn_loader):
if args.short_epoch and (i > 10): break
batch_num = i + 1
timer.batch_start()
scheduler.update_lr(epoch, i + 1, len(trn_loader))
# compute output
output = model(input)
loss = criterion(output, target)
# compute gradient and do SGD step
if args.fp16:
loss = loss * args.loss_scale
model.zero_grad()
loss.backward()
model_grads_to_master_grads(model_params, master_params)
for param in master_params:
param.grad.data = param.grad.data / args.loss_scale
optimizer.step()
master_params_to_model_params(model_params, master_params)
loss = loss / args.loss_scale
else:
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Train batch done. Logging results
timer.batch_end()
corr1, corr5 = correct(output.data, target, topk=(1, 5))
reduced_loss, batch_total = to_python_float(
loss.data), to_python_float(input.size(0))
if args.distributed: # Must keep track of global batch size, since not all machines are guaranteed equal batches at the end of an epoch
metrics = torch.tensor([batch_total, reduced_loss, corr1,
corr5]).float().cuda()
batch_total, reduced_loss, corr1, corr5 = dist_utils.sum_tensor(
metrics).cpu().numpy()
reduced_loss = reduced_loss / dist_utils.env_world_size()
top1acc = to_python_float(corr1) * (100.0 / batch_total)
top5acc = to_python_float(corr5) * (100.0 / batch_total)
losses.update(reduced_loss, batch_total)
top1.update(top1acc, batch_total)
top5.update(top5acc, batch_total)
should_print = (batch_num % args.print_freq
== 0) or (batch_num == len(trn_loader))
if args.local_rank == 0 and should_print:
tb.log_memory()
tb.log_trn_times(timer.batch_time.val, timer.data_time.val,
input.size(0))
tb.log_trn_loss(losses.val, top1.val, top5.val)
recv_gbit, transmit_gbit = net_meter.update_bandwidth()
tb.log("sizes/batch_total", batch_total)
tb.log('net/recv_gbit', recv_gbit)
tb.log('net/transmit_gbit', transmit_gbit)
output = (
f'Epoch: [{epoch}][{batch_num}/{len(trn_loader)}]\t'
f'Time {timer.batch_time.val:.3f} ({timer.batch_time.avg:.3f})\t'
f'Loss {losses.val:.4f} ({losses.avg:.4f})\t'
f'Acc@1 {top1.val:.3f} ({top1.avg:.3f})\t'
f'Acc@5 {top5.val:.3f} ({top5.avg:.3f})\t'
f'Data {timer.data_time.val:.3f} ({timer.data_time.avg:.3f})\t'
f'BW {recv_gbit:.3f} {transmit_gbit:.3f}')
log.verbose(output)
tb.update_step_count(batch_total)
def process(self):
acc = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
losses = AverageMeter()
log_file = os.path.join(self.data_folder, 'test.log')
logger = Logger('test', log_file)
# switch to evaluate mode
self.model.eval()
start_time = time.clock()
print("Begin testing")
predicted, probs = [], []
for i, (images, labels) in enumerate(self.test_loader):
if check_gpu() > 0:
images = images.cuda(async=True)
labels = labels.cuda(async=True)
images = torch.autograd.Variable(images)
labels = torch.autograd.Variable(labels)
if self.tencrop:
# Due to ten-cropping, input batch is a 5D Tensor
batch_size, number_of_crops, number_of_channels, height, width = images.size(
)
# Fuse batch size and crops
images = images.view(-1, number_of_channels, height, width)
# Compute model output
output_batch_crops = self.model(images)
# Average predictions for each set of crops
output_batch = output_batch_crops.view(batch_size,
number_of_crops,
-1).mean(1)
label_repeated = labels.repeat(10, 1).transpose(
1, 0).contiguous().view(-1, 1).squeeze()
loss = self.criterion(output_batch_crops, label_repeated)
else:
output_batch = self.model(images)
loss = self.criterion(output_batch, labels)
# measure accuracy and record loss
prec1, prec5 = accuracy(output_batch.data, labels, topk=(1, 5))
# print(prec1, prec5)
losses.update(loss.item(), images.size(0))
acc.update(prec1.item(), images.size(0))
top1.update(prec1.item(), images.size(0))
top5.update(prec5.item(), images.size(0))
if i % self.print_freq == 0:
print('TestVal: [{0}/{1}]\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
i,
len(self.test_loader),
loss=losses,
top1=top1,
top5=top5))
print(
' * Accuracy {acc.avg:.3f} Acc@5 {top5.avg:.3f} Loss {loss.avg:.3f}'
.format(acc=acc, top5=top5, loss=losses))
end_time = time.clock()
print("Total testing time %.2gs" % (end_time - start_time))
logger.info("Total testing time %.2gs" % (end_time - start_time))
logger.info(' * Accuracy {acc.avg:.3f} Loss {loss.avg:.3f}'.format(
acc=acc, top5=top5, loss=losses))
def train(args):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
tfboard_writer = SummaryWriter()
logname = '{}'.format(datetime.datetime.now().strftime('%Y-%m-%d-%H:%M'))
logger = Logger(args.save_path, logname)
logger.log('Arguments : -------------------------------')
for name, value in args._get_kwargs():
logger.log('{:16} : {:}'.format(name, value))
# Data Augmentation
mean_fill = tuple([int(x * 255) for x in [0.485, 0.456, 0.406]])
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_transform = [
transforms.AugTransBbox(args.transbbox_prob, args.transbbox_percent)
]
train_transform += [transforms.PreCrop(args.pre_crop_expand)]
train_transform += [
transforms.TrainScale2WH((args.crop_width, args.crop_height))
]
#train_transform += [transforms.AugHorizontalFlip(args.flip_prob)]
#train_transform += [transforms.AugScale(args.scale_prob, args.scale_min, args.scale_max)]
#train_transform += [transforms.AugCrop(args.crop_width, args.crop_height, args.crop_perturb_max, mean_fill)]
if args.rotate_max:
train_transform += [transforms.AugRotate(args.rotate_max)]
train_transform += [
transforms.AugGaussianBlur(args.gaussianblur_prob,
args.gaussianblur_kernel_size,
args.gaussianblur_sigma)
]
train_transform += [transforms.ToTensor(), normalize]
train_transform = transforms.Compose(train_transform)
eval_transform = transforms.Compose([
transforms.PreCrop(args.pre_crop_expand),
transforms.TrainScale2WH((args.crop_width, args.crop_height)),
transforms.ToTensor(), normalize
])
# Training datasets
train_data = GeneralDataset(args.num_pts, train_transform,
args.train_lists)
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
# Evaluation Dataloader
eval_loaders = []
for eval_ilist in args.eval_lists:
eval_idata = GeneralDataset(args.num_pts, eval_transform, eval_ilist)
eval_iloader = torch.utils.data.DataLoader(eval_idata,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
eval_loaders.append(eval_iloader)
net = Model(args.num_pts)
logger.log("=> network :\n {}".format(net))
logger.log('arguments : {:}'.format(args))
optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad,
net.parameters()),
lr=args.LR,
momentum=args.momentum,
weight_decay=args.decay,
nesterov=args.nesterov)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
milestones=args.schedule,
gamma=args.gamma)
criterion = wing_loss(args)
# criterion = torch.nn.MSELoss(reduce=True)
net = net.cuda()
criterion = criterion.cuda()
net = torch.nn.DataParallel(net)
last_info = logger.last_info()
if last_info.exists():
logger.log("=> loading checkpoint of the last-info '{:}' start".format(
last_info))
last_info = torch.load(last_info)
start_epoch = last_info['epoch'] + 1
checkpoint = torch.load(last_info['last_checkpoint'])
assert last_info['epoch'] == checkpoint[
'epoch'], 'Last-Info is not right {:} vs {:}'.format(
last_info, checkpoint['epoch'])
net.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
logger.log("=> load-ok checkpoint '{:}' (epoch {:}) done".format(
logger.last_info(), checkpoint['epoch']))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch = 0
for epoch in range(start_epoch, args.epochs):
scheduler.step()
net.train()
# train
img_prediction = []
img_target = []
train_losses = AverageMeter()
for i, (inputs, target) in enumerate(train_loader):
target = target.squeeze(1)
inputs = inputs.cuda()
target = target.cuda()
#print(inputs.size())
#ssert 1==0
prediction = net(inputs)
loss = criterion(prediction, target)
train_losses.update(loss.item(), inputs.size(0))
prediction = prediction.detach().to(torch.device('cpu')).numpy()
target = target.detach().to(torch.device('cpu')).numpy()
for idx in range(inputs.size()[0]):
img_prediction.append(prediction[idx, :])
img_target.append(target[idx, :])
optimizer.zero_grad()
loss.backward()
optimizer.step()
if i % args.print_freq == 0 or i + 1 == len(train_loader):
logger.log(
'[train Info]: [epoch-{}-{}][{:04d}/{:04d}][Loss:{:.2f}]'.
format(epoch, args.epochs, i, len(train_loader),
loss.item()))
train_nme = compute_nme(args.num_pts, img_prediction, img_target)
logger.log('epoch {:02d} completed!'.format(epoch))
logger.log(
'[train Info]: [epoch-{}-{}][Avg Loss:{:.6f}][NME:{:.2f}]'.format(
epoch, args.epochs, train_losses.avg, train_nme * 100))
tfboard_writer.add_scalar('Average Loss', train_losses.avg, epoch)
tfboard_writer.add_scalar('NME', train_nme * 100,
epoch) # traing data nme
# save checkpoint
filename = 'epoch-{}-{}.pth'.format(epoch, args.epochs)
save_path = logger.path('model') / filename
torch.save(
{
'epoch': epoch,
'args': deepcopy(args),
'state_dict': net.state_dict(),
'scheduler': scheduler.state_dict(),
'optimizer': optimizer.state_dict(),
},
logger.path('model') / filename)
logger.log('save checkpoint into {}'.format(filename))
last_info = torch.save({
'epoch': epoch,
'last_checkpoint': save_path
}, logger.last_info())
# eval
logger.log('Basic-Eval-All evaluates {} dataset'.format(
len(eval_loaders)))
for i, loader in enumerate(eval_loaders):
eval_losses = AverageMeter()
eval_prediction = []
eval_target = []
with torch.no_grad():
net.eval()
for i_batch, (inputs, target) in enumerate(loader):
target = target.squeeze(1)
inputs = inputs.cuda()
target = target.cuda()
prediction = net(inputs)
loss = criterion(prediction, target)
eval_losses.update(loss.item(), inputs.size(0))
prediction = prediction.detach().to(
torch.device('cpu')).numpy()
target = target.detach().to(torch.device('cpu')).numpy()
for idx in range(inputs.size()[0]):
eval_prediction.append(prediction[idx, :])
eval_target.append(target[idx, :])
if i_batch % args.print_freq == 0 or i + 1 == len(loader):
logger.log(
'[Eval Info]: [epoch-{}-{}][{:04d}/{:04d}][Loss:{:.2f}]'
.format(epoch, args.epochs, i, len(loader),
loss.item()))
eval_nme = compute_nme(args.num_pts, eval_prediction, eval_target)
logger.log(
'[Eval Info]: [evaluate the {}/{}-th dataset][epoch-{}-{}][Avg Loss:{:.6f}][NME:{:.2f}]'
.format(i, len(eval_loaders), epoch, args.epochs,
eval_losses.avg, eval_nme * 100))
tfboard_writer.add_scalar('eval_nme/{}'.format(i), eval_nme * 100,
epoch)
logger.close()
def train(self, logger, epoch):
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
rate = get_learning_rate(self.optimizer)[0]
# switch to train mode
self.model.train()
end = time.time()
for i, (images, target) in enumerate(self.train_loader):
# adjust learning rate scheduler step
self.scheduler.batch_step()
# measure data loading time
data_time.update(time.time() - end)
if check_gpu() > 0:
images = images.cuda(async=True)
target = target.cuda(async=True)
image_var = torch.autograd.Variable(images)
label_var = torch.autograd.Variable(target)
self.optimizer.zero_grad()
# compute y_pred
y_pred = self.model(image_var)
if self.model_type == 'I3D':
y_pred = y_pred[0]
loss = self.criterion(y_pred, label_var)
# measure accuracy and record loss
prec1, prec5 = accuracy(y_pred.data, target, topk=(1, 5))
losses.update(loss.item(), images.size(0))
acc.update(prec1.item(), images.size(0))
top1.update(prec1.item(), images.size(0))
top5.update(prec5.item(), images.size(0))
# compute gradient and do SGD step
loss.backward()
self.optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % self.print_freq == 0:
print('Epoch: [{0}/{1}][{2}/{3}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Lr {rate:.5f}\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
epoch,
self.epochs,
i,
len(self.train_loader),
batch_time=batch_time,
data_time=data_time,
rate=rate,
loss=losses,
top1=top1,
top5=top5))
logger.info('Epoch: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Lr {rate:.5f}\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
epoch,
self.epochs,
batch_time=batch_time,
data_time=data_time,
rate=rate,
loss=losses,
top1=top1,
top5=top5))
return losses, acc
def validate(self, logger):
batch_time = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to evaluate mode
self.model.eval()
end = time.time()
for i, (images, labels) in enumerate(self.val_loader):
if check_gpu() > 0:
images = images.cuda(async=True)
labels = labels.cuda(async=True)
images = torch.autograd.Variable(images)
labels = torch.autograd.Variable(labels)
if self.tencrop:
# Due to ten-cropping, input batch is a 5D Tensor
batch_size, number_of_crops, number_of_channels, height, width = images.size(
)
# Fuse batch size and crops
images = images.view(-1, number_of_channels, height, width)
# Compute model output
output_batch_crops = self.model(images)
# Average predictions for each set of crops
output_batch = output_batch_crops.view(batch_size,
number_of_crops,
-1).mean(1)
label_repeated = labels.repeat(10, 1).transpose(
1, 0).contiguous().view(-1, 1).squeeze()
loss = self.criterion(output_batch_crops, label_repeated)
else:
output_batch = self.model(images)
loss = self.criterion(output_batch, labels)
# measure accuracy and record loss
prec1, prec5 = accuracy(output_batch.data, labels, topk=(1, 5))
losses.update(loss.item(), images.size(0))
acc.update(prec1.item(), images.size(0))
top1.update(prec1.item(), images.size(0))
top5.update(prec5.item(), images.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % self.print_freq == 0:
print('TrainVal: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
i,
len(self.val_loader),
batch_time=batch_time,
loss=losses,
top1=top1,
top5=top5))
print(' * Accuracy {acc.avg:.3f} Loss {loss.avg:.3f}'.format(
acc=acc, loss=losses))
logger.info(' * Accuracy {acc.avg:.3f} Loss {loss.avg:.3f}'.format(
acc=acc, loss=losses))
return losses, acc
drop_last=True,
num_workers=opt.num_workers)
# setup tracker
tracker = TrackerSiamFC(name=opt.name,
weight=opt.weight,
device=opt.device)
# training loop
itr = 0
num_itrs = int((opt.num_epochs * len(loader)) / opt.print_freq) + 1
loss_logger = Logger(os.path.join(opt.log_dir, 'loss.csv'), num_itrs)
loss_meter = AverageMeter()
for epoch in range(opt.num_epochs):
for step, batch in enumerate(loader):
loss = tracker.step(batch, backward=True, update_lr=(step == 0))
itr += 1
loss_meter.update(loss)
if itr % opt.print_freq == 0:
print('Epoch [{}/{}] itr [{}]: Loss: {:.5f}'.format(
epoch + 1, opt.num_epochs, itr, loss_meter.avg))
sys.stdout.flush()
loss_logger.set(itr / opt.print_freq, loss_meter.avg)
loss_meter = AverageMeter()
# save checkpoint
net_path = os.path.join(opt.log_dir, 'model_e%d.pth' % (epoch + 1))
torch.save(tracker.net.state_dict(), net_path)
def train_walk(walk_file,
w,
data_batches,
valid_batches,
model,
num_epochs,
verbose=False):
# for param in model.parameters():
# param.requires_grad = False # freeze the model
print("START TRAINING:", walk_file)
opt = optim.SGD([w], lr=0.01)
# opt = optim.Adam([w], lr=0.01, momentum=0.9)
start_time = time.perf_counter()
meter = AverageMeter()
loss_hist_before = []
loss_hist_during = []
for e in range(num_epochs):
avg_loss_before = average_loss(w, data_batches, model, verbose)
model.train()
total_loss = 0
nsents = 0
meter.clear()
indices = list(range(len(data_batches)))
random.shuffle(indices)
for i, idx in enumerate(indices):
opt.zero_grad()
x, x_edit = data_batches[idx]
# encode the input x
mu, logvar = model.encode(x)
z = reparameterize(mu, logvar)
# add w to compute new latent
new_latent = z + alpha * w
# decode the new latent
logits, hidden = model.decode(new_latent, x)
# compute the loss wrt to the edit
loss = model.loss_rec(logits, x_edit).mean()
#print("LOSS", idx, ":", loss)
loss.backward()
opt.step()
total_loss += loss * x.shape[1]
nsents += x.shape[1]
meter.update(loss, x.shape[1])
print("---------------------------")
avg_loss_after = average_loss(w, data_batches, model)
print("FINISHED EPOCH", e)
print("avg loss before:", avg_loss_before)
print("avg train loss: ", total_loss / nsents)
# print("meter loss", meter.avg)
loss_hist_before.append((e, avg_loss_before.item()))
loss_hist_during.append((e, meter.avg.item()))
if verbose:
print("loss", loss)
print("nsents", nsents)
val_loss = average_loss(w, valid_batches, model, False)
print("avg valid loss: ", val_loss)
epoch_time = time.perf_counter()
print("time: ", epoch_time - start_time)
print("=" * 60)
#print(torch.cuda.memory_summary(device=None, abbreviated=False))
print("FINISHED TRAINING")
best_before_loss = min(loss_hist_before, key=lambda x: x[1])
best_during_loss = min(loss_hist_during, key=lambda x: x[1])
print("best_before_loss:", best_before_loss,
loss_hist_during[best_before_loss[0]])
print("best_during_loss:", best_during_loss,
loss_hist_before[best_during_loss[0]])
plot_series([loss_hist_before, loss_hist_during], walk_file)
print(w)
torch.save(w, results_dir + walk_file)
return w
def average_loss(w, data_batches, model, verbose=False):
meter = AverageMeter()
model.eval()
with torch.no_grad():
total_loss = 0
B = len(data_batches)
nsents = 0
for idx in range(len(data_batches)):
x, x_edit = data_batches[idx]
mu, logvar = model.encode(x)
z = reparameterize(mu, logvar)
new_latent = z + alpha * w
logits, hidden = model.decode(new_latent, x)
loss = model.loss_rec(logits, x_edit).mean()
if verbose:
# losses = model.autoenc(x, x_edit)
# print("autoenc", idx, ":", losses['rec'], "shapes", x.shape, x_edit.shape)
print("my loss", idx, ":", loss)
print("x", x.shape, "| x_edit", x_edit.shape)
sents = []
edited_sents = []
walk_sents = []
batch_len = x.shape[1]
max_len = 35
dec = 'greedy'
outputs = model.generate(new_latent, max_len, dec).t()
for i in range(batch_len):
x_i = x[:, i]
sents.append([vocab.idx2word[id] for id in x_i])
xe_i = x_edit[:, i]
edited_sents.append([vocab.idx2word[id] for id in xe_i])
output_i = outputs[i]
walk_sents.append([vocab.idx2word[id] for id in output_i])
for i in range(batch_len):
x_i = torch.unsqueeze(x[:, i], dim=1)
xe_i = torch.unsqueeze(x_edit[:, i], dim=1)
loss_i = compute_loss(w, x_i, xe_i, model)
print("batch", idx, ":", loss, "| sentence", i, ":",
loss_i)
print("--SENT:", sents[i])
print(x[:, i])
print("--EDIT:", edited_sents[i])
print(x_edit[:, i])
print("--WALK:", walk_sents[i])
print(outputs[i])
if print_outputs_flag:
if idx == 4:
print("batch", idx, "length", x.shape[1])
edited_sents = []
walked_sents = []
sents = []
max_len = 35
dec = 'greedy'
outputs = model.generate(new_latent, max_len, dec).t()
print("outputs", outputs.shape)
print("x", x.shape)
print("x_edit", x_edit.shape)
print("z", z.shape)
for i in range(batch_len):
output_i = outputs[i]
walked_sents.append(
[vocab.idx2word[id] for id in output_i])
x_i = x[:, i]
sents.append([vocab.idx2word[id] for id in x_i])
xe_i = x_edit[:, i]
edited_sents.append(
[vocab.idx2word[id] for id in xe_i])
walked_sents = strip_eos(walked_sents)
edited_sents = strip_eos(edited_sents)
sents = strip_eos(sents)
for i in range(batch_len):
print(i)
print("--SENT:", sents[i])
print("--EDIT:", edited_sents[i])
print("--WALK:", walked_sents[i])
total_loss += loss * x.shape[1]
nsents += x.shape[1]
#breakpoint()
meter.update(loss.item(), x.shape[1])
avg_loss = total_loss / nsents
if verbose:
print("avg_loss meter loss vs avg_loss", meter.avg, avg_loss)
#print("average loss", avg_loss)
#print("=" * 60)
return avg_loss
decoder_losses = AverageMeter()
# train
for ibatch, (img, noise) in enumerate(train_loader):
img = img.cuda()
noise = noise.cuda()
z_mu, z_sig = vae_encoder(img) # z_mu , z_sig : N*1*Dz
zl = torch.exp(0.5 * z_sig) * noise + z_mu # zl : N*L*Dz
x_ber = vae_decoder(zl) # N*L*Dx
#encoder_loss, decoder_loss = vae_loss(img, x_mu, x_sig, z_mu, z_sig)
encoder_loss, decoder_loss = vae_loss(img, x_ber, z_mu, z_sig)
encoder_losses.update(encoder_loss.item())
decoder_losses.update(decoder_loss.item())
opt_encoder.zero_grad()
opt_decoder.zero_grad()
encoder_loss.backward()
opt_encoder.step()
opt_decoder.step()
if ibatch % args.print_freq == 0 or ibatch + 1 == len(train_loader):
logger.log(
'[train Info]: [epoch-{}-{}][{:04d}/{:04d}][Encoder Loss:{:.2f}][Decoder Loss:{:.2f}]'
.format(ep, args.epoch, ibatch, len(train_loader),
encoder_loss.item(), decoder_loss.item()))
tfboard_writer.add_scalar('Encoder_Loss', encoder_losses.avg, ep)