def get_dataset_and_model(dataset_name, model_name, pretrained_model,
input_size):
if dataset_name == 'cityscapes':
dataset = CityscapesSemanticSegmentationDataset(
split='val', label_resolution='fine')
label_names = cityscapes_semantic_segmentation_label_names
elif dataset_name == 'ade20k':
dataset = ADE20KSemanticSegmentationDataset(split='val')
label_names = ade20k_semantic_segmentation_label_names
elif dataset_name == 'camvid':
dataset = CamVidDataset(split='test')
label_names = camvid_label_names
n_class = len(label_names)
if pretrained_model:
pretrained_model = pretrained_model
else:
pretrained_model = dataset_name
if model_name == 'pspnet_resnet101':
model = PSPNetResNet101(n_class=n_class,
pretrained_model=pretrained_model,
input_size=input_size)
elif model_name == 'pspnet_resnet50':
model = PSPNetResNet50(n_class=n_class,
pretrained_model=pretrained_model,
input_size=input_size)
elif model_name == 'segnet':
model = SegNetBasic(n_class=n_class, pretrained_model=pretrained_model)
return dataset, label_names, model
def calc_bn_statistics(model, gpu):
model.to_gpu(gpu)
d = CamVidDataset(split='train')
it = chainer.iterators.SerialIterator(d, 24, repeat=False, shuffle=False)
bn_params = {}
num_iterations = 0
for batch in it:
imgs, labels = concat_examples(batch, device=gpu)
model(imgs)
for name, link in model.namedlinks():
if name.endswith('_bn'):
if name not in bn_params:
bn_params[name] = [
cuda.to_cpu(link.avg_mean),
cuda.to_cpu(link.avg_var)
]
else:
bn_params[name][0] += cuda.to_cpu(link.avg_mean)
bn_params[name][1] += cuda.to_cpu(link.avg_var)
num_iterations += 1
for name, params in bn_params.items():
bn_params[name][0] /= num_iterations
bn_params[name][1] /= num_iterations
for name, link in model.namedlinks():
if name.endswith('_bn'):
link.avg_mean = bn_params[name][0]
link.avg_var = bn_params[name][1]
model.to_cpu()
return model
def recalculate_bn_statistics(model, batchsize, dtype='float32'):
print(
'==> Recalculating BN statistics (batchsize={}) ...'.format(batchsize))
train = CamVidDataset(split='train')
it = chainer.iterators.SerialIterator(train,
batchsize,
repeat=False,
shuffle=False)
bn_avg_mean = defaultdict(np.float32)
bn_avg_var = defaultdict(np.float32)
if dtype == 'mixed16':
dtype = 'float16'
n_iter = 0
for batch in it:
imgs, _ = concat_examples(batch)
model(F.cast(model.xp.array(imgs), dtype))
for name, link in model.namedlinks():
if name.endswith('_bn'):
bn_avg_mean[name] += link.avg_mean
bn_avg_var[name] += link.avg_var
n_iter += 1
for name, link in model.namedlinks():
if name.endswith('_bn'):
link.avg_mean = bn_avg_mean[name] / n_iter
link.avg_var = bn_avg_var[name] / n_iter
return model
def calc_bn_statistics(model, batchsize):
train = CamVidDataset(split="train")
it = chainer.iterators.SerialIterator(train,
batchsize,
repeat=False,
shuffle=False)
bn_avg_mean = defaultdict(np.float32)
bn_avg_var = defaultdict(np.float32)
n_iter = 0
for batch in it:
imgs, _ = concat_examples(batch)
model(model.xp.array(imgs))
for name, link in model.namedlinks():
if name.endwith("_bn"):
bn_avg_mean[name] += link.avg_mean
bn_avg_var[name] += link.avg_var
n_iter += 1
for name, link in model.namedlinks():
if name.endwith("_bn"):
link.avg_mean = bn_avg_mean[name] / n_iter
link.avg_var = bn_avg_var[name] / n_iter
return model
def CamVid_loader(dataset_dir=root_dir):
# Dataset
train = CamVidDataset(dataset_dir, split='train')
test = CamVidDataset(dataset_dir, split='val')
train = concat_examples(train)
test = concat_examples(test)
train_images = train[:][0]
train_labels = train[:][1]
test_images = test[:][0]
test_labels = test[:][1]
train_images /= 255.0
test_images /= 255.0
return train_images, test_images, train_labels, test_labels
def main():
args = arg()
data = CamVidDataset(split=args.type)
img = data[args.index][0]
img /= 255.0
#img = np.clip(img, -1.0, 1.0)
import matplotlib.pyplot as plt
img1 = img.transpose(1, 2, 0)
plt.imshow(img1)
plt.show()
def setup(dataset, model, pretrained_model, batchsize, input_size):
dataset_name = dataset
if dataset_name == 'cityscapes':
dataset = CityscapesSemanticSegmentationDataset(
split='val', label_resolution='fine')
label_names = cityscapes_semantic_segmentation_label_names
elif dataset_name == 'ade20k':
dataset = ADE20KSemanticSegmentationDataset(split='val')
label_names = ade20k_semantic_segmentation_label_names
elif dataset_name == 'camvid':
dataset = CamVidDataset(split='test')
label_names = camvid_label_names
elif dataset_name == 'voc':
dataset = VOCSemanticSegmentationDataset(split='val')
label_names = voc_semantic_segmentation_label_names
def eval_(out_values, rest_values):
pred_labels, = out_values
gt_labels, = rest_values
result = eval_semantic_segmentation(pred_labels, gt_labels)
for iu, label_name in zip(result['iou'], label_names):
print('{:>23} : {:.4f}'.format(label_name, iu))
print('=' * 34)
print('{:>23} : {:.4f}'.format('mean IoU', result['miou']))
print('{:>23} : {:.4f}'.format('Class average accuracy',
result['mean_class_accuracy']))
print('{:>23} : {:.4f}'.format('Global average accuracy',
result['pixel_accuracy']))
cls, pretrained_models, default_batchsize = models[model]
if pretrained_model is None:
pretrained_model = pretrained_models.get(dataset_name, dataset_name)
if input_size is None:
input_size = None
else:
input_size = (input_size, input_size)
kwargs = {
'n_class': len(label_names),
'pretrained_model': pretrained_model,
}
if model in ['pspnet_resnet50', 'pspnet_resnet101']:
kwargs.update({'input_size': input_size})
elif model == 'deeplab_v3plus_xception65':
kwargs.update({'min_input_size': input_size})
model = cls(**kwargs)
if batchsize is None:
batchsize = default_batchsize
return dataset, eval_, model, batchsize
def main():
n_class = 11
dataset = CamVidDataset(split="train")
n_cls_pixels = np.zeros((n_class, ))
n_img_pizels = np.zeros((n_class, ))
for img, label in dataset:
for cls_i in np.unique(label):
if cls_i == -1:
continue
n_cls_pixels[cls_i] += np.sum(label == cls_i)
n_img_pizels[cls_i] += label.size
freq = n_cls_pixels / n_img_pizels
median_freq = np.median(freq)
np.save("class_weight", median_freq / freq)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--pretrained_model', type=str, default='camvid')
parser.add_argument('--batchsize', type=int, default=24)
args = parser.parse_args()
model = SegNetBasic(n_class=len(camvid_label_names),
pretrained_model=args.pretrained_model)
if args.gpu >= 0:
model.to_gpu(args.gpu)
model = calc_bn_statistics(model, args.batchsize)
chainer.config.train = False
test = CamVidDataset(split='test')
it = chainer.iterators.SerialIterator(test,
batch_size=args.batchsize,
repeat=False,
shuffle=False)
imgs, pred_values, gt_values = apply_prediction_to_iterator(
model.predict, it)
# Delete an iterator of images to save memory usage.
del imgs
pred_labels, = pred_values
gt_labels, = gt_values
confusion = calc_semantic_segmentation_confusion(pred_labels, gt_labels)
ious = calc_semantic_segmentation_iou(confusion)
pixel_accuracy = np.diag(confusion).sum() / confusion.sum()
mean_pixel_accuracy = np.mean(
np.diag(confusion) / np.sum(confusion, axis=1))
for iou, label_name in zip(ious, camvid_label_names):
print('{:>23} : {:.4f}'.format(label_name, iou))
print('=' * 34)
print('{:>23} : {:.4f}'.format('mean IoU', np.nanmean(ious)))
print('{:>23} : {:.4f}'.format('Class average accuracy',
mean_pixel_accuracy))
print('{:>23} : {:.4f}'.format('Global average accuracy', pixel_accuracy))
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--gpu", type=int, default=-1)
parser.add_argument("--pretrained_model", type=str, default="camvid")
parser.add_argument("-batchsize", type=int, default=24)
args = parser.parse_args()
model = SegNetBasic(n_class=len(camvid_label_names),
pretrained_model=args.pretrained_model)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
model = calc_bn_statistics(model, args.batchsize)
test = CamVidDataset(split="test")
it = chainer.iterators.SerialIterator(test,
batch_size=args.batchsize,
repeat=False,
shuffle=False)
imgs, pred_values, gt_values = apply_prediction_to_iterator(
model.predict, it)
# Delete an iterator of iamges to save memory usage.
del imgs
pred_labels, = pred_values
gt_labels, = gt_values
result = eval_semantic_segmentation(pred_labels, gt_labels)
for iu, label_name in zip(result["iou"], camvid_label_names):
print("{:>23} : {:.4f}".format(label_name, iu))
print("=" * 34)
print("{:>23} : {:.4f}".format("mean IoU", result["miou"]))
print("{:>23} : {:.4f}".format("Class average accuracy",
result["mean_calss_accuracy"]))
print("{:>23} : {:.4f}".format("Global average accuracy",
result["pixe;_accuracy"]))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--pretrained_model', type=str, default='camvid')
parser.add_argument('--batchsize', type=int, default=24)
args = parser.parse_args()
model = SegNetBasic(n_class=len(camvid_label_names),
pretrained_model=args.pretrained_model)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
model = calc_bn_statistics(model, args.batchsize)
chainer.config.train = False
test = CamVidDataset(split='test')
it = chainer.iterators.SerialIterator(test,
batch_size=args.batchsize,
repeat=False,
shuffle=False)
imgs, pred_values, gt_values = apply_prediction_to_iterator(
model.predict, it)
# Delete an iterator of images to save memory usage.
del imgs
pred_labels, = pred_values
gt_labels, = gt_values
result = eval_semantic_segmentation(pred_labels, gt_labels)
for iu, label_name in zip(result['iou'], camvid_label_names):
print('{:>23} : {:.4f}'.format(label_name, iu))
print('=' * 34)
print('{:>23} : {:.4f}'.format('mean IoU', result['miou']))
print('{:>23} : {:.4f}'.format('Class average accuracy',
result['mean_class_accuracy']))
print('{:>23} : {:.4f}'.format('Global average accuracy',
result['pixel_accuracy']))
def main():
args = arg()
model = CamVidAutoEncoder(F.mean_squared_error)
chainer.serializers.load_npz(args.model, model)
train = CamVidDataset(split='test')
img = train[0][0]
img /= 255.0
ch, w, h = img.shape
example_input = img.reshape(1, ch, w, h)
x = chainer.Variable(example_input)
y = model.predict(x)
import matplotlib.pyplot as plt
img1 = img.transpose(1, 2, 0)
plt.subplot(1,2,1)
plt.imshow(img1)
y = y.data.reshape(ch, w, h)
y = y.transpose(1, 2, 0).clip(0, 1.0)
plt.subplot(1,2,2)
plt.imshow(y)
plt.show()
def CamVid_loader(dataset_dir=root_dir):
# Dataset
train = CamVidDataset(dataset_dir, split='train')
train = TransformDataset(train, transform)
test = CamVidDataset(dataset_dir, split='val')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--batchsize', type=int, default=12)
parser.add_argument('--class_weight', type=str, default='class_weight.npy')
parser.add_argument('--out', type=str, default='result')
args = parser.parse_args()
# Triggers
log_trigger = (50, 'iteration')
validation_trigger = (2000, 'iteration')
end_trigger = (16000, 'iteration')
# Dataset
train = CamVidDataset(split='train')
train = TransformDataset(train, transform)
val = CamVidDataset(split='val')
# Iterator
train_iter = iterators.MultiprocessIterator(train, args.batchsize)
val_iter = iterators.MultiprocessIterator(val,
args.batchsize,
shuffle=False,
repeat=False)
# Model
class_weight = np.load(args.class_weight)
model = SegNetBasic(n_class=len(camvid_label_names))
model = PixelwiseSoftmaxClassifier(model, class_weight=class_weight)
if args.gpu >= 0:
# Make a specified GPU current
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu() # Copy the model to the GPU
# Optimizer
optimizer = optimizers.MomentumSGD(lr=0.1, momentum=0.9)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer_hooks.WeightDecay(rate=0.0005))
# Updater
updater = training.updaters.StandardUpdater(train_iter,
optimizer,
device=args.gpu)
# Trainer
trainer = training.Trainer(updater, end_trigger, out=args.out)
trainer.extend(extensions.LogReport(trigger=log_trigger))
trainer.extend(extensions.observe_lr(), trigger=log_trigger)
trainer.extend(extensions.dump_graph('main/loss'))
if extensions.PlotReport.available():
trainer.extend(
extensions.PlotReport(['main/loss'],
x_key='iteration',
file_name='loss.png'))
trainer.extend(
extensions.PlotReport(['validation/main/miou'],
x_key='iteration',
file_name='miou.png'))
trainer.extend(extensions.snapshot_object(
model.predictor, filename='model_iteration-{.updater.iteration}'),
trigger=end_trigger)
trainer.extend(extensions.PrintReport([
'epoch', 'iteration', 'elapsed_time', 'lr', 'main/loss',
'validation/main/miou', 'validation/main/mean_class_accuracy',
'validation/main/pixel_accuracy'
]),
trigger=log_trigger)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(SemanticSegmentationEvaluator(val_iter, model.predictor,
camvid_label_names),
trigger=validation_trigger)
trainer.run()
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--gpu", type=int, default=-1)
parser.add_argument("--batchsize", type=int, default=12)
parser.add_argument("--class_weight", type=str, default="class_weight.npy")
parser.add_argument("--out", type=str, default="result")
args = parser.parse_args()
# Triggers
log_trigger = (50, "iteration")
validation_trigger = (2000, "iteration")
end_trigger = (16000, "iteration")
# Dataset
train = CamVidDataset(split="train")
train = TransformDataset(dataset=train, transform=transform)
val = CamVidDataset(split="val")
# Iterator
train_iter = iterators.MultiprocessIterator(train, args.batchsize)
val_iter = iterators.MultiprocessIterator(val, args.batchsize,
repeat=False,
shuffle=False)
# Model
class_weight = np.load(args.class_weight)
model = SegNetBasic(n_class=11)
model = PixelwiseSoftmaxClassifier(
model, class_weight=class_weight
)
if args.gpu >= 0:
# Make a specified GPU current
chainer.cuda.get_device_from_id(args.gpu).use()
# Copy the model to the GPU
model.to_gpu()
# Optimizer
optimizer = optimizers.MomentumSGD(lr=0.1, momentum=0.9)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(rate=0.0005))
# Updater
updater = training.StandardUpdater(train_iter, optimizer, device=args.gpu)
# Trainer
trainer = training.Trainer(updater=updater, stop_trigger=end_trigger, out=args.out)
trainer.extend(extensions.LogReport(trigger=log_trigger))
trainer.extend(extensions.observe_lr(), trigger=log_trigger)
trainer.extend(extensions.dump_graph("main/loss"))
if extensions.PlotReport.available():
trainer.extend(extensions.PlotReport(
["main/loss"], x_key="iteration",
file_name="loss.png"
))
trainer.extend(extensions.PlotReport(
["validation/loss"], x_key="iteration",
file_name="miou.png"
))
trainer.extend(extensions.snapshot_object(
model.predictor, filename="model_iteration-{.updater.iteration}"),
trigger=end_trigger)
trainer.extend(extensions.PrintReport(
["epoch", "iteration", "elapsed_time", "lr",
"main/loss", "validation/main/miou",
"validation/main/mean_class_accuracy",
"validation/main/pixel_accuracy"
]
), trigger=log_trigger)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(SemanticSegmentationEvaluator(
val_iter, model.predictor,
camvid_label_names), trigger=validation_trigger)
trainer.run()
def main():
# This follows evaluation code used in SegNet.
# https://github.com/alexgkendall/SegNet-Tutorial/blob/master/
# # Scripts/compute_test_results.m
parser = argparse.ArgumentParser()
parser.add_argument('gpu', type=int, default=-1)
parser.add_argument('snapshot', type=str)
parser.add_argument('--batchsize', type=int, default=24)
args = parser.parse_args()
n_class = 11
ignore_labels = [-1]
model = SegNetBasic(n_class=n_class)
serializers.load_npz(args.snapshot, model)
model = calc_bn_statistics(model, args.gpu)
model.train = False
if args.gpu >= 0:
model.to_gpu(args.gpu)
test = CamVidDataset(split='test')
it = chainer.iterators.SerialIterator(test,
batch_size=args.batchsize,
repeat=False,
shuffle=False)
n_positive = [0 for _ in range(n_class)]
n_true = [0 for _ in range(n_class)]
n_true_positive = [0 for _ in range(n_class)]
for batch in it:
img, gt_label = concat_examples(batch, args.gpu)
img = chainer.Variable(img, volatile=True)
pred_label = F.argmax(F.softmax(model(img)), axis=1)
pred_label = cuda.to_cpu(pred_label.data)
gt_label = cuda.to_cpu(gt_label)
for cls_i in range(n_class):
if cls_i in ignore_labels:
continue
n_positive[cls_i] += np.sum(pred_label == cls_i)
n_true[cls_i] += np.sum(gt_label == cls_i)
n_true_positive[cls_i] += np.sum(
(pred_label == cls_i) * (gt_label == cls_i))
ious = []
mean_accs = []
pixel_accs = []
for cls_i in range(n_class):
if cls_i in ignore_labels:
continue
deno = n_positive[cls_i] + n_true[cls_i] - n_true_positive[cls_i]
iou = n_true_positive[cls_i] / deno
ious.append(iou)
print('{:>23} : {:.4f}'.format(camvid_label_names[cls_i], iou))
mean_accs.append(n_true_positive[cls_i] / n_true[cls_i])
pixel_accs.append([n_true_positive[cls_i], n_true[cls_i]])
print('=' * 34)
print('{:>23} : {:.4f}'.format('mean IoU', np.mean(ious)))
print('{:>23} : {:.4f}'.format('Class average accuracy',
np.mean(mean_accs)))
pixel_accs = np.asarray(pixel_accs)
pixel_accs = pixel_accs[:, 0].sum() / pixel_accs[:, 1].sum()
print('{:>23} : {:.4f}'.format('Global average accuracy', pixel_accs))
def setUp(self):
self.dataset = CamVidDataset(split=self.split)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--batchsize', type=int, default=12)
parser.add_argument('--class-weight', type=str, default='class_weight.npy')
parser.add_argument('--val-iter', type=int, default=2000)
parser.add_argument('--out', type=str, default='result')
parser.add_argument('--iter', type=int, default=16000)
parser.add_argument('--dtype', type=str, default='float32')
parser.add_argument('--init-scale', type=float, default=1.0)
parser.add_argument('--n-uf', type=float, default=1e-5)
parser.add_argument('--dynamic-interval', type=int, default=None)
parser.add_argument('--loss-scale-method',
type=str,
default='approx_range')
parser.add_argument('--verbose', action='store_true', default=False)
args = parser.parse_args()
# Data type
print('==> Using data type: {}'.format(args.dtype))
chainer.global_config.dtype = dtypes[args.dtype]
# Triggers
log_trigger = (50, 'iteration')
validation_trigger = (args.val_iter, 'iteration')
end_trigger = (args.iter, 'iteration')
# Dataset
train = CamVidDataset(split='train')
train = TransformDataset(train, transform)
val = CamVidDataset(split='val')
# Iterator
train_iter = iterators.MultiprocessIterator(train, args.batchsize)
val_iter = iterators.MultiprocessIterator(val,
args.batchsize,
shuffle=False,
repeat=False)
# Model
class_weight = np.load(args.class_weight)
model = SegNetBasic(n_class=len(camvid_label_names),
dtype=dtypes[args.dtype])
# adaptive loss scaling
if args.dtype != 'float32':
recorder = AdaLossRecorder(sample_per_n_iter=100)
model = AdaLossScaled(model,
init_scale=args.init_scale,
transforms=[
AdaLossTransformLinear(),
AdaLossTransformConvolution2D(),
AdaLossTransformBatchNormalization(),
],
cfg={
'loss_scale_method': args.loss_scale_method,
'scale_upper_bound': 65504,
'accum_upper_bound': 65504,
'update_per_n_iteration': 100,
'recorder': recorder,
'n_uf_threshold': args.n_uf,
},
verbose=args.verbose)
model = PixelwiseSoftmaxClassifier(model, class_weight=class_weight)
if args.gpu >= 0:
# Make a specified GPU current
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu() # Copy the model to the GPU
# Optimizer
optimizer = optimizers.MomentumSGD(lr=0.1, momentum=0.9)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer_hooks.WeightDecay(rate=0.0005))
if args.dtype == 'mixed16':
optimizer.use_fp32_update()
# Updater
updater = training.updaters.StandardUpdater(train_iter,
optimizer,
device=args.gpu)
if args.dtype == 'mixed16':
if args.dynamic_interval is not None:
print('==> Using dynamic loss scaling (interval={}) ...'.format(
args.dynamic_interval))
optimizer.loss_scaling(interval=args.dynamic_interval, scale=None)
optimizer._loss_scale_max = 32770.0
else:
if args.loss_scale_method == 'approx_range':
print('==> Using adaptive loss scaling ...')
else:
print('==> Using default fixed loss scaling (scale={}) ...'.
format(args.init_scale))
optimizer.loss_scaling(interval=float('inf'), scale=None)
optimizer._loss_scale_max = 1.0 # to prevent actual loss scaling
# Trainer
trainer = training.Trainer(updater, end_trigger, out=args.out)
# warmup_attr_ratio = 0.1 if args.dtype != 'float32' else None
# # NOTE: this is confusing but it means n_iter
# warmup_n_epoch = 1000 if args.dtype != 'float32' else None
# lr_shift = chainerlp.extensions.ExponentialShift(
# 'lr',
# 0.1,
# init=0.1 * warmup_attr_ratio,
# warmup_attr_ratio=warmup_attr_ratio,
# warmup_n_epoch=warmup_n_epoch)
# trainer.extend(lr_shift, trigger=(1, 'iteration'))
trainer.extend(extensions.LogReport(trigger=log_trigger))
trainer.extend(extensions.observe_lr(), trigger=log_trigger)
if args.dtype != 'float32':
trainer.extend(extensions.observe_value(
'loss_scale',
lambda trainer: trainer.updater.get_optimizer('main')._loss_scale),
trigger=log_trigger)
trainer.extend(extensions.dump_graph('main/loss'))
# snapshot the trainer after each
trainer.extend(extensions.snapshot(), trigger=validation_trigger)
# snapshot the model itself
trainer.extend(extensions.snapshot_object(
get_snapshot_model(model), 'model_iter_{.updater.iteration}'),
trigger=validation_trigger)
if extensions.PlotReport.available():
trainer.extend(
extensions.PlotReport(['main/loss'],
x_key='iteration',
file_name='loss.png'))
trainer.extend(
extensions.PlotReport(['validation/main/miou'],
x_key='iteration',
file_name='miou.png'))
metrics = [
'epoch', 'iteration', 'elapsed_time', 'lr', 'main/loss',
'validation/main/miou', 'validation/main/mean_class_accuracy',
'validation/main/pixel_accuracy'
]
if args.dtype != 'float32':
metrics.append('loss_scale')
trainer.extend(extensions.PrintReport(metrics), trigger=log_trigger)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(SemanticSegmentationEvaluator(val_iter, model.predictor,
camvid_label_names),
trigger=validation_trigger)
# snapshot the best validation result
trainer.extend(extensions.snapshot_object(get_snapshot_model(model),
'model_best'),
trigger=chainer.training.triggers.MaxValueTrigger(
'validation/main/miou', trigger=validation_trigger))
hooks = []
if args.dtype != 'float32':
hook = AdaLossMonitor(sample_per_n_iter=100,
verbose=args.verbose,
includes=['Grad', 'Deconvolution'])
recorder.trainer = trainer
hook.trainer = trainer
hooks.append(hook)
with ExitStack() as stack:
for hook in hooks:
stack.enter_context(hook)
trainer.run()
chainer.serializers.save_npz(
os.path.join(args.out, 'snapshot_model.npz'),
recalculate_bn_statistics(model.predictor,
args.batchsize,
dtype=args.dtype))
if args.dtype != 'float32':
recorder.export().to_csv(os.path.join(args.out, 'loss_scale.csv'))
from __future__ import division
import numpy as np
from chainercv.datasets import CamVidDataset
n_class = 11
dataset = CamVidDataset(split='train')
n_cls_pixels = np.zeros((n_class, ))
n_img_pixels = np.zeros((n_class, ))
for img, label in dataset:
for cls_i in np.unique(label):
if cls_i == -1:
continue
n_cls_pixels[cls_i] += np.sum(label == cls_i)
n_img_pixels[cls_i] += label.size
freq = n_cls_pixels / n_img_pixels
median_freq = np.median(freq)
np.save('class_weight', median_freq / freq)
