def __init__(self, masking_from_epoch, num_epochs, moving_mask_percent, masking_linear_increase):
self.masking_from_epoch = masking_from_epoch
self.num_epochs = num_epochs
self.moving_mask_percent = moving_mask_percent
self.masking_linear_increase = masking_linear_increase
self.segmentation_input_key = ('color_aug', 0, 0)
self.logits_key = ('segmentation_logits', 0)
self.metric_model_moving = SegmentationRunningScore(2)
self.iou_thresh = dict()
self.iou_thresh['non_moving'] = 0.0
self.iou_thresh['moving'] = 0.0
self.iou_log = dict()
self.iou_log['non_moving'] = list()
self.iou_log['moving'] = list()
def _run_segmentation_validation(self, images_to_keep=0):
scores = dict()
images = dict()
# torch.no_grad() = disable gradient calculation
with torch.no_grad(), self.state.model_manager.get_eval() as model:
for batch in self.segmentation_validation_loader:
domain = batch[0]['domain'][0]
num_classes = batch[0]['num_classes'][0].item()
if domain not in scores:
scores[domain] = SegmentationRunningScore(num_classes)
images[domain] = list()
_ = self._validate_batch_segmentation(model, batch,
scores[domain],
images[domain])
images[domain] = images[domain][:images_to_keep]
return scores, images
def __init__(self, options, model=None):
if __name__ == "__main__":
print(" -> Executing script", os.path.basename(__file__))
self.opt = options
self.device = torch.device("cpu" if self.opt.no_cuda else "cuda")
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LABELS
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
assert self.opt.train_set in {1, 2, 3, 12, 123}, "Invalid train_set!"
assert self.opt.task_to_val in {0, 1, 2, 3, 12, 123}, "Invalid task!"
keys_to_load = ['color', 'segmentation']
# Labels
labels = self._get_labels_cityscapes()
# Train IDs
self.train_ids = set([labels[i].trainId for i in range(len(labels))])
self.train_ids.remove(255)
self.train_ids = sorted(list(self.train_ids))
self.num_classes_model = len(self.train_ids)
# Task handling
if self.opt.task_to_val != 0:
labels_task = self._get_task_labels_cityscapes()
train_ids_task = set(
[labels_task[i].trainId for i in range(len(labels_task))])
train_ids_task.remove(255)
self.task_low = min(train_ids_task)
self.task_high = max(train_ids_task) + 1
labels = labels_task
self.train_ids = sorted(list(train_ids_task))
else:
self.task_low = 0
self.task_high = self.num_classes_model
self.opt.task_to_val = self.opt.train_set
# Number of classes for the SegmentationRunningScore
self.num_classes_score = self.task_high - self.task_low
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# DATASET DEFINITIONS
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Data augmentation
test_data_transforms = [
mytransforms.CreateScaledImage(),
mytransforms.Resize((self.opt.height, self.opt.width),
image_types=['color']),
mytransforms.ConvertSegmentation(),
mytransforms.CreateColoraug(new_element=True,
scales=self.opt.scales),
mytransforms.RemoveOriginals(),
mytransforms.ToTensor(),
mytransforms.NormalizeZeroMean(),
]
# If hyperparameter search, only load the respective validation set. Else, load the full validation set.
if self.opt.hyperparameter:
trainvaltest_split = 'train'
folders_to_load = CitySet.get_city_set(-1)
else:
trainvaltest_split = 'validation'
folders_to_load = None
test_dataset = CityscapesDataset(dataset='cityscapes',
split=self.opt.dataset_split,
trainvaltest_split=trainvaltest_split,
video_mode='mono',
stereo_mode='mono',
scales=self.opt.scales,
labels_mode='fromid',
labels=labels,
keys_to_load=keys_to_load,
data_transforms=test_data_transforms,
video_frames=self.opt.video_frames,
folders_to_load=folders_to_load)
self.test_loader = DataLoader(dataset=test_dataset,
batch_size=self.opt.batch_size,
shuffle=False,
num_workers=self.opt.num_workers,
pin_memory=True,
drop_last=False)
print(
"++++++++++++++++++++++ INIT VALIDATION ++++++++++++++++++++++++")
print("Using dataset\n ", self.opt.dataset, "with split",
self.opt.dataset_split)
print("There are {:d} validation items\n ".format(len(test_dataset)))
print("Validating classes up to train set\n ", self.opt.train_set)
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LOGGING OPTIONS
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# If no model is passed, standalone validation is to be carried out. The log_path needs to be set before
# self.load_model() is invoked.
if model is None:
self.opt.validate = False
self.opt.model_name = self.opt.load_model_name
path_getter = GetPath()
log_path = path_getter.get_checkpoint_path()
self.log_path = os.path.join(log_path, 'erfnet', self.opt.model_name)
# All outputs will be saved to save_path
self.save_path = self.log_path
# Create output path for standalone validation
if not self.opt.validate:
save_dir = 'eval_{}'.format(self.opt.dataset)
if self.opt.hyperparameter:
save_dir = save_dir + '_hyper'
save_dir = save_dir + '_task_to_val{}'.format(self.opt.task_to_val)
self.save_path = os.path.join(self.log_path, save_dir)
if not os.path.exists(self.save_path):
os.makedirs(self.save_path)
# Copy this file to save_path
shutil.copy2(__file__, self.save_path)
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# MODEL DEFINITION
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Standalone validation
if not self.opt.validate:
# Create a conventional ERFNet
self.model = ERFNet(self.num_classes_model, self.opt)
self.load_model()
self.model.to(self.device)
# Validate while training
else:
self.model = model
self.model.eval()
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LOGGING OPTIONS II
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# self.called is used to decide which file mode shall be used when writing metrics to disk.
self.called = False
self.metric_model = SegmentationRunningScore(self.num_classes_score)
# Metrics are only saved if val_frequency > 0!
if self.opt.val_frequency != 0:
print("Saving metrics to\n ", self.save_path)
# Set up colour output. Coloured images are only output if standalone validation is carried out!
if not self.opt.validate and self.opt.save_pred_to_disk:
# Output path
self.img_path = os.path.join(
self.save_path, 'output_{}'.format(self.opt.weights_epoch))
if self.opt.pred_wout_blend:
self.img_path += '_wout_blend'
if not os.path.exists(self.img_path):
os.makedirs(self.img_path)
print("Saving prediction images to\n ", self.img_path)
print("Save frequency\n ", self.opt.pred_frequency)
# Get the colours from dataset.
colors = [
(label.trainId - self.task_low, label.color)
for label in labels
if label.trainId != 255 and label.trainId in self.train_ids
]
colors.append((255, (0, 0, 0))) # void class
self.id_color = dict(colors)
self.id_color_keys = [key for key in self.id_color.keys()]
self.id_color_vals = [val for val in self.id_color.values()]
# Ongoing index to name the outputs
self.img_idx = 0
# Set up probability output. Probabilities are only output if standalone validation is carried out!
if not self.opt.validate and self.opt.save_probs_to_disk:
# Output path
self.logit_path = os.path.join(
self.save_path,
'probabilities_{}'.format(self.opt.weights_epoch))
if not os.path.exists(self.logit_path):
os.makedirs(self.logit_path)
print("Saving probabilities to\n ", self.logit_path)
print("Save frequency\n ", self.opt.probs_frequency)
# Ongoing index to name the probability outputs
self.probs_idx = 0
print(
"+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++")
# Save all options to disk and print them to stdout
self._print_options()
self._save_opts(len(test_dataset))
class Evaluator:
def __init__(self, options, model=None):
if __name__ == "__main__":
print(" -> Executing script", os.path.basename(__file__))
self.opt = options
self.device = torch.device("cpu" if self.opt.no_cuda else "cuda")
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LABELS
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
assert self.opt.train_set in {1, 2, 3, 12, 123}, "Invalid train_set!"
assert self.opt.task_to_val in {0, 1, 2, 3, 12, 123}, "Invalid task!"
keys_to_load = ['color', 'segmentation']
# Labels
labels = self._get_labels_cityscapes()
# Train IDs
self.train_ids = set([labels[i].trainId for i in range(len(labels))])
self.train_ids.remove(255)
self.train_ids = sorted(list(self.train_ids))
self.num_classes_model = len(self.train_ids)
# Task handling
if self.opt.task_to_val != 0:
labels_task = self._get_task_labels_cityscapes()
train_ids_task = set(
[labels_task[i].trainId for i in range(len(labels_task))])
train_ids_task.remove(255)
self.task_low = min(train_ids_task)
self.task_high = max(train_ids_task) + 1
labels = labels_task
self.train_ids = sorted(list(train_ids_task))
else:
self.task_low = 0
self.task_high = self.num_classes_model
self.opt.task_to_val = self.opt.train_set
# Number of classes for the SegmentationRunningScore
self.num_classes_score = self.task_high - self.task_low
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# DATASET DEFINITIONS
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Data augmentation
test_data_transforms = [
mytransforms.CreateScaledImage(),
mytransforms.Resize((self.opt.height, self.opt.width),
image_types=['color']),
mytransforms.ConvertSegmentation(),
mytransforms.CreateColoraug(new_element=True,
scales=self.opt.scales),
mytransforms.RemoveOriginals(),
mytransforms.ToTensor(),
mytransforms.NormalizeZeroMean(),
]
# If hyperparameter search, only load the respective validation set. Else, load the full validation set.
if self.opt.hyperparameter:
trainvaltest_split = 'train'
folders_to_load = CitySet.get_city_set(-1)
else:
trainvaltest_split = 'validation'
folders_to_load = None
test_dataset = CityscapesDataset(dataset='cityscapes',
split=self.opt.dataset_split,
trainvaltest_split=trainvaltest_split,
video_mode='mono',
stereo_mode='mono',
scales=self.opt.scales,
labels_mode='fromid',
labels=labels,
keys_to_load=keys_to_load,
data_transforms=test_data_transforms,
video_frames=self.opt.video_frames,
folders_to_load=folders_to_load)
self.test_loader = DataLoader(dataset=test_dataset,
batch_size=self.opt.batch_size,
shuffle=False,
num_workers=self.opt.num_workers,
pin_memory=True,
drop_last=False)
print(
"++++++++++++++++++++++ INIT VALIDATION ++++++++++++++++++++++++")
print("Using dataset\n ", self.opt.dataset, "with split",
self.opt.dataset_split)
print("There are {:d} validation items\n ".format(len(test_dataset)))
print("Validating classes up to train set\n ", self.opt.train_set)
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LOGGING OPTIONS
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# If no model is passed, standalone validation is to be carried out. The log_path needs to be set before
# self.load_model() is invoked.
if model is None:
self.opt.validate = False
self.opt.model_name = self.opt.load_model_name
path_getter = GetPath()
log_path = path_getter.get_checkpoint_path()
self.log_path = os.path.join(log_path, 'erfnet', self.opt.model_name)
# All outputs will be saved to save_path
self.save_path = self.log_path
# Create output path for standalone validation
if not self.opt.validate:
save_dir = 'eval_{}'.format(self.opt.dataset)
if self.opt.hyperparameter:
save_dir = save_dir + '_hyper'
save_dir = save_dir + '_task_to_val{}'.format(self.opt.task_to_val)
self.save_path = os.path.join(self.log_path, save_dir)
if not os.path.exists(self.save_path):
os.makedirs(self.save_path)
# Copy this file to save_path
shutil.copy2(__file__, self.save_path)
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# MODEL DEFINITION
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Standalone validation
if not self.opt.validate:
# Create a conventional ERFNet
self.model = ERFNet(self.num_classes_model, self.opt)
self.load_model()
self.model.to(self.device)
# Validate while training
else:
self.model = model
self.model.eval()
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LOGGING OPTIONS II
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# self.called is used to decide which file mode shall be used when writing metrics to disk.
self.called = False
self.metric_model = SegmentationRunningScore(self.num_classes_score)
# Metrics are only saved if val_frequency > 0!
if self.opt.val_frequency != 0:
print("Saving metrics to\n ", self.save_path)
# Set up colour output. Coloured images are only output if standalone validation is carried out!
if not self.opt.validate and self.opt.save_pred_to_disk:
# Output path
self.img_path = os.path.join(
self.save_path, 'output_{}'.format(self.opt.weights_epoch))
if self.opt.pred_wout_blend:
self.img_path += '_wout_blend'
if not os.path.exists(self.img_path):
os.makedirs(self.img_path)
print("Saving prediction images to\n ", self.img_path)
print("Save frequency\n ", self.opt.pred_frequency)
# Get the colours from dataset.
colors = [
(label.trainId - self.task_low, label.color)
for label in labels
if label.trainId != 255 and label.trainId in self.train_ids
]
colors.append((255, (0, 0, 0))) # void class
self.id_color = dict(colors)
self.id_color_keys = [key for key in self.id_color.keys()]
self.id_color_vals = [val for val in self.id_color.values()]
# Ongoing index to name the outputs
self.img_idx = 0
# Set up probability output. Probabilities are only output if standalone validation is carried out!
if not self.opt.validate and self.opt.save_probs_to_disk:
# Output path
self.logit_path = os.path.join(
self.save_path,
'probabilities_{}'.format(self.opt.weights_epoch))
if not os.path.exists(self.logit_path):
os.makedirs(self.logit_path)
print("Saving probabilities to\n ", self.logit_path)
print("Save frequency\n ", self.opt.probs_frequency)
# Ongoing index to name the probability outputs
self.probs_idx = 0
print(
"+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++")
# Save all options to disk and print them to stdout
self._print_options()
self._save_opts(len(test_dataset))
def _get_labels_cityscapes(self, id=None):
if id is None:
id = self.opt.train_set
if id == 1:
labels = labels_cityscape_seg_train1.getlabels()
elif id == 2:
labels = labels_cityscape_seg_train2_eval.getlabels()
elif id == 12:
labels = labels_cityscape_seg_train2_eval.getlabels()
elif id == 3:
labels = labels_cityscape_seg_train3_eval.getlabels()
elif id == 123:
labels = labels_cityscape_seg_train3_eval.getlabels()
return labels
def _get_task_labels_cityscapes(self, id=None):
if id is None:
id = self.opt.task_to_val
if id == 1:
labels_task = labels_cityscape_seg_train1.getlabels()
elif id == 2:
labels_task = labels_cityscape_seg_train2.getlabels()
elif id == 12:
labels_task = labels_cityscape_seg_train2_eval.getlabels()
elif id == 3:
labels_task = labels_cityscape_seg_train3.getlabels()
elif id == 123:
labels_task = labels_cityscape_seg_train3_eval.getlabels()
return labels_task
def load_model(self):
"""Load model(s) from disk
"""
base_path = os.path.split(self.log_path)[0]
checkpoint_path = os.path.join(
base_path, self.opt.load_model_name, 'models',
'weights_{}'.format(self.opt.weights_epoch))
assert os.path.isdir(checkpoint_path), \
"Cannot find folder {}".format(checkpoint_path)
print("loading model from folder {}".format(checkpoint_path))
path = os.path.join(checkpoint_path, "{}.pth".format('model'))
model_dict = self.model.state_dict()
if self.opt.no_cuda:
pretrained_dict = torch.load(path, map_location='cpu')
else:
pretrained_dict = torch.load(path)
pretrained_dict = {
k: v
for k, v in pretrained_dict.items() if k in model_dict
}
model_dict.update(pretrained_dict)
self.model.load_state_dict(model_dict)
def calculate_metrics(self, epoch=None):
print("-> Computing predictions with input size {}x{}".format(
self.opt.height, self.opt.width))
print("-> Evaluating")
for data in self.test_loader:
with torch.no_grad():
input_color = data[("color_aug", 0, 0)]
gt_seg = data[('segmentation', 0, 0)][:, 0, :, :].numpy()
input_color = {
("color_aug", 0, 0): input_color.to(self.device)
}
output = self.model(input_color)
pred_seg = output['segmentation_logits'].float()
# Apply task reduction for argmax
if self.opt.task_to_val != 0:
pred_seg = pred_seg[:, self.task_low:self.task_high, ...]
gt_seg -= self.task_low # gt_seg trainIDs must be in range(0, self.num_classes_score) to map them with torch.argmax output
gt_seg[
gt_seg == 255 - self.
task_low] = 255 # maintaining the background trainID
# Save probabilities to disk
if not self.opt.validate and self.opt.save_probs_to_disk:
self._save_probs_to_disk(
F.softmax(pred_seg, dim=1).cpu().numpy())
pred_seg = F.interpolate(pred_seg,
gt_seg[0].shape,
mode='nearest')
# Select most probable class
pred_seg = torch.argmax(pred_seg, dim=1)
pred_seg = pred_seg.cpu().numpy()
self.metric_model.update(gt_seg, pred_seg)
# Save predictions to disk
if not self.opt.validate and self.opt.save_pred_to_disk:
self._save_pred_to_disk(pred_seg, gt_seg)
metrics = self.metric_model.get_scores()
# Save metrics
if self.opt.val_frequency != 0:
# Local epoch will not be specified if the validation is carried out standalone.
if not self.opt.validate and epoch is None:
epoch = int(self.opt.weights_epoch)
self._save_metrics(epoch, metrics)
self.metric_model.reset()
print("\n " + ("{:>8} | " * 2).format("miou", "maccuracy"))
print(("&{: 8.3f} " *
2).format(metrics['meaniou'], metrics['meanacc']) + "\\\\")
print("\n-> Done!")
def _save_metrics(self, epoch, metrics):
''' Save metrics (class-wise) to disk as HDF5 file.
'''
# If a single model is validated, the output file will carry its epoch number in its file name. If a learning
# process is validated "on the go", the output filename will just be "validation.h5".
if not self.opt.validate:
filename = 'validation_{:d}.h5'.format(epoch)
else:
filename = 'validation.h5'
save_path = os.path.join(self.save_path, filename)
# When _save_metrics is invoked for the first time, the HDF file will be opened in "w" mode overwriting any
# existing file. In case of another invocation, the file will be opened in "a" mode not overwriting any
# existing file but appending the data.
if not self.called:
mode = 'w'
self.called = True
else:
mode = 'a'
# If a single model is validated, all datasets reside in the first layer of the HDF file. If a learning process
# is validated "on the go", each validated model will have its own group named after the epoch of the model.
with h5.File(save_path, mode) as f:
if self.opt.validate:
grp = f.create_group('epoch_{:d}'.format(epoch))
else:
grp = f
# Write mean_IoU, mean_acc and mean prec to file / group
dset = grp.create_dataset('mean_IoU', data=metrics['meaniou'])
dset.attrs[
'Description'] = 'See trainIDs for information on the classes'
dset = grp.create_dataset('mean_recall', data=metrics['meanacc'])
dset.attrs[
'Description'] = 'See trainIDs for information on the classes'
dset.attrs['AKA'] = 'Accuracy -> TP / (TP + FN)'
dset = grp.create_dataset('mean_precision',
data=metrics['meanprec'])
dset.attrs[
'Description'] = 'See trainIDs for information on the classes'
dset.attrs['AKA'] = 'Precision -> TP / (TP + FP)'
# If in 'w' mode, allocate memory for class_id dataset
if mode == 'w':
ids = np.zeros(shape=(len(metrics['iou'])), dtype=np.uint32)
class_iou = np.zeros(shape=(len(metrics['iou'])), dtype=np.float64)
class_acc = np.zeros(shape=(len(metrics['acc'])), dtype=np.float64)
class_prec = np.zeros(shape=(len(metrics['prec'])),
dtype=np.float64)
# Disassemble the dictionary
for key, i in zip(sorted(metrics['iou']),
range(len(metrics['iou']))):
if mode == 'w':
ids[i] = self.train_ids[i] # int(key)
class_iou[i] = metrics['iou'][key]
class_acc[i] = metrics['acc'][key]
class_prec[i] = metrics['prec'][key]
# Create class_id dataset only once in first layer of HDF5 file when in 'w' mode
if mode == 'w':
dset = f.create_dataset('trainIDs', data=ids)
dset.attrs['Description'] = 'trainIDs of classes'
dset = f.create_dataset('first_epoch_in_file',
data=np.array([epoch
]).astype(np.uint32))
dset.attrs[
'Description'] = 'First epoch that has been saved in this file.'
dset = grp.create_dataset('class_IoU', data=class_iou)
dset.attrs[
'Description'] = 'See trainIDs for information on the class order'
dset = grp.create_dataset('class_recall', data=class_acc)
dset.attrs[
'Description'] = 'See trainIDs for information on the class order'
dset.attrs['AKA'] = 'Accuracy -> TP / (TP + FN)'
dset = grp.create_dataset('class_precision', data=class_prec)
dset.attrs[
'Description'] = 'See trainIDs for information on the class order'
dset.attrs['AKA'] = 'Precision -> TP / (TP + FP)'
def _save_pred_to_disk(self, pred, gt):
''' Save a correctly coloured image of the prediction (batch) to disk. Only every self.opt.pred_frequency-th
prediction is saved to disk!
'''
for i in range(gt.shape[0]):
if self.img_idx % self.opt.pred_frequency == 0:
o_size = gt[i].shape # original image shape
single_pred = pred[i].flatten()
single_gt = gt[i].flatten()
# Copy voids from ground truth to prediction
if not self.opt.pred_wout_blend:
single_pred[single_gt == 255] = 255
# Convert to colour
single_pred = self._convert_to_colour(single_pred, o_size)
single_gt = self._convert_to_colour(single_gt, o_size)
# Save predictions to disk using an ongoing index
cv2.imwrite(
os.path.join(self.img_path,
'pred_val_{}.png'.format(self.img_idx)),
single_pred)
cv2.imwrite(
os.path.join(self.img_path,
'gt_val_{}.png'.format(self.img_idx)),
single_gt)
self.img_idx += 1
def _convert_to_colour(self, img, o_size):
''' Replace trainIDs in prediction with colours from dict, reshape it afterwards to input dimensions and
convert RGB to BGR to match openCV's colour system.
'''
sort_idx = np.argsort(self.id_color_keys)
idx = np.searchsorted(self.id_color_keys, img, sorter=sort_idx)
img = np.asarray(self.id_color_vals)[sort_idx][idx]
img = img.astype(np.uint8)
img = np.reshape(img, newshape=(o_size[0], o_size[1], 3))
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
return img
def _save_probs_to_disk(self, output):
''' Save the network output as numpy npy-file to disk. Only every self.opt.probs_frequency-th image is saved
to disk!
'''
for i in range(output.shape[0]):
if self.probs_idx % self.opt.probs_frequency == 0:
np.save(
os.path.join(self.logit_path,
'seg_logit_{}'.format(self.probs_idx)),
output[i])
self.probs_idx += 1
def _print_options(self):
''' Print validation options to stdout
'''
# Convert namespace to dictionary
opts = vars(self.opt)
# Get max key length for left justifying
max_len = max([len(key) for key in opts.keys()])
# Print options to stdout
print(
"+++++++++++++++++++++++++++ OPTIONS +++++++++++++++++++++++++++")
for item in sorted(opts.items()):
print(item[0].ljust(max_len), item[1])
print(
"+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++")
def _save_opts(self, n_eval):
"""Save options to disk so we know what we ran this experiment with
"""
to_save = self.opt.__dict__.copy()
to_save['n_eval'] = n_eval
if self.opt.validate:
filename = 'eval_opt.json'
else:
filename = 'eval_opt_{}.json'.format(self.opt.weights_epoch)
with open(os.path.join(self.save_path, filename), 'w') as f:
json.dump(to_save, f, indent=2)
class DCMasking(object):
def __init__(self, masking_from_epoch, num_epochs, moving_mask_percent, masking_linear_increase):
self.masking_from_epoch = masking_from_epoch
self.num_epochs = num_epochs
self.moving_mask_percent = moving_mask_percent
self.masking_linear_increase = masking_linear_increase
self.segmentation_input_key = ('color_aug', 0, 0)
self.logits_key = ('segmentation_logits', 0)
self.metric_model_moving = SegmentationRunningScore(2)
self.iou_thresh = dict()
self.iou_thresh['non_moving'] = 0.0
self.iou_thresh['moving'] = 0.0
self.iou_log = dict()
self.iou_log['non_moving'] = list()
self.iou_log['moving'] = list()
def _moving_class_criterion(self, segmentation):
# TODO this is valid for the Cityscapes class definitions and has to be adapted for other datasets
# to be more generic
mask = (segmentation > 10) & (segmentation < 100)
return mask
def compute_segmentation_frames(self, batch, model):
batch_masking = deepcopy(batch)
# get the depth indices
batch_indices = tuple([idx_batch for idx_batch, sub_batch in enumerate(batch_masking)
if any('depth' in purpose_tuple
for purpose_tuple in sub_batch['purposes'])
])
# get the depth images
batch_masking = tuple([sub_batch for sub_batch in batch_masking
if any('depth' in purpose_tuple
for purpose_tuple in sub_batch['purposes'])
])
# replace the purpose to segmentation
for idx1, sub_batch in enumerate(batch_masking):
for idx2, purpose_tuple in enumerate(sub_batch['purposes']):
batch_masking[idx1]['purposes'][idx2] = tuple([purpose.replace('depth', 'segmentation')
for purpose in purpose_tuple])
# generate the correct keys and outputs
input_image_keys = [key for key in batch_masking[0].keys() if 'color_aug' in key]
output_segmentation_keys = [('segmentation', key[1], key[2]) for key in input_image_keys]
outputs_masked = list(dict() for i in range(len(batch)))
# pass all depth image frames through the network to get the segmentation outputs
for in_key, out_key in zip(input_image_keys, output_segmentation_keys):
wanted_keys = ['domain', 'purposes', 'domain_idx', in_key]
batch_masking_key = deepcopy(batch_masking)
batch_masking_key = tuple([{key: sub_batch[key] for key in sub_batch.keys()
if key in wanted_keys}
for sub_batch in batch_masking_key])
for idx1 in range(len(batch_masking_key)):
batch_masking_key[idx1][self.segmentation_input_key] = \
batch_masking_key[idx1][in_key].clone()
if in_key != self.segmentation_input_key:
del batch_masking_key[idx1][in_key]
outputs_masked_key = model(batch_masking_key)
cur_idx_outputs = 0
for idx_batch in range(len(outputs_masked)):
if idx_batch in batch_indices:
outputs_masked[idx_batch][out_key] = outputs_masked_key[cur_idx_outputs][self.logits_key].argmax(1)
cur_idx_outputs += 1
else:
outputs_masked[idx_batch] = None
outputs_masked = tuple(outputs_masked)
return outputs_masked
def compute_moving_mask(self, output_masked):
"""Compute moving mask and iou
"""
segmentation = output_masked[("segmentation", 0, 0)]
# Create empty mask
moving_mask_combined = torch.zeros(segmentation.shape).to(segmentation.device)
# Create binary mask moving in t = 0, movable object = 1, non_movable = 0
# Create binary masks (moving / non-moving)
moving_mask = dict()
moving_mask[0] = self._moving_class_criterion(segmentation).float()
for key in output_masked.keys():
if key[0] == "segmentation_warped":
moving_mask[key[1]] = self._moving_class_criterion(output_masked[("segmentation_warped", key[1], 0)])
# Calculate IoU for each frame separately
for i in range(moving_mask[0].shape[0]):
# Average score over frames
for frame_id in moving_mask.keys():
if frame_id == 0:
continue
# For binary class
self.metric_model_moving.update(
np.array(moving_mask[frame_id][i].cpu()), np.array(moving_mask[0][i].cpu()))
scores = self.metric_model_moving.get_scores()
if not np.isnan(scores['iou'][0]):
self.iou_log['non_moving'].append(scores['iou'][0])
if not np.isnan(scores['iou'][1]):
self.iou_log['moving'].append(scores['iou'][1])
# Calculate Mask if scores of moving objects is not NaN
# mask every moving class, were the iou is smaller than threshold
if scores['iou'][1] < self.iou_thresh['moving']:
# Add moving mask of t = 0
moving_mask_combined[i] += self._moving_class_criterion(segmentation[i]).float()
# Add moving mask of segmentation mask of t!=0 warped to t=0
for frame_id in moving_mask.keys():
if frame_id == 0:
continue
moving_mask_combined[i] += self._moving_class_criterion(
output_masked[("segmentation_warped", frame_id, 0)][i]).float()
# mask moving in t != 0
self.metric_model_moving.reset()
# movable object = 0, non_movable = 1
output_masked['moving_mask'] = (moving_mask_combined < 1).float().detach()
def clear_iou_log(self):
self.iou_log = dict()
self.iou_log['non_moving'] = list()
self.iou_log['moving'] = list()
def calculate_iou_threshold(self, current_epoch):
if self.masking_from_epoch <= current_epoch:
self.iou_thresh = dict()
if self.masking_linear_increase:
percentage = 1 - (1 / (self.num_epochs - 1 - self.masking_from_epoch) * (
current_epoch + 1 - self.masking_from_epoch)) # Mask 100 % to 0 %
else:
percentage = self.moving_mask_percent
try:
self.iou_thresh['non_moving'] = np.percentile(self.iou_log['non_moving'], (100 * percentage)).item()
except Exception as e:
self.iou_thresh['non_moving'] = 0.0
print('Error calculating percentile of non_moving')
print(e)
try:
self.iou_thresh['moving'] = np.percentile(self.iou_log['moving'], (100 * percentage)).item()
except Exception as e:
self.iou_thresh['moving'] = 0.0
print('Error calculating percentile of moving')
print(e)
def __init__(self, options):
self.opt = options
self.device = torch.device("cpu" if self.opt.no_cuda else "cuda")
# Assertions
assert os.path.isfile(self.opt.image), "Invalid image!"
self.opt.image.replace('/', os.sep)
self.opt.image.replace('\\', os.sep)
self.image_name = self.opt.image.split(os.sep)[-1]
if self.opt.model_stage == 1:
assert self.opt.task in {1}, "Invalid task!"
assert not self.opt.with_weights, "Weights for stage 1 not available"
elif self.opt.model_stage == 2:
assert self.opt.task in {1, 2, 12}, "Invalid task!"
elif self.opt.model_stage == 3:
assert self.opt.task in {1, 2, 3, 12, 123}, "Invalid task!"
# Model and task set-up
self.num_classes_model = {1: 5, 2: 11, 3: 19}[self.opt.model_stage]
self.task_low, self.task_high = {1: (0, 5), 2: (5, 11), 3: (11, 19), 12: (0, 11), 123: (0, 19)}[self.opt.task]
# Create a conventional ERFNet
self.model = ERFNet(self.num_classes_model, self.opt)
self._load_model()
self.model.to(self.device)
self.model.eval()
# Ground truth
self.metrics = False
if self.opt.ground_truth:
assert os.path.isfile(self.opt.ground_truth), "Invalid ground truth!"
self.metrics = True
self.num_classes_score = self.task_high - self.task_low
self.metric_model = SegmentationRunningScore(self.num_classes_score)
# Output directory
if self.opt.output_path:
if not os.path.isdir(self.opt.output_path):
os.makedirs(self.opt.output_path)
else:
self.opt.output_path = os.path.join(self.opt.image.split(os.sep)[:-1])
image_extension_idx = self.image_name.rfind('.')
segmentation_name = self.image_name[:image_extension_idx] + \
"_seg_stage_{}_task_{}".format(self.opt.model_stage, self.opt.task) + \
self.image_name[image_extension_idx:]
self.output_image = os.path.join(self.opt.output_path, segmentation_name)
ground_truth_name = self.image_name[:image_extension_idx] + \
"_gt_stage_{}_task_{}".format(self.opt.model_stage, self.opt.task) + \
self.image_name[image_extension_idx:]
self.output_gt = os.path.join(self.opt.output_path, ground_truth_name)
# stdout output
print("++++++++++++++++++++++ INIT DEMO ++++++++++++++++++++++++")
print("Image:\t {}".format(self.opt.image))
print("GT:\t {}".format(self.opt.ground_truth))
print("Output:\t {}".format(self.opt.output_path))
print("Stage:\t {}".format(self.opt.model_stage))
print("Weights: {}".format(self.opt.with_weights))
print("Task:\t {}".format(self.opt.task))
print("!!! MIND THAT THE MODELS WERE TRAINED USING AN IMAGE RESOLUTION OF 1024x512px !!!")
# Class colours
labels = labels_cityscape_seg_train3_eval.getlabels()
colors = [(label.trainId - self.task_low, label.color) for label in labels if
label.trainId != 255 and label.trainId in range(0, 19)]
colors.append((255, (0, 0, 0))) # void class
self.id_color = dict(colors)
self.id_color_keys = [key for key in self.id_color.keys()]
self.id_color_vals = [val for val in self.id_color.values()]
class Demo(object):
def __init__(self, options):
self.opt = options
self.device = torch.device("cpu" if self.opt.no_cuda else "cuda")
# Assertions
assert os.path.isfile(self.opt.image), "Invalid image!"
self.opt.image.replace('/', os.sep)
self.opt.image.replace('\\', os.sep)
self.image_name = self.opt.image.split(os.sep)[-1]
if self.opt.model_stage == 1:
assert self.opt.task in {1}, "Invalid task!"
assert not self.opt.with_weights, "Weights for stage 1 not available"
elif self.opt.model_stage == 2:
assert self.opt.task in {1, 2, 12}, "Invalid task!"
elif self.opt.model_stage == 3:
assert self.opt.task in {1, 2, 3, 12, 123}, "Invalid task!"
# Model and task set-up
self.num_classes_model = {1: 5, 2: 11, 3: 19}[self.opt.model_stage]
self.task_low, self.task_high = {1: (0, 5), 2: (5, 11), 3: (11, 19), 12: (0, 11), 123: (0, 19)}[self.opt.task]
# Create a conventional ERFNet
self.model = ERFNet(self.num_classes_model, self.opt)
self._load_model()
self.model.to(self.device)
self.model.eval()
# Ground truth
self.metrics = False
if self.opt.ground_truth:
assert os.path.isfile(self.opt.ground_truth), "Invalid ground truth!"
self.metrics = True
self.num_classes_score = self.task_high - self.task_low
self.metric_model = SegmentationRunningScore(self.num_classes_score)
# Output directory
if self.opt.output_path:
if not os.path.isdir(self.opt.output_path):
os.makedirs(self.opt.output_path)
else:
self.opt.output_path = os.path.join(self.opt.image.split(os.sep)[:-1])
image_extension_idx = self.image_name.rfind('.')
segmentation_name = self.image_name[:image_extension_idx] + \
"_seg_stage_{}_task_{}".format(self.opt.model_stage, self.opt.task) + \
self.image_name[image_extension_idx:]
self.output_image = os.path.join(self.opt.output_path, segmentation_name)
ground_truth_name = self.image_name[:image_extension_idx] + \
"_gt_stage_{}_task_{}".format(self.opt.model_stage, self.opt.task) + \
self.image_name[image_extension_idx:]
self.output_gt = os.path.join(self.opt.output_path, ground_truth_name)
# stdout output
print("++++++++++++++++++++++ INIT DEMO ++++++++++++++++++++++++")
print("Image:\t {}".format(self.opt.image))
print("GT:\t {}".format(self.opt.ground_truth))
print("Output:\t {}".format(self.opt.output_path))
print("Stage:\t {}".format(self.opt.model_stage))
print("Weights: {}".format(self.opt.with_weights))
print("Task:\t {}".format(self.opt.task))
print("!!! MIND THAT THE MODELS WERE TRAINED USING AN IMAGE RESOLUTION OF 1024x512px !!!")
# Class colours
labels = labels_cityscape_seg_train3_eval.getlabels()
colors = [(label.trainId - self.task_low, label.color) for label in labels if
label.trainId != 255 and label.trainId in range(0, 19)]
colors.append((255, (0, 0, 0))) # void class
self.id_color = dict(colors)
self.id_color_keys = [key for key in self.id_color.keys()]
self.id_color_vals = [val for val in self.id_color.values()]
def _load_model(self):
"""Load model from disk
"""
path = self.opt.checkpoint_path
# checkpoint_path = os.path.join("models", "stage_{}".format(self.opt.model_stage))
#assert os.path.isdir(checkpoint_path), \
# "Cannot find folder {}".format(checkpoint_path)
# path = os.path.join(checkpoint_path, "{}.pth".format("with_weights" if self.opt.with_weights else "wout_weights"))
model_dict = self.model.state_dict()
if self.opt.no_cuda:
pretrained_dict = torch.load(path, map_location='cpu')
else:
pretrained_dict = torch.load(path)
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
model_dict.update(pretrained_dict)
self.model.load_state_dict(model_dict)
def process_image(self):
# Required image transformations
resize_interp = transforms.Resize((512, 1024), interpolation=pil.BILINEAR)
transformer = transforms.ToTensor()
normalize = transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225))
# Load Image
image = cv2.imread(self.opt.image)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = pil.fromarray(image)
native_image_size = image.size
# Transform image
image = resize_interp(image)
image = transformer(image)
image = normalize(image).unsqueeze(0).to(self.device)
# Process image
input_rgb = {("color_aug", 0, 0): image}
output = self.model(input_rgb)
# Process network output
pred_seg = output['segmentation_logits'].float()
pred_seg = pred_seg[:, self.task_low:self.task_high, ...]
pred_seg = F.interpolate(pred_seg, (native_image_size[1], native_image_size[0]), mode='nearest')
pred_seg = torch.argmax(pred_seg, dim=1)
pred_seg = pred_seg.cpu().numpy()
# Process ground truth
gt = None
if self.opt.ground_truth:
gt = cv2.imread(self.opt.ground_truth, 0)
gt[gt < self.task_low] = 255
gt[gt >= self.task_high] = 255
gt -= self.task_low
gt[gt == 255 - self.task_low] = 255
gt = np.expand_dims(gt, 0)
self.metric_model.update(gt, pred_seg)
metrics = self.metric_model.get_scores()
self._save_metrics(metrics)
print("\n " + ("{:>8} | " * 2).format("miou", "maccuracy"))
print(("&{: 8.3f} " * 2).format(metrics['meaniou'], metrics['meanacc']) + "\\\\")
# Save prediction to disk
self._save_pred_to_disk(pred_seg, gt)
print("\n-> Done!")
def _save_metrics(self, metrics):
''' Save metrics (class-wise) to disk as HDF5 file.
'''
save_path = os.path.join(self.opt.output_path, "demo.h5")
with h5.File(save_path, 'w') as f:
grp = f
# Write mean_IoU, mean_acc and mean prec to file / group
dset = grp.create_dataset('mean_IoU', data=metrics['meaniou'])
dset.attrs['Description'] = 'See trainIDs for information on the classes'
dset = grp.create_dataset('mean_recall', data=metrics['meanacc'])
dset.attrs['Description'] = 'See trainIDs for information on the classes'
dset.attrs['AKA'] = 'Accuracy -> TP / (TP + FN)'
dset = grp.create_dataset('mean_precision', data=metrics['meanprec'])
dset.attrs['Description'] = 'See trainIDs for information on the classes'
dset.attrs['AKA'] = 'Precision -> TP / (TP + FP)'
ids = np.zeros(shape=(len(metrics['iou'])), dtype=np.uint32)
class_iou = np.zeros(shape=(len(metrics['iou'])), dtype=np.float64)
class_acc = np.zeros(shape=(len(metrics['acc'])), dtype=np.float64)
class_prec = np.zeros(shape=(len(metrics['prec'])), dtype=np.float64)
# Disassemble the dictionary
for key, i in zip(sorted(metrics['iou']), range(len(metrics['iou']))):
class_iou[i] = metrics['iou'][key]
class_acc[i] = metrics['acc'][key]
class_prec[i] = metrics['prec'][key]
# Create class_id dataset only once in first layer of HDF5 file when in 'w' mode
dset = f.create_dataset('trainIDs', data=ids)
dset.attrs['Description'] = 'trainIDs of classes'
dset = grp.create_dataset('class_IoU', data=class_iou)
dset.attrs['Description'] = 'See trainIDs for information on the class order'
dset = grp.create_dataset('class_recall', data=class_acc)
dset.attrs['Description'] = 'See trainIDs for information on the class order'
dset.attrs['AKA'] = 'Accuracy -> TP / (TP + FN)'
dset = grp.create_dataset('class_precision', data=class_prec)
dset.attrs['Description'] = 'See trainIDs for information on the class order'
dset.attrs['AKA'] = 'Precision -> TP / (TP + FP)'
def _save_pred_to_disk(self, pred, gt=None):
''' Save a correctly coloured image of the prediction (batch) to disk.
'''
pred = pred[0]
o_size = pred.shape
single_pred = pred.flatten()
if gt is not None:
single_gt = gt[0].flatten()
single_pred[single_gt == 255] = 255
single_gt = self._convert_to_colour(single_gt, o_size)
cv2.imwrite(self.output_gt, single_gt)
single_pred = self._convert_to_colour(single_pred, o_size)
cv2.imwrite(self.output_image, single_pred)
def _convert_to_colour(self, img, o_size):
''' Replace trainIDs in prediction with colours from dict, reshape it afterwards to input dimensions and
convert RGB to BGR to match openCV's colour system.
'''
sort_idx = np.argsort(self.id_color_keys)
idx = np.searchsorted(self.id_color_keys, img, sorter=sort_idx)
img = np.asarray(self.id_color_vals)[sort_idx][idx]
img = img.astype(np.uint8)
img = np.reshape(img, newshape=(o_size[0], o_size[1], 3))
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
return img
def __init__(self, options):
print(" -> Executing script", os.path.basename(__file__))
self.opt = options
self.device = torch.device("cpu" if self.opt.no_cuda else "cuda")
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LABELS AND CITIES
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
assert self.opt.train_set in {123, 1}, "Invalid train_set!"
keys_to_load = ['color', 'segmentation']
# Labels
if self.opt.train_set == 1:
labels = labels_cityscape_seg_train1.getlabels()
else:
labels = labels_cityscape_seg_train3_eval.getlabels()
# Train IDs
self.train_ids = set([labels[i].trainId for i in range(len(labels))])
self.train_ids.remove(255)
self.num_classes = len(self.train_ids)
# Apply city filter
folders_to_train = CitySet.get_city_set(0)
if self.opt.city:
folders_to_train = CitySet.get_city_set(self.opt.train_set)
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# DATASET DEFINITIONS
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Data augmentation
train_data_transforms = [
mytransforms.RandomHorizontalFlip(),
mytransforms.CreateScaledImage(),
mytransforms.Resize((self.opt.height, self.opt.width),
image_types=keys_to_load),
mytransforms.RandomRescale(1.5),
mytransforms.RandomCrop(
(self.opt.crop_height, self.opt.crop_width)),
mytransforms.ConvertSegmentation(),
mytransforms.CreateColoraug(new_element=True,
scales=self.opt.scales),
mytransforms.ColorJitter(brightness=0.2,
contrast=0.2,
saturation=0.2,
hue=0.1,
gamma=0.0),
mytransforms.RemoveOriginals(),
mytransforms.ToTensor(),
mytransforms.NormalizeZeroMean(),
]
train_dataset = CityscapesDataset(
dataset="cityscapes",
trainvaltest_split='train',
video_mode='mono',
stereo_mode='mono',
scales=self.opt.scales,
labels_mode='fromid',
labels=labels,
keys_to_load=keys_to_load,
data_transforms=train_data_transforms,
video_frames=self.opt.video_frames,
folders_to_load=folders_to_train,
)
self.train_loader = DataLoader(dataset=train_dataset,
batch_size=self.opt.batch_size,
shuffle=True,
num_workers=self.opt.num_workers,
pin_memory=True,
drop_last=True)
val_data_transforms = [
mytransforms.CreateScaledImage(),
mytransforms.Resize((self.opt.height, self.opt.width),
image_types=keys_to_load),
mytransforms.ConvertSegmentation(),
mytransforms.CreateColoraug(new_element=True,
scales=self.opt.scales),
mytransforms.RemoveOriginals(),
mytransforms.ToTensor(),
mytransforms.NormalizeZeroMean(),
]
val_dataset = CityscapesDataset(
dataset=self.opt.dataset,
trainvaltest_split="train",
video_mode='mono',
stereo_mode='mono',
scales=self.opt.scales,
labels_mode='fromid',
labels=labels,
keys_to_load=keys_to_load,
data_transforms=val_data_transforms,
video_frames=self.opt.video_frames,
folders_to_load=CitySet.get_city_set(-1))
self.val_loader = DataLoader(dataset=val_dataset,
batch_size=self.opt.batch_size,
shuffle=False,
num_workers=self.opt.num_workers,
pin_memory=True,
drop_last=True)
self.val_iter = iter(self.val_loader)
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LOGGING OPTIONS
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
print(
"++++++++++++++++++++++ INIT TRAINING ++++++++++++++++++++++++++")
print("Using dataset:\n ", self.opt.dataset, "with split",
self.opt.dataset_split)
print(
"There are {:d} training items and {:d} validation items\n".format(
len(train_dataset), len(val_dataset)))
path_getter = GetPath()
log_path = path_getter.get_checkpoint_path()
self.log_path = os.path.join(log_path, 'erfnet', self.opt.model_name)
self.writers = {}
for mode in ["train", "validation"]:
self.writers[mode] = SummaryWriter(
os.path.join(self.log_path, mode))
# Copy this file to log dir
shutil.copy2(__file__, self.log_path)
print("Training model named:\n ", self.opt.model_name)
print("Models and tensorboard events files are saved to:\n ",
self.log_path)
print("Training is using:\n ", self.device)
print("Training takes place on train set:\n ", self.opt.train_set)
print(
"+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++")
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# MODEL DEFINITION
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Instantiate model
self.model = ERFNet(self.num_classes, self.opt)
self.model.to(self.device)
self.parameters_to_train = self.model.parameters()
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# OPTIMIZER SET-UP
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
self.model_optimizer = optim.Adam(params=self.parameters_to_train,
lr=self.opt.learning_rate,
weight_decay=self.opt.weight_decay)
lambda1 = lambda epoch: pow((1 -
((epoch - 1) / self.opt.num_epochs)), 0.9)
self.model_lr_scheduler = optim.lr_scheduler.LambdaLR(
self.model_optimizer, lr_lambda=lambda1)
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LOSSES
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
self.crossentropy = CrossEntropyLoss(ignore_background=True,
device=self.device)
self.crossentropy.to(self.device)
self.metric_model = SegmentationRunningScore(self.num_classes)
# Save all options to disk and print them to stdout
self.save_opts(len(train_dataset), len(val_dataset))
self._print_options()
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# EVALUATOR DEFINITION
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
if self.opt.validate:
self.evaluator = Evaluator(self.opt, self.model)
class Trainer:
def __init__(self, options):
print(" -> Executing script", os.path.basename(__file__))
self.opt = options
self.device = torch.device("cpu" if self.opt.no_cuda else "cuda")
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LABELS AND CITIES
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
assert self.opt.train_set in {123, 1}, "Invalid train_set!"
keys_to_load = ['color', 'segmentation']
# Labels
if self.opt.train_set == 1:
labels = labels_cityscape_seg_train1.getlabels()
else:
labels = labels_cityscape_seg_train3_eval.getlabels()
# Train IDs
self.train_ids = set([labels[i].trainId for i in range(len(labels))])
self.train_ids.remove(255)
self.num_classes = len(self.train_ids)
# Apply city filter
folders_to_train = CitySet.get_city_set(0)
if self.opt.city:
folders_to_train = CitySet.get_city_set(self.opt.train_set)
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# DATASET DEFINITIONS
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Data augmentation
train_data_transforms = [
mytransforms.RandomHorizontalFlip(),
mytransforms.CreateScaledImage(),
mytransforms.Resize((self.opt.height, self.opt.width),
image_types=keys_to_load),
mytransforms.RandomRescale(1.5),
mytransforms.RandomCrop(
(self.opt.crop_height, self.opt.crop_width)),
mytransforms.ConvertSegmentation(),
mytransforms.CreateColoraug(new_element=True,
scales=self.opt.scales),
mytransforms.ColorJitter(brightness=0.2,
contrast=0.2,
saturation=0.2,
hue=0.1,
gamma=0.0),
mytransforms.RemoveOriginals(),
mytransforms.ToTensor(),
mytransforms.NormalizeZeroMean(),
]
train_dataset = CityscapesDataset(
dataset="cityscapes",
trainvaltest_split='train',
video_mode='mono',
stereo_mode='mono',
scales=self.opt.scales,
labels_mode='fromid',
labels=labels,
keys_to_load=keys_to_load,
data_transforms=train_data_transforms,
video_frames=self.opt.video_frames,
folders_to_load=folders_to_train,
)
self.train_loader = DataLoader(dataset=train_dataset,
batch_size=self.opt.batch_size,
shuffle=True,
num_workers=self.opt.num_workers,
pin_memory=True,
drop_last=True)
val_data_transforms = [
mytransforms.CreateScaledImage(),
mytransforms.Resize((self.opt.height, self.opt.width),
image_types=keys_to_load),
mytransforms.ConvertSegmentation(),
mytransforms.CreateColoraug(new_element=True,
scales=self.opt.scales),
mytransforms.RemoveOriginals(),
mytransforms.ToTensor(),
mytransforms.NormalizeZeroMean(),
]
val_dataset = CityscapesDataset(
dataset=self.opt.dataset,
trainvaltest_split="train",
video_mode='mono',
stereo_mode='mono',
scales=self.opt.scales,
labels_mode='fromid',
labels=labels,
keys_to_load=keys_to_load,
data_transforms=val_data_transforms,
video_frames=self.opt.video_frames,
folders_to_load=CitySet.get_city_set(-1))
self.val_loader = DataLoader(dataset=val_dataset,
batch_size=self.opt.batch_size,
shuffle=False,
num_workers=self.opt.num_workers,
pin_memory=True,
drop_last=True)
self.val_iter = iter(self.val_loader)
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LOGGING OPTIONS
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
print(
"++++++++++++++++++++++ INIT TRAINING ++++++++++++++++++++++++++")
print("Using dataset:\n ", self.opt.dataset, "with split",
self.opt.dataset_split)
print(
"There are {:d} training items and {:d} validation items\n".format(
len(train_dataset), len(val_dataset)))
path_getter = GetPath()
log_path = path_getter.get_checkpoint_path()
self.log_path = os.path.join(log_path, 'erfnet', self.opt.model_name)
self.writers = {}
for mode in ["train", "validation"]:
self.writers[mode] = SummaryWriter(
os.path.join(self.log_path, mode))
# Copy this file to log dir
shutil.copy2(__file__, self.log_path)
print("Training model named:\n ", self.opt.model_name)
print("Models and tensorboard events files are saved to:\n ",
self.log_path)
print("Training is using:\n ", self.device)
print("Training takes place on train set:\n ", self.opt.train_set)
print(
"+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++")
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# MODEL DEFINITION
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# Instantiate model
self.model = ERFNet(self.num_classes, self.opt)
self.model.to(self.device)
self.parameters_to_train = self.model.parameters()
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# OPTIMIZER SET-UP
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
self.model_optimizer = optim.Adam(params=self.parameters_to_train,
lr=self.opt.learning_rate,
weight_decay=self.opt.weight_decay)
lambda1 = lambda epoch: pow((1 -
((epoch - 1) / self.opt.num_epochs)), 0.9)
self.model_lr_scheduler = optim.lr_scheduler.LambdaLR(
self.model_optimizer, lr_lambda=lambda1)
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# LOSSES
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
self.crossentropy = CrossEntropyLoss(ignore_background=True,
device=self.device)
self.crossentropy.to(self.device)
self.metric_model = SegmentationRunningScore(self.num_classes)
# Save all options to disk and print them to stdout
self.save_opts(len(train_dataset), len(val_dataset))
self._print_options()
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# EVALUATOR DEFINITION
# +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
if self.opt.validate:
self.evaluator = Evaluator(self.opt, self.model)
def set_train(self):
"""Convert all models to training mode
"""
self.model.train()
def set_eval(self):
"""Convert all models to testing/evaluation mode
"""
self.model.eval()
def train(self):
"""Run the entire training pipeline
"""
self.epoch = 0
self.step = 0
self.start_time = time.time()
for self.epoch in range(self.opt.num_epochs):
self.run_epoch()
if (self.epoch + 1) % self.opt.save_frequency == 0:
self.save_model()
if self.opt.validate and (self.epoch +
1) % self.opt.val_frequency == 0:
self.run_eval()
def run_epoch(self):
"""Run a single epoch of training and validation
"""
print("Training")
self.set_train()
for batch_idx, inputs in enumerate(self.train_loader):
before_op_time = time.time()
outputs, losses = self.process_batch(inputs)
self.model_optimizer.zero_grad()
losses["loss"].backward()
self.model_optimizer.step()
duration = time.time() - before_op_time
# log less frequently after the first 2000 steps to save time & disk space
early_phase = batch_idx % self.opt.log_frequency == 0 and self.step < 2000
late_phase = self.step % 2000 == 0
if early_phase or late_phase:
if ('segmentation', 0, 0) in inputs.keys():
metrics = self.compute_segmentation_losses(inputs, outputs)
self.log_time(batch_idx, duration,
losses["loss"].cpu().data,
metrics["meaniou"], metrics["meanacc"])
else:
self.log_time(batch_idx, duration,
losses["loss"].cpu().data, 0, 0)
metrics = {}
self.log("train", losses, metrics)
self.val()
self.step += 1
self.model_lr_scheduler.step()
def run_eval(self):
print("Validating on full validation set")
self.set_eval()
self.evaluator.calculate_metrics(self.epoch)
def val(self):
"""Validate the model on a single minibatch
"""
self.set_eval()
try:
inputs_val = self.val_iter.next()
except StopIteration:
self.val_iter = iter(self.val_loader)
inputs_val = self.val_iter.next()
with torch.no_grad():
outputs_val, losses_val = self.process_batch(inputs_val)
if ('segmentation', 0, 0) in inputs_val:
metrics_val = self.compute_segmentation_losses(
inputs_val, outputs_val)
else:
metrics_val = {}
self.log("validation", losses_val, metrics_val)
self.set_train()
def process_batch(self, inputs):
"""Pass a minibatch through the network and generate images and losses
"""
for key, ipt in inputs.items():
inputs[key] = ipt.to(self.device)
outputs = self.model(inputs)
losses = self.compute_losses(inputs, outputs)
return outputs, losses
def compute_losses(self, inputs, outputs):
"""Compute the reprojection and smoothness losses for a minibatch
"""
losses = {}
preds = F.log_softmax(outputs['segmentation_logits'].float(), dim=1)
targets = inputs[('segmentation', 0, 0)][:, 0, :, :].long()
cross_loss = self.crossentropy(preds, targets)
losses["loss"] = cross_loss
return losses
def compute_segmentation_losses(self, inputs, outputs):
"""Compute the loss metrics based on the current prediction
"""
label_true = np.array(inputs[('segmentation', 0, 0)].cpu())[:, 0, :, :]
label_pred = np.array(outputs['segmentation'].detach().cpu())
self.metric_model.update(label_true, label_pred)
metrics = self.metric_model.get_scores()
self.metric_model.reset()
return metrics
def log_time(self, batch_idx, duration, loss, miou, acc):
"""Print a logging statement to the terminal
"""
samples_per_sec = self.opt.batch_size / duration
print_string = "epoch {:>3} | batch {:>6} | examples/s: {:5.1f}" + \
" | loss: {:.5f}| meaniou: {:.5f}| meanacc: {:.5f}"
print(
print_string.format(self.epoch, batch_idx, samples_per_sec, loss,
miou, acc))
def log(self, mode, losses, metrics):
"""Write an event to the tensorboard events file
"""
writer = self.writers[mode]
for l, v in losses.items():
writer.add_scalar("{}".format(l), v, self.step)
for l, v in metrics.items():
if l in {'iou', 'acc', 'prec'}:
continue
writer.add_scalar("{}".format(l), v, self.step)
def save_opts(self, n_train, n_eval):
"""Save options to disk so we know what we ran this experiment with
"""
models_dir = os.path.join(self.log_path, "models")
if not os.path.exists(models_dir):
os.makedirs(models_dir)
to_save = self.opt.__dict__.copy()
to_save['n_train'] = n_train
to_save['n_eval'] = n_eval
with open(os.path.join(models_dir, 'opt.json'), 'w') as f:
json.dump(to_save, f, indent=2)
def save_model(self):
"""Save model weights to disk
"""
save_folder = os.path.join(self.log_path, "models",
"weights_{}".format(self.epoch))
if not os.path.exists(save_folder):
os.makedirs(save_folder)
save_path = os.path.join(save_folder, "{}.pth".format("model"))
to_save = self.model.state_dict()
torch.save(to_save, save_path)
save_path = os.path.join(save_folder, "{}.pth".format("optim"))
torch.save(self.model_optimizer.state_dict(), save_path)
def load_model(self, adam=True):
"""Load model(s) from disk
:param adam: whether to load the Adam state too
"""
base_path = os.path.split(self.log_path)[0]
checkpoint_path = os.path.join(
base_path, self.opt.load_model_name, 'models',
'weights_{}'.format(self.opt.weights_epoch))
assert os.path.isdir(checkpoint_path), \
"Cannot find folder {}".format(checkpoint_path)
print("loading model from folder {}".format(checkpoint_path))
path = os.path.join(checkpoint_path, "{}.pth".format('model'))
model_dict = self.model.state_dict()
pretrained_dict = torch.load(path)
pretrained_dict = {
k: v
for k, v in pretrained_dict.items() if k in model_dict
}
model_dict.update(pretrained_dict)
self.model.load_state_dict(model_dict)
if adam:
# loading adam state
optimizer_load_path = os.path.join(checkpoint_path,
"{}.pth".format("optim"))
if os.path.isfile(optimizer_load_path):
print("Loading Adam weights")
optimizer_dict = torch.load(optimizer_load_path)
self.model_optimizer.load_state_dict(optimizer_dict)
else:
print(
"Cannot find Adam weights so Adam is randomly initialized")
def _print_options(self):
"""Print training options to stdout so that they appear in the SLURM log
"""
# Convert namespace to dictionary
opts = vars(self.opt)
# Get max key length for left justifying
max_len = max([len(key) for key in opts.keys()])
# Print options to stdout
print(
"+++++++++++++++++++++++++++ OPTIONS +++++++++++++++++++++++++++")
for item in sorted(opts.items()):
print(item[0].ljust(max_len), item[1])
print(
"+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++")
def __init__(self, remaps=('none',)):
self.scores = dict(
(remap_name, SegmentationRunningScore(self._remap_len(remap_name))) # scores = dict mit remap name und leerer confusion matrix
for remap_name in remaps
)
def __init__(self, remaps=('none',)):
self.scores = dict(
(remap_name, SegmentationRunningScore(self._remap_len(remap_name)))
for remap_name in remaps
)