def inferHuman36Data(self, batch, model_type, device, config, randomize_n_views,
min_n_views,
max_n_views):
"""
For batch inferences
"""
outputBatch = {}
inputBatch = {}
collatFunction = dataset_utils.make_collate_fn(randomize_n_views,
min_n_views,
max_n_views)
batch = collatFunction(batch)
images_batch, keypoints_3d_gt, keypoints_3d_validity_gt, proj_matricies_batch = dataset_utils.prepare_batch(batch, device, config)
#print(proj_matricies_batch,proj_matricies_batch.shape,len(batch),images_batch.shape)
keypoints_2d_pred, cuboids_pred, base_points_pred, volumes_pred, coord_volumes_pred = None, None, None, None, None
if model_type == "alg" or model_type == "ransac":
keypoints_3d_pred, keypoints_2d_pred, heatmaps_pred, confidences_pred = self.model(images_batch, proj_matricies_batch, batch)
elif model_type == "vol":
keypoints_3d_pred, heatmaps_pred, volumes_pred, confidences_pred, cuboids_pred, coord_volumes_pred, base_points_pred = self.model(images_batch, proj_matricies_batch, batch)
outputBatch["keypoints_3d_pred"] = keypoints_3d_pred
outputBatch["heatmaps_pred"] = heatmaps_pred
outputBatch["volumes_pred"] = volumes_pred
outputBatch["confidences_pred"] = confidences_pred
outputBatch["cuboids_pred"] = confidences_pred
outputBatch["coord_volumes_pred"] = coord_volumes_pred
outputBatch["base_points_pred"] = base_points_pred
inputBatch["images_batch"] = images_batch
inputBatch["proj_matricies_batch"] = proj_matricies_batch
return outputBatch, inputBatch
def setup_cmu_dataloaders(config):
# val
val_dataset = cmupanoptic.CMUPanopticDataset(
cmu_root=config.dataset.val.cmu_root,
pred_results_path=config.dataset.val.pred_results_path if hasattr(
config.dataset.val, "pred_results_path") else None,
train=False,
test=True,
image_shape=config.image_shape if hasattr(config, "image_shape") else
(256, 256),
labels_path=config.dataset.val.labels_path,
retain_every_n_frames_in_test=config.dataset.val.
retain_every_n_frames_in_test,
scale_bbox=config.dataset.val.scale_bbox,
square_bbox=config.dataset.val.square_bbox if hasattr(
config.dataset.val, "square_bbox") else True,
kind=config.kind,
choose_cameras=config.dataset.val.choose_cameras if hasattr(
config.dataset.val, "choose_cameras") else [],
ignore_cameras=config.dataset.val.ignore_cameras if hasattr(
config.dataset.val, "ignore_cameras") else [],
crop=config.dataset.val.crop
if hasattr(config.dataset.val, "crop") else True,
frames_split_file=config.opt.frames_split_file if hasattr(
config.opt, "frames_split_file") else None)
val_dataloader = DataLoader(
val_dataset,
batch_size=config.opt.val_batch_size if hasattr(
config.opt, "val_batch_size") else config.opt.batch_size,
shuffle=config.dataset.val.shuffle,
collate_fn=dataset_utils.make_collate_fn(
randomize_n_views=config.dataset.val.randomize_n_views,
min_n_views=config.dataset.val.min_n_views,
max_n_views=config.dataset.val.max_n_views),
num_workers=config.dataset.val.num_workers,
worker_init_fn=dataset_utils.worker_init_fn,
pin_memory=True,
drop_last=False)
return val_dataloader
def setup_human36m_dataloaders(config):
# val
val_dataset = human36m.Human36MMultiViewDataset(
h36m_root=config.dataset.val.h36m_root,
pred_results_path=config.dataset.val.pred_results_path if hasattr(
config.dataset.val, "pred_results_path") else None,
train=False,
test=True,
image_shape=config.image_shape if hasattr(config, "image_shape") else
(256, 256),
labels_path=config.dataset.val.labels_path,
with_damaged_actions=config.dataset.val.with_damaged_actions,
retain_every_n_frames_in_test=config.dataset.val.
retain_every_n_frames_in_test,
scale_bbox=config.dataset.val.scale_bbox,
kind=config.kind,
undistort_images=config.dataset.val.undistort_images,
ignore_cameras=config.dataset.val.ignore_cameras if hasattr(
config.dataset.val, "ignore_cameras") else [],
crop=config.dataset.val.crop
if hasattr(config.dataset.val, "crop") else True,
)
val_dataloader = DataLoader(
val_dataset,
batch_size=config.opt.val_batch_size if hasattr(
config.opt, "val_batch_size") else config.opt.batch_size,
shuffle=config.dataset.val.shuffle,
collate_fn=dataset_utils.make_collate_fn(
randomize_n_views=config.dataset.val.randomize_n_views,
min_n_views=config.dataset.val.min_n_views,
max_n_views=config.dataset.val.max_n_views),
num_workers=config.dataset.val.num_workers,
worker_init_fn=dataset_utils.worker_init_fn,
pin_memory=True,
drop_last=False)
return val_dataloader
def setup_human36m_dataloaders(config, is_train, distributed_train):
train_dataloader = None
if is_train:
# train
train_dataset = human36m.Human36MMultiViewDataset(
h36m_root=config.dataset.train.h36m_root,
pred_results_path=config.dataset.train.pred_results_path if hasattr(config.dataset.train, "pred_results_path") else None,
train=True,
test=False,
image_shape=config.image_shape if hasattr(config, "image_shape") else (256, 256),
labels_path=config.dataset.train.labels_path,
with_damaged_actions=config.dataset.train.with_damaged_actions,
scale_bbox=config.dataset.train.scale_bbox,
kind=config.kind,
undistort_images=config.dataset.train.undistort_images,
ignore_cameras=config.dataset.train.ignore_cameras if hasattr(config.dataset.train, "ignore_cameras") else [],
crop=config.dataset.train.crop if hasattr(config.dataset.train, "crop") else True,
)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) if distributed_train else None
train_dataloader = DataLoader(
train_dataset,
batch_size=config.opt.batch_size,
shuffle=config.dataset.train.shuffle and (train_sampler is None), # debatable
sampler=train_sampler,
collate_fn=dataset_utils.make_collate_fn(randomize_n_views=config.dataset.train.randomize_n_views,
min_n_views=config.dataset.train.min_n_views,
max_n_views=config.dataset.train.max_n_views),
num_workers=config.dataset.train.num_workers,
worker_init_fn=dataset_utils.worker_init_fn,
pin_memory=True
)
# val
val_dataset = human36m.Human36MMultiViewDataset(
h36m_root=config.dataset.val.h36m_root,
pred_results_path=config.dataset.val.pred_results_path if hasattr(config.dataset.val, "pred_results_path") else None,
train=False,
test=True,
image_shape=config.image_shape if hasattr(config, "image_shape") else (256, 256),
labels_path=config.dataset.val.labels_path,
with_damaged_actions=config.dataset.val.with_damaged_actions,
retain_every_n_frames_in_test=config.dataset.val.retain_every_n_frames_in_test,
scale_bbox=config.dataset.val.scale_bbox,
kind=config.kind,
undistort_images=config.dataset.val.undistort_images,
ignore_cameras=config.dataset.val.ignore_cameras if hasattr(config.dataset.val, "ignore_cameras") else [],
crop=config.dataset.val.crop if hasattr(config.dataset.val, "crop") else True,
)
val_dataloader = DataLoader(
val_dataset,
batch_size=config.opt.val_batch_size if hasattr(config.opt, "val_batch_size") else config.opt.batch_size,
shuffle=config.dataset.val.shuffle,
collate_fn=dataset_utils.make_collate_fn(randomize_n_views=config.dataset.val.randomize_n_views,
min_n_views=config.dataset.val.min_n_views,
max_n_views=config.dataset.val.max_n_views),
num_workers=config.dataset.val.num_workers,
worker_init_fn=dataset_utils.worker_init_fn,
pin_memory=True
)
return train_dataloader, val_dataloader, train_sampler
def setup_cmu_dataloaders(config, is_train, distributed_train):
train_dataloader = None
if is_train:
# train
train_dataset = cmupanoptic.CMUPanopticDataset(
cmu_root=config.dataset.train.cmu_root,
pred_results_path=config.dataset.train.pred_results_path if hasattr(config.dataset.train, "pred_results_path") else None,
train=True,
test=False,
image_shape=config.image_shape if hasattr(config, "image_shape") else (256, 256),
labels_path=config.dataset.train.labels_path,
scale_bbox=config.dataset.train.scale_bbox,
square_bbox=config.dataset.train.square_bbox if hasattr(config.dataset.train, "square_bbox") else True,
kind=config.kind,
transfer_cmu_to_human36m=config.model.transfer_cmu_to_human36m if hasattr(config.model, "transfer_cmu_to_human36m") else False,
choose_cameras=config.dataset.train.choose_cameras if hasattr(config.dataset.train, "choose_cameras") else [],
ignore_cameras=config.dataset.train.ignore_cameras if hasattr(config.dataset.train, "ignore_cameras") else [],
crop=config.dataset.train.crop if hasattr(config.dataset.train, "crop") else True,
frames_split_file=config.opt.frames_split_file if hasattr(config.opt, "frames_split_file") else None
)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) if distributed_train else None
train_dataloader = DataLoader(
train_dataset,
batch_size=config.opt.batch_size,
shuffle=config.dataset.train.shuffle and (train_sampler is None), # debatable
sampler=train_sampler,
collate_fn=dataset_utils.make_collate_fn(randomize_n_views=config.dataset.train.randomize_n_views,
min_n_views=config.dataset.train.min_n_views,
max_n_views=config.dataset.train.max_n_views),
num_workers=config.dataset.train.num_workers,
worker_init_fn=dataset_utils.worker_init_fn,
pin_memory=True,
drop_last=False
)
# val
val_dataset = cmupanoptic.CMUPanopticDataset(
cmu_root=config.dataset.val.cmu_root,
pred_results_path=config.dataset.val.pred_results_path if hasattr(
config.dataset.val, "pred_results_path") else None,
train=False,
test=True,
image_shape=config.image_shape if hasattr(config, "image_shape") else (256, 256),
labels_path=config.dataset.val.labels_path,
retain_every_n_frames_in_test=config.dataset.val.retain_every_n_frames_in_test,
scale_bbox=config.dataset.val.scale_bbox,
square_bbox=config.dataset.val.square_bbox if hasattr(config.dataset.val, "square_bbox") else True,
kind=config.kind,
choose_cameras=config.dataset.val.choose_cameras if hasattr(config.dataset.val, "choose_cameras") else [],
ignore_cameras=config.dataset.val.ignore_cameras if hasattr(config.dataset.val, "ignore_cameras") else [],
crop=config.dataset.val.crop if hasattr(config.dataset.val, "crop") else True,
frames_split_file=config.opt.frames_split_file if hasattr(config.opt, "frames_split_file") else None
)
val_dataloader = DataLoader(
val_dataset,
batch_size=config.opt.val_batch_size if hasattr(config.opt, "val_batch_size") else config.opt.batch_size,
shuffle=config.dataset.val.shuffle,
collate_fn=dataset_utils.make_collate_fn(randomize_n_views=config.dataset.val.randomize_n_views,
min_n_views=config.dataset.val.min_n_views,
max_n_views=config.dataset.val.max_n_views),
num_workers=config.dataset.val.num_workers,
worker_init_fn=dataset_utils.worker_init_fn,
pin_memory=True,
drop_last=False
)
return train_dataloader, val_dataloader, train_sampler
def setup_human36m_dataloaders(config,
is_train,
distributed_train,
rank=None,
world_size=None):
train_dataloader = None
if is_train:
# train
train_dataset = eval('datasets.' + config.dataset.train_dataset)(
root=config.dataset.root,
pred_results_path=config.train.pred_results_path,
train=True,
test=False,
image_shape=config.model.image_shape,
labels_path=config.dataset.train_labels_path,
with_damaged_actions=config.train.with_damaged_actions,
scale_bbox=config.train.scale_bbox,
kind=config.kind,
undistort_images=config.train.undistort_images,
ignore_cameras=config.train.ignore_cameras,
crop=config.train.crop,
erase=config.train.erase,
data_format=config.dataset.data_format)
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset) if distributed_train else None
train_dataloader = DataLoader(
train_dataset,
batch_size=config.train.batch_size,
shuffle=config.train.shuffle
and (train_sampler is None), # debatable
sampler=train_sampler,
collate_fn=dataset_utils.make_collate_fn(
randomize_n_views=config.train.randomize_n_views,
min_n_views=config.train.min_n_views,
max_n_views=config.train.max_n_views),
num_workers=config.train.num_workers,
worker_init_fn=dataset_utils.worker_init_fn,
pin_memory=True)
# val
val_dataset = eval('datasets.' + config.dataset.val_dataset)(
root=config.dataset.root,
pred_results_path=config.val.pred_results_path,
train=False,
test=True,
image_shape=config.model.image_shape,
labels_path=config.dataset.val_labels_path,
with_damaged_actions=config.val.with_damaged_actions,
retain_every_n_frames_in_test=config.val.retain_every_n_frames_in_test,
scale_bbox=config.val.scale_bbox,
kind=config.kind,
undistort_images=config.val.undistort_images,
ignore_cameras=config.val.ignore_cameras,
crop=config.val.crop,
erase=config.val.erase,
rank=rank,
world_size=world_size,
data_format=config.dataset.data_format)
val_dataloader = DataLoader(
val_dataset,
batch_size=config.val.batch_size,
shuffle=config.val.shuffle,
collate_fn=dataset_utils.make_collate_fn(
randomize_n_views=config.val.randomize_n_views,
min_n_views=config.val.min_n_views,
max_n_views=config.val.max_n_views),
num_workers=config.val.num_workers,
worker_init_fn=dataset_utils.worker_init_fn,
pin_memory=True)
return train_dataloader, val_dataloader, train_sampler, val_dataset.dist_size