print(
f'VGE-NV-ImageOnly training: image limit: {IMAGE_LIMIT} bs: {BATCH_SIZE} lr gamma: {LR_GAMMA} embed-dim: {EMBED_DIM} shuffle: {SHUFFLE} margin: {MARGIN} lr:{LR}'
)
transform = transforms.Compose([
#transforms.Resize((950,1000)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224,
0.225]),
])
data_set = Semantic3dDataset('data/pointcloud_images_o3d_merged',
'train',
transform=transform,
image_limit=IMAGE_LIMIT,
load_viewObjects=True,
load_sceneGraphs=True,
return_graph_data=True)
#Option: shuffle, pin_memory crashes on my system, CARE: shuffle for PairWiseRankingLoss(!)
data_loader = DataLoader(data_set,
batch_size=BATCH_SIZE,
num_workers=2,
pin_memory=False,
shuffle=SHUFFLE)
loss_dict = {}
best_loss = np.inf
best_model = None
#for lr in (5e-4,1e-4,5e-5,1e-5):
# #Append the average pos&ori. errors *for the cases that the scene was hit*
# pos_results[k].append( np.mean( topk_pos_dists[scene_correct==True]) if np.sum(scene_correct)>0 else None )
# ori_results[k].append( np.mean( topk_ori_dists[scene_correct==True]) if np.sum(scene_correct)>0 else None )
# scene_results[k].append( np.mean(scene_correct) ) #Always append the scene-scores
# assert len(pos_results[k])==len(ori_results[k])==len(scene_results[k])==len(test_indices)
# print('Saving retrieval results...')
# pickle.dump(retrieval_dict, open('retrievals_netvlad_plus_sceneGraph.pkl','wb'))
# return evaluate_topK(pos_results, ori_results, scene_results)
if __name__ == "__main__":
IMAGE_LIMIT=3000
dataset_train=Semantic3dDataset('data/pointcloud_images_o3d_merged','train',transform=None, image_limit=IMAGE_LIMIT, load_viewObjects=True, load_sceneGraphs=True)
dataset_test =Semantic3dDataset('data/pointcloud_images_o3d_merged','test', transform=None, image_limit=IMAGE_LIMIT, load_viewObjects=True, load_sceneGraphs=True)
if 'gather' in sys.argv:
gather_sceneGraph2viewObjects(dataset_train, dataset_test, ablation=None)
# gather_sceneGraph2viewObjects(dataset_train, dataset_test, ablation='colors')
# gather_sceneGraph2viewObjects(dataset_train, dataset_test, ablation='relationships')
#gather_sceneGraph2sceneGraph(dataset_train, dataset_test)
if 'gather-occ' in sys.argv:
IMAGE_LIMIT=3000
dataset_train=Semantic3dDataset('data/pointcloud_images_o3d_merged_occ','train',transform=None, image_limit=IMAGE_LIMIT, load_viewObjects=True, load_sceneGraphs=True)
dataset_test =Semantic3dDataset('data/pointcloud_images_o3d_merged_occ','test', transform=None, image_limit=IMAGE_LIMIT, load_viewObjects=True, load_sceneGraphs=True)
gather_sceneGraph2viewObjects(dataset_train, dataset_test, ablation=None)
cv2.imwrite(f'cases_hit_{i}.png', img_hit)
cv2.imwrite(f'cases_miss_{i}.png', img_miss)
cv2.waitKey()
#Results sanity-checked (NetVLAD) ✓
#CARE: retrievals in main or retrieval dir, re-create if in doubt!
if __name__ == "__main__":
IMAGE_LIMIT = 3000
BATCH_SIZE = 6
NUM_CLUSTERS = 8
TEST_SPLIT = 4
ALPHA = 10.0
dataset_train = Semantic3dDataset('data/pointcloud_images_o3d_merged',
'train',
transform=None,
image_limit=IMAGE_LIMIT)
dataset_test = Semantic3dDataset('data/pointcloud_images_o3d_merged',
'test',
transform=None,
image_limit=IMAGE_LIMIT)
retrievals_netvlad = pickle.load(open('retrievals_NV-S3D.pkl', 'rb'))
retrievals_sg_scoring = pickle.load(open('retrievals_PureSG.pkl', 'rb'))
quit()
results_hit, results_miss = get_top3_all_scene_hit_miss(
retrievals_sg_scoring, retrievals_netvlad, dataset_train, dataset_test)
print(len(results_hit), len(results_miss))
for key in results_hit.keys():
print('hit', results_hit[key], 'miss', results_miss[key])