Upload pretrained models

lido-kitti/MinkowskiNet-semantickitti.yaml

@@ -0,0 +1,69 @@
+# Config format schema number
+format_version: 1
+
+###################
+## Model options
+model_params:
+  model_architecture: "minkowskinet"
+
+  input_dims: 4
+  voxel_size: 0.05 # 0.05 for SemanticKITTI (?)
+  cr: 1
+  layer_num:
+    - 32
+    - 32
+    - 64
+    - 128
+    - 256
+    - 256
+    - 128
+    - 96
+    - 96
+
+###################
+## Dataset options
+dataset:
+  pc_dataset_type: "SemanticKITTI"
+  collate_type: "sparse_collate_fn"
+  ignore_label: 0
+  data_config: "./config/labels/semantic-kitti.yaml" # semantic-kitti-sdata.yaml for small data setup
+  num_classes: 20
+  sensor:
+    name: "HDL64"
+    type: "spherical" # projective
+    fov_up: 3
+    fov_down: -25
+
+###################
+## Train params
+train:
+  epochs: 64 # default 128
+  learning_rate: 2.4e-1 # 1.0e-2 for SGD | 1.0e-4 for AdamW
+  weight_decay: 1.0e-4
+  optimizer: SGD # [SGD, AdamW]
+  batch_size: 4   # batch size (default 8)
+  workers: 4     # number of threads to get data
+  epsilon_w: 0.001       # class weight w = 1 / (content + epsilon_w)
+  momentum: 0.9
+  nesterov: True
+
+  mav_loss: True  # features loss
+  cont_loss: True  # contrastive loss
+  obj_loss: True  # objectosphere loss
+
+  anomaly_ratio: [0.0, 1.0] # ratio of anomalous samples in train/val
+
+  scheduler:
+    name: "WarmupCosine"  # [OneCycle, WarmupCosine]
+    #OneCycleLR: # Old decay with warmup and cosine annealing 
+    max_lr: 0.01 # Equal to optimizer.lr
+    total_steps: 1000 # Equal to max_epochs * iterations_per_epoch
+    pct_start: 0.02 # The percentage of the cycle (in number of steps) spent increasing the learning rate (warmup).
+
+  report_epoch: 1 # report every epoch
+  show_scans: False # show scans during training
+
+  #lr_scheduler: CosineAnnealingWarmRestarts  # [StepLR, ReduceLROnPlateau, CosineAnnealingLR, CosineAnnealingWarmRestarts]
+  #momentum: 0.9
+  #nesterov: True
+  #weight_decay: 1.0e-4

lido-kitti/checkpoint-best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bcac3e8525d7d45fde94e33506c35e1119105d34e78b5b8e1a9cdd10f186f4b1
+size 174020824

lido-kitti/mavs.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c67fe7f61aaf3a8b9c8fbd739a469dffa23240bcc19a3dedd8479be46f83407
+size 7611

lido-kitti/vars.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c46b2188ddc7d41cc24ef0169fb8383020b978d3401845b960e3ec6167bbb034
+size 7535

lido-nuscenes/MinkowskiNet-nuscenes.yaml

@@ -0,0 +1,69 @@
+# Config format schema number
+format_version: 1
+
+###################
+## Model options
+model_params:
+  model_architecture: "minkowskinet"
+
+  input_dims: 4
+  voxel_size: 0.05 # 0.05 for SemanticKITTI (?)
+  cr: 1
+  layer_num:
+    - 32
+    - 32
+    - 64
+    - 128
+    - 256
+    - 256
+    - 128
+    - 96
+    - 96
+
+###################
+## Dataset options
+dataset:
+  pc_dataset_type: "nuScenes"
+  collate_type: "sparse_collate_fn"
+  ignore_label: 0
+  data_config: "./config/labels/nuscenes.yaml"
+  num_classes: 17
+  sensor:
+    name: "HDL32"
+    type: "spherical" # projective
+    fov_up: 10
+    fov_down: -30
+
+###################
+## Train params
+train:
+  epochs: 64 # default 128
+  learning_rate: 2.4e-1 # 1.0e-2 for SGD | 1.0e-4 for AdamW
+  weight_decay: 1.0e-4
+  optimizer: SGD # [SGD, AdamW]
+  batch_size: 4   # batch size (default 8)
+  workers: 4     # number of threads to get data
+  epsilon_w: 0.001       # class weight w = 1 / (content + epsilon_w)
+  momentum: 0.9
+  nesterov: True
+
+  mav_loss: True  # features loss
+  cont_loss: True  # contrastive loss
+  obj_loss: True  # objectosphere loss
+
+  anomaly_ratio: [0.0, 0.0] # ratio of anomalous samples in train/val
+
+  scheduler:
+    name: "WarmupCosine"  # [OneCycle, WarmupCosine]
+    #OneCycleLR: # Old decay with warmup and cosine annealing 
+    max_lr: 0.01 # Equal to optimizer.lr
+    total_steps: 1000 # Equal to max_epochs * iterations_per_epoch
+    pct_start: 0.02 # The percentage of the cycle (in number of steps) spent increasing the learning rate (warmup).
+
+  report_epoch: 1 # report every epoch
+  show_scans: False # show scans during training
+
+  #lr_scheduler: CosineAnnealingWarmRestarts  # [StepLR, ReduceLROnPlateau, CosineAnnealingLR, CosineAnnealingWarmRestarts]
+  #momentum: 0.9
+  #nesterov: True
+  #weight_decay: 1.0e-4

lido-nuscenes/checkpoint-best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3cc99b3694b6dbab30d19b65e1b49d0fb3a9ea0a1634cb5b3f97b5809c73464
+size 174016216

lido-nuscenes/mavs.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a821a0d3025f127bdf4887503861b27b70c1453b3f393405b1e5be4d3838cfde
+size 6282

lido-nuscenes/vars.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:399e72044d473eaa77abd2ebe9375ab37666b3526e19f207ab1cc94b56da0502
+size 6218

lido-poss/MinkowskiNet-semanticposs.yaml

@@ -0,0 +1,69 @@
+# Config format schema number
+format_version: 1
+
+###################
+## Model options
+model_params:
+  model_architecture: "minkowskinet"
+
+  input_dims: 4
+  voxel_size: 0.05 # 0.05 for SemanticPOSS (?)
+  cr: 1
+  layer_num:
+    - 32
+    - 32
+    - 64
+    - 128
+    - 256
+    - 256
+    - 128
+    - 96
+    - 96
+
+###################
+## Dataset options
+dataset:
+  pc_dataset_type: "SemanticPOSS"
+  collate_type: "sparse_collate_fn"
+  ignore_label: 0
+  data_config: "./config/labels/semantic-poss.yaml"
+  num_classes: 14
+  sensor:
+    name: "HDL64"
+    type: "spherical" # projective
+    fov_up: 7
+    fov_down: -16
+
+###################
+## Train params
+train:
+  epochs: 64 # default 128
+  learning_rate: 2.4e-1 # 1.0e-2 for SGD | 1.0e-4 for AdamW
+  weight_decay: 1.0e-4
+  optimizer: SGD # [SGD, AdamW]
+  batch_size: 4   # batch size (default 8)
+  workers: 4     # number of threads to get data
+  epsilon_w: 0.001       # class weight w = 1 / (content + epsilon_w)
+  momentum: 0.9
+  nesterov: True
+
+  mav_loss: True  # features loss
+  cont_loss: True  # contrastive loss
+  obj_loss: True  # objectosphere loss
+
+  anomaly_ratio: [0.0, 0.0] # ratio of anomalous samples in train/val
+
+  scheduler:
+    name: "WarmupCosine"  # [OneCycle, WarmupCosine]
+    #OneCycleLR: # Old decay with warmup and cosine annealing 
+    max_lr: 0.01 # Equal to optimizer.lr
+    total_steps: 1000 # Equal to max_epochs * iterations_per_epoch
+    pct_start: 0.02 # The percentage of the cycle (in number of steps) spent increasing the learning rate (warmup).
+
+  report_epoch: 1 # report every epoch
+  show_scans: False # show scans during training
+
+  #lr_scheduler: CosineAnnealingWarmRestarts  # [StepLR, ReduceLROnPlateau, CosineAnnealingLR, CosineAnnealingWarmRestarts]
+  #momentum: 0.9
+  #nesterov: True
+  #weight_decay: 1.0e-4

lido-poss/checkpoint-best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:518c84e0d07d2225965f2ce3f9c9610670d0153ee45cec3ca15b459419362b49
+size 174011608

lido-poss/mavs.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:adb74423b340313bc6fef71c613dcd88204e486f7eb109673f92564405260bd9
+size 5025

lido-poss/vars.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e69b439fb29dea4e5c5b100190ba25600bedb438d92117a8a8f6141d2c40933
+size 4973

lido-stu/MinkowskiNet-semantickitti.yaml

@@ -0,0 +1,69 @@
+# Config format schema number
+format_version: 1
+
+###################
+## Model options
+model_params:
+  model_architecture: "minkowskinet"
+
+  input_dims: 4
+  voxel_size: 0.05 # 0.05 for SemanticKITTI (?)
+  cr: 1
+  layer_num:
+    - 32
+    - 32
+    - 64
+    - 128
+    - 256
+    - 256
+    - 128
+    - 96
+    - 96
+
+###################
+## Dataset options
+dataset:
+  pc_dataset_type: "SemanticKITTI"
+  collate_type: "sparse_collate_fn"
+  ignore_label: 0
+  data_config: "./config/labels/semantic-kitti.yaml" # semantic-kitti-sdata.yaml for small data setup
+  num_classes: 20
+  sensor:
+    name: "HDL64"
+    type: "spherical" # projective
+    fov_up: 3
+    fov_down: -25
+
+###################
+## Train params
+train:
+  epochs: 64 # default 128
+  learning_rate: 2.4e-1 # 1.0e-2 for SGD | 1.0e-4 for AdamW
+  weight_decay: 1.0e-4
+  optimizer: SGD # [SGD, AdamW]
+  batch_size: 4   # batch size (default 8)
+  workers: 4     # number of threads to get data
+  epsilon_w: 0.001       # class weight w = 1 / (content + epsilon_w)
+  momentum: 0.9
+  nesterov: True
+
+  mav_loss: True  # features loss
+  cont_loss: True  # contrastive loss
+  obj_loss: True  # objectosphere loss
+
+  anomaly_ratio: [0.0, 0.0] # ratio of anomalous samples in train/val
+
+  scheduler:
+    name: "WarmupCosine"  # [OneCycle, WarmupCosine]
+    #OneCycleLR: # Old decay with warmup and cosine annealing 
+    max_lr: 0.01 # Equal to optimizer.lr
+    total_steps: 1000 # Equal to max_epochs * iterations_per_epoch
+    pct_start: 0.02 # The percentage of the cycle (in number of steps) spent increasing the learning rate (warmup).
+
+  report_epoch: 1 # report every epoch
+  show_scans: False # show scans during training
+
+  #lr_scheduler: CosineAnnealingWarmRestarts  # [StepLR, ReduceLROnPlateau, CosineAnnealingLR, CosineAnnealingWarmRestarts]
+  #momentum: 0.9
+  #nesterov: True
+  #weight_decay: 1.0e-4

lido-stu/checkpoint-best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1937df434810447d132203169e9ce1005c42c505b500a0915681e8271ee5e3f1
+size 174020824

lido-stu/mavs.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba194e2a48b684bf1710fdaaeba36af4ba2df617ac05ce060e5a3ef45ad10b07
+size 7611

lido-stu/vars.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c55e80b956584ed2967d9aeee95ac194efd590c05f864103aca8cac90d78f3c3
+size 7535