Upload t1_spformer/config.py with huggingface_hub

t1_spformer/config.py

@@ -0,0 +1,277 @@
+weight = 'weights/volt-base-scannetpp.pth'
+resume = False
+evaluate = True
+test_only = False
+seed = 20260816
+save_path = 'exp/default'
+num_worker = 16
+batch_size = 2
+gradient_accumulation_steps = 8
+batch_size_val = None
+batch_size_test = None
+epoch = 400
+eval_epoch = 20
+clip_grad = 10.0
+use_ema = True
+ema_decay = 0.999
+dataset_ratios = None
+sync_bn = False
+enable_amp = True
+amp_dtype = 'bfloat16'
+empty_cache = False
+empty_cache_per_epoch = False
+find_unused_parameters = True
+enable_wandb = False
+wandb_project = 'Volt'
+wandb_key = None
+mix_prob = 0.0
+param_dicts = None
+hooks = [
+    dict(type='CheckpointLoader'),
+    dict(type='IterationTimer', warmup_iter=2),
+    dict(type='InformationWriter'),
+    dict(
+        type='InsSegEvaluator',
+        segment_ignore_index=(-1, 0),
+        min_region_size=1),
+    dict(type='CheckpointSaver', save_freq=2)
+]
+train = dict(type='DefaultTrainer')
+test = dict(
+    type='InsSegTester',
+    verbose=True,
+    segment_ignore_index=(-1, 0),
+    instance_ignore_index=-1)
+SUPERPOINT_PARAMS = dict(k_thresh=0.005, seg_min=5)
+num_classes = 3
+segment_ignore_index = (-1, 0)
+semantic_num_classes = 2
+num_channels = 256
+model = dict(
+    type='SPFormer-v1m1',
+    backbone=dict(
+        type='Volt',
+        in_channels=6,
+        embed_dim=768,
+        depth=12,
+        num_heads=12,
+        mlp_ratio=4,
+        init_values=None,
+        qk_norm=True,
+        drop_path=0.3,
+        stride=5,
+        kernel_size=5,
+        increase_drop_path=True,
+        up_mlp_dim=256),
+    decoder=dict(
+        type='SPFormerDecoder',
+        num_class=2,
+        in_channel=256,
+        num_layer=6,
+        num_query=200,
+        d_model=384,
+        nhead=8,
+        hidden_dim=1024,
+        dropout=0.0,
+        activation_fn='gelu',
+        iter_pred=True,
+        attn_mask=True,
+        use_query_pos=False,
+        use_score=True,
+        use_param_query=True),
+    criterion=dict(
+        type='SPFormerCriterion',
+        matcher=dict(
+            type='SPFormerHungarianMatcher',
+            costs=[
+                dict(type='SPFormerQueryClassificationCost', weight=0.5),
+                dict(type='SPFormerMaskBCECost', weight=1.0),
+                dict(type='SPFormerMaskDiceCost', weight=1.0)
+            ]),
+        loss_weight=[0.2, 1.0, 1.0, 0.5],
+        num_classes=2,
+        non_object_weight=0.1,
+        fix_dice_loss_weight=False,
+        iter_matcher=True,
+        fix_mean_loss=True),
+    semantic_num_classes=2,
+    semantic_ignore_index=-1,
+    segment_ignore_index=(-1, 0),
+    instance_ignore_index=-1,
+    topk_insts=200,
+    score_thr=0.0,
+    npoint_thr=10,
+    nms=True)
+optimizer = dict(type='AdamW', lr=0.0003, weight_decay=0.1)
+scheduler = dict(
+    type='OneCycleLR',
+    max_lr=0.0003,
+    pct_start=0.05,
+    anneal_strategy='cos',
+    div_factor=10.0,
+    final_div_factor=1000.0)
+dataset_type = 'ArtiInsSegDataset'
+data_root = 'data/pointcept_mov'
+GRID = 0.02
+data = dict(
+    num_classes=3,
+    ignore_index=-1,
+    names=['background', 'rotation', 'translation'],
+    train=dict(
+        type='ArtiInsSegDataset',
+        split='train',
+        data_root='data/pointcept_mov',
+        motion_targets=False,
+        require_assets=('coord', 'color', 'normal', 'segment', 'instance',
+                        'superpoint'),
+        transform=[
+            dict(type='SphereCrop', point_max=400000, mode='random'),
+            dict(type='CenterShift', apply_z=True),
+            dict(
+                type='RandomDropout',
+                dropout_ratio=0.2,
+                dropout_application_ratio=0.2),
+            dict(
+                type='RandomRotate',
+                angle=[-1, 1],
+                axis='z',
+                center=[0, 0, 0],
+                p=0.95),
+            dict(
+                type='RandomRotate',
+                angle=[-0.015625, 0.015625],
+                axis='x',
+                p=0.95),
+            dict(
+                type='RandomRotate',
+                angle=[-0.015625, 0.015625],
+                axis='y',
+                p=0.95),
+            dict(type='RandomScale', scale=[0.9, 1.1]),
+            dict(type='ChromaticAutoContrast', p=0.2, blend_factor=None),
+            dict(type='ChromaticTranslation', p=0.95, ratio=0.05),
+            dict(type='ChromaticJitter', p=0.95, std=0.05),
+            dict(type='SphereCrop', sample_rate=0.6, mode='random'),
+            dict(
+                type='Copy',
+                keys_dict=dict(
+                    coord='origin_coord',
+                    instance='origin_instance',
+                    segment='origin_segment')),
+            dict(
+                type='Update',
+                keys_dict=dict(index_valid_keys=[
+                    'coord', 'color', 'normal', 'segment', 'instance'
+                ])),
+            dict(
+                type='GridSample',
+                grid_size=0.02,
+                hash_type='fnv',
+                mode='train',
+                return_grid_coord=True,
+                return_inverse=True),
+            dict(type='NormalizeColor'),
+            dict(
+                type='InstanceParser',
+                segment_ignore_index=(-1, 0),
+                instance_ignore_index=-1),
+            dict(type='ToTensor'),
+            dict(
+                type='Collect',
+                keys=('coord', 'origin_coord', 'grid_coord', 'segment',
+                      'origin_segment', 'instance', 'origin_instance',
+                      'superpoint', 'inverse'),
+                feat_keys=('color', 'normal'),
+                offset_keys_dict=dict(
+                    offset='coord', origin_offset='origin_coord'))
+        ],
+        test_mode=False,
+        loop=20),
+    val=dict(
+        type='ArtiInsSegDataset',
+        split='validation',
+        data_root='data/pointcept_mov',
+        motion_targets=False,
+        require_assets=('coord', 'color', 'normal', 'segment', 'instance',
+                        'superpoint'),
+        transform=[
+            dict(type='CenterShift', apply_z=True),
+            dict(
+                type='Copy',
+                keys_dict=dict(
+                    coord='origin_coord',
+                    instance='origin_instance',
+                    segment='origin_segment')),
+            dict(
+                type='Update',
+                keys_dict=dict(index_valid_keys=[
+                    'coord', 'color', 'normal', 'segment', 'instance'
+                ])),
+            dict(
+                type='GridSampleDeterministic',
+                grid_size=0.02,
+                hash_type='fnv',
+                return_grid_coord=True,
+                return_inverse=True),
+            dict(type='NormalizeColor'),
+            dict(
+                type='InstanceParser',
+                segment_ignore_index=(-1, 0),
+                instance_ignore_index=-1),
+            dict(type='ToTensor'),
+            dict(
+                type='Collect',
+                keys=('coord', 'origin_coord', 'grid_coord', 'segment',
+                      'origin_segment', 'instance', 'origin_instance',
+                      'superpoint', 'inverse', 'name'),
+                feat_keys=('color', 'normal'),
+                offset_keys_dict=dict(
+                    offset='coord', origin_offset='origin_coord'))
+        ],
+        test_mode=False),
+    test=dict(
+        type='ArtiInsSegDataset',
+        split='validation',
+        data_root='data/pointcept_mov',
+        motion_targets=False,
+        require_assets=('coord', 'color', 'normal', 'segment', 'instance',
+                        'superpoint'),
+        transform=[
+            dict(type='CenterShift', apply_z=True),
+            dict(
+                type='Copy',
+                keys_dict=dict(
+                    coord='origin_coord',
+                    instance='origin_instance',
+                    segment='origin_segment')),
+            dict(
+                type='Update',
+                keys_dict=dict(index_valid_keys=[
+                    'coord', 'color', 'normal', 'segment', 'instance'
+                ])),
+            dict(
+                type='GridSampleDeterministic',
+                grid_size=0.02,
+                hash_type='fnv',
+                return_grid_coord=True,
+                return_inverse=True),
+            dict(type='NormalizeColor'),
+            dict(
+                type='InstanceParser',
+                segment_ignore_index=(-1, 0),
+                instance_ignore_index=-1),
+            dict(type='ToTensor'),
+            dict(
+                type='Collect',
+                keys=('coord', 'origin_coord', 'grid_coord', 'segment',
+                      'origin_segment', 'instance', 'origin_instance',
+                      'superpoint', 'inverse', 'name'),
+                feat_keys=('color', 'normal'),
+                offset_keys_dict=dict(
+                    offset='coord', origin_offset='origin_coord'))
+        ],
+        test_mode=False))
+num_worker_per_gpu = 16
+batch_size_per_gpu = 2
+batch_size_val_per_gpu = 1
+batch_size_test_per_gpu = 1