Buddhi19 commited on
Commit
3db00d5
·
1 Parent(s): ba3de5c

Release MambaCD checkpoints and metrics

Browse files
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ .nfs*
README.md CHANGED
@@ -1,3 +1,85 @@
1
  ---
2
  license: mit
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: mit
3
+ library_name: pytorch
4
+ pipeline_tag: image-segmentation
5
+ tags:
6
+ - remote-sensing
7
+ - change-detection
8
+ - mamba
9
+ - vmamba
10
  ---
11
+
12
+ # ORBIT-Mamba / MambaCD checkpoints
13
+
14
+ Validation-selected MambaCD checkpoints for binary, bi-temporal remote-sensing change detection. Each release contains the checkpoint, the exact saved training configuration, the original test metrics, and a normalized `metrics.json`.
15
+
16
+ All accuracy metrics are fractions in `[0, 1]`. The F1 score is for the changed class. `mIoU` is the mean of changed and unchanged class IoU. GPU memory is the peak reserved **training** memory recorded in the source run, not inference-only memory.
17
+
18
+ ## Released models
19
+
20
+ | Dataset | Backbone | F1 | mIoU | Overall accuracy | Recall | Precision | BF1 | BmIoU | GFLOPs | GPU GB | FPS | Parameters | Download |
21
+ | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- |
22
+ | LEVIR-CD+ | VMamba-base | 0.8728 | 0.8817 | 0.9895 | 0.8815 | 0.8642 | 0.8321 | — | — | 12.23 | — | 96.0M | [model.pth](https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/levir-cd-plus/model.pth?download=true) |
23
+ | WHU-CD | VMamba-base | 0.9617 | 0.9612 | 0.9963 | 0.9573 | 0.9662 | 0.9321 | — | — | 12.23 | — | 96.0M | [model.pth](https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/whu-cd/model.pth?download=true) |
24
+ | WildFireS2 | VMamba-tiny | 0.8645 | 0.6477 | 0.8126 | 0.9080 | 0.8250 | — | — | — | 8.68 | — | 44.0M | [model.pth](https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/wildfires2/model.pth?download=true) |
25
+ | SYSU-CD | VMamba-base | 0.8457 | 0.8213 | 0.9279 | 0.8382 | 0.8533 | 0.4600 | — | — | 12.23 | — | 96.0M | [model.pth](https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/sysu-cd/model.pth?download=true) |
26
+ | DSIFN-CD | VMamba-small | 0.7127 | 0.7223 | 0.9039 | 0.7012 | 0.7245 | 0.3705 | — | — | 10.47 | — | 57.2M | [model.pth](https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/dsifn-cd/model.pth?download=true) |
27
+ | KATE-CD-256 | VMamba-small | 0.6612 | 0.7376 | 0.9816 | 0.6357 | 0.6889 | 0.6411 | — | — | 10.61 | — | 57.9M | [model.pth](https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/kate-cd-256/model.pth?download=true) |
28
+
29
+ `—` means the value was not measured in the source run. BmIoU was not implemented by the evaluator. GFLOPs and FPS were requested in the saved configs but were not emitted by the runs, and this release does not estimate them. The WildFireS2 run predates the corrected symmetric boundary evaluator, so its recorded BF1 value of `0.0` is retained in `source_test_metrics.json` but treated as unavailable in the normalized table.
30
+
31
+ ## Download
32
+
33
+ Download one checkpoint with Python:
34
+
35
+ ```python
36
+ from huggingface_hub import hf_hub_download
37
+
38
+ checkpoint = hf_hub_download(
39
+ repo_id="dineth18/ORBIT-Mamba",
40
+ filename="models/whu-cd/model.pth",
41
+ )
42
+ ```
43
+
44
+ Or clone every checkpoint (git-xet is recommended for the large files):
45
+
46
+ ```bash
47
+ git xet install
48
+ git clone git@hf.co:dineth18/ORBIT-Mamba
49
+ ```
50
+
51
+ ## Repository layout
52
+
53
+ ```text
54
+ models/<dataset>/
55
+ model.pth # validation-selected checkpoint
56
+ config.yaml # exact configuration saved by the run
57
+ resolved_config.yaml # included when the run emitted one
58
+ source_test_metrics.json # unmodified evaluator output
59
+ metrics.json # normalized release metrics and provenance
60
+ ```
61
+
62
+ See [`model-index.json`](model-index.json) for a machine-readable index and direct download URLs.
63
+
64
+ ## Loading
65
+
66
+ Use the matching MambaCD source code and the included configuration. Checkpoints are PyTorch training checkpoints rather than standalone TorchScript or ONNX exports. Local dataset, cache, and pretrained-weight paths in the saved configs document the original run and must be changed for a new machine.
67
+
68
+ ```python
69
+ import torch
70
+
71
+ state = torch.load("models/whu-cd/model.pth", map_location="cpu", weights_only=False)
72
+ print(state.keys())
73
+ ```
74
+
75
+ ## Metric provenance
76
+
77
+ - F1, mIoU, overall accuracy, recall, precision, and BF1 come directly from each run's `metrics/test_metrics.json`.
78
+ - Thresholds were selected on validation data and then applied to the test split.
79
+ - Parameter totals come from the model summary written during training.
80
+ - GPU GB is the maximum `gpu_memory` value in the run's `metrics/train_history.json` and represents peak reserved training memory.
81
+ - BmIoU, GFLOPs, and FPS are explicitly null when no reproducible measurement exists.
82
+
83
+ ## License
84
+
85
+ This Hugging Face repository is marked MIT. The source MambaCD repository did not contain a root license file in the inspected checkout, so users must also review the MambaCD, VMamba, pretrained-weight, and dataset terms before redistribution or commercial use.
model-index.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "repository": "dineth18/ORBIT-Mamba",
4
+ "models": [
5
+ {"dataset": "LEVIR-CD+", "path": "models/levir-cd-plus", "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/levir-cd-plus/model.pth?download=true"},
6
+ {"dataset": "WHU-CD", "path": "models/whu-cd", "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/whu-cd/model.pth?download=true"},
7
+ {"dataset": "WildFireS2", "path": "models/wildfires2", "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/wildfires2/model.pth?download=true"},
8
+ {"dataset": "SYSU-CD", "path": "models/sysu-cd", "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/sysu-cd/model.pth?download=true"},
9
+ {"dataset": "DSIFN-CD", "path": "models/dsifn-cd", "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/dsifn-cd/model.pth?download=true"},
10
+ {"dataset": "KATE-CD-256", "path": "models/kate-cd-256", "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/kate-cd-256/model.pth?download=true"}
11
+ ]
12
+ }
models/dsifn-cd/config.yaml ADDED
@@ -0,0 +1,273 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ project:
2
+ name: mamba-cd
3
+ seed: 42
4
+ storage_root: /new-home/buddhiw/mamba-cd
5
+ output_root: /new-home/buddhiw/mamba-cd/outputs
6
+ run_name: auto
7
+ notes: ""
8
+
9
+ cache:
10
+ root: /new-home/buddhiw/mamba-cd/cache
11
+
12
+ # Change only this line to switch datasets.
13
+ dataset_config: configs/datasets/dsifn.yaml
14
+
15
+ # ─── SOTA Training Configuration (Task 2) ────────────────────────────────────
16
+ sota_training:
17
+ enabled: true
18
+ registry: configs/sota_models/sota_model_registry.yaml
19
+ output_root: /new-home/buddhiw/mamba-cd/sota_runs
20
+ datasets:
21
+ - configs/datasets/levir.yaml
22
+ - configs/datasets/whu.yaml
23
+ - configs/datasets/sysu.yaml
24
+ - configs/datasets/dsifn.yaml
25
+ - configs/datasets/kate.yaml
26
+ - configs/datasets/wildfire.yaml
27
+ models_to_train:
28
+ - auto
29
+ train_policy:
30
+ use_existing_script_if_available: true
31
+ prefer_wrapper_if_existing_script_incompatible: true
32
+ write_adapter_if_missing: true
33
+ no_command_line_arguments: true
34
+ validation_best_only: true
35
+ run_final_test: true
36
+ run_test_sweep: true
37
+ save_predictions: false
38
+ common_training:
39
+ iterations: 50000
40
+ lr: 0.0001
41
+ weight_decay: 0.01
42
+ optimizer: adamw
43
+ scheduler: cosine
44
+ warmup_iters: 1000
45
+ batch_size: null
46
+ amp: true
47
+ grad_clip: 1.0
48
+ log_every: 50
49
+ val_every: 5000
50
+ eval:
51
+ threshold: 0.5
52
+ sweep_val_threshold: true
53
+ threshold_min: 0.05
54
+ threshold_max: 0.95
55
+ threshold_step: 0.05
56
+ use_val_threshold_for_test: true
57
+ also_report_test_sweep: true
58
+ boundary_tolerance: 2
59
+ execution:
60
+ skip_completed: true
61
+ continue_on_error: true
62
+ max_runs: null
63
+ dry_run: false
64
+ debug_fast: false
65
+ debug_iterations: 200
66
+ debug_val_every: 100
67
+ debug_max_models: 2
68
+ debug_max_datasets: 1
69
+
70
+ # ─── Global Defaults (overridden by dataset YAML best_config) ────────────────
71
+ defaults:
72
+ model:
73
+ name: MambaCD
74
+ encoder: vmamba
75
+ encoder_variant: small
76
+ pretrained: true
77
+ freeze_encoder: false
78
+ decoder_channels: 192
79
+ temporal_input:
80
+ region: signed_abs
81
+ boundary: signed_abs
82
+ use_drbi: true
83
+ use_decoder: true
84
+ use_boundary_refinement: true
85
+ drbi:
86
+ type: attentive
87
+ use_depthwise: true
88
+ gate_hidden_ratio: 0.25
89
+ region_gate_min: 0.2
90
+ region_gate_max: 0.8
91
+ boundary_gate_min: 0.0
92
+ boundary_gate_max: 0.4
93
+ share_region_boundary_projection: true
94
+ share_region_boundary_spatial: true
95
+ attentive:
96
+ sources: [abs, signed, pair]
97
+ fusion: softmax
98
+ use_channel_attention: false
99
+ use_spatial_attention: true
100
+ region_gate_type: residual
101
+ region_gate_alpha: 0.5
102
+ boundary_gate_type: hybrid_sobel_learned
103
+ boundary_sobel_init_weight: 0.7
104
+ boundary_gate_max: 0.5
105
+ decoder:
106
+ type: arf_fpn
107
+ use_dilations: true
108
+ dilation_rates: [1, 2, 3]
109
+ fusion: concat
110
+ arf_stages: [2, 3, 4]
111
+ deep_supervision: true
112
+ boundary_refinement:
113
+ type: default
114
+ enabled: true
115
+ residual_scale: 0.05
116
+ use_depthwise: true
117
+ zero_init_residual: true
118
+ input_stages: [1, 2]
119
+ stage_indexing: one_based
120
+ empty_stage_policy: error
121
+ training:
122
+ iterations: 50000
123
+ lr: 0.0001
124
+ weight_decay: 0.01
125
+ optimizer: adamw
126
+ scheduler: cosine
127
+ warmup_iters: 1000
128
+ amp: true
129
+ grad_clip: 1.0
130
+ log_every: 50
131
+ val_every: 5000
132
+ save_best_only: true
133
+ best_metric: F1
134
+ higher_is_better: true
135
+ loss:
136
+ bce: 1.0
137
+ dice: 0.5
138
+ aux: 0.1
139
+ boundary: 0.05
140
+ residual_reg: 0.01
141
+ use_pos_weight: false
142
+ pos_weight: auto
143
+ auto_pos_weight: true
144
+ pos_weight_cache: true
145
+ max_pos_weight: 3.0
146
+ boundary_width: 2
147
+ boundary_dice: 0.05
148
+ deep_supervision: 0.10
149
+ deep_weights: [0.4, 0.3, 0.2, 0.1]
150
+ eval:
151
+ threshold: 0.5
152
+ sweep_val_threshold: true
153
+ threshold_min: 0.05
154
+ threshold_max: 0.95
155
+ threshold_step: 0.05
156
+ use_val_threshold_for_test: true
157
+ also_report_test_sweep: true
158
+ save_predictions: true
159
+ test_split: test
160
+ allow_val_as_test: false
161
+ boundary_tolerance: 2
162
+ checkpoint_eval:
163
+ enabled: true
164
+ run_dir: latest
165
+ checkpoint_dir: null
166
+ test_split: test
167
+ checkpoint_glob: "*.pth"
168
+ sort_by_iteration: true
169
+ threshold_mode: both
170
+ threshold_min: 0.05
171
+ threshold_max: 0.95
172
+ threshold_step: 0.05
173
+ amp: true
174
+ save_predictions: false
175
+ save_probability_maps: false
176
+ batch_size_override: null
177
+ skip_completed: true
178
+ continue_on_error: true
179
+ output_subdir: checkpoint_test_evaluation
180
+ diagnostic_only: true
181
+ max_checkpoints: null
182
+ ablation_runner:
183
+ official_selection_metric: val_F1
184
+ official_selection_split: val
185
+ prohibit_test_based_selection: true
186
+
187
+ pretrained:
188
+ dir: /new-home/buddhiw/mamba-cd/pretrained_weights
189
+ auto_download: true
190
+ strict_load: false
191
+ allow_random_init_if_failed: false
192
+ vmamba:
193
+ tiny:
194
+ url: "https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm1_tiny_0230s_ckpt_epoch_264.pth"
195
+ filename: "vssm1_tiny_0230s_ckpt_epoch_264.pth"
196
+ local_path: null
197
+ small:
198
+ url: "https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_small_0229_ckpt_epoch_222.pth"
199
+ filename: "vssm_small_0229_ckpt_epoch_222.pth"
200
+ local_path: null
201
+ base:
202
+ url: "https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_base_0229_ckpt_epoch_237.pth"
203
+ filename: "vssm_base_0229_ckpt_epoch_237.pth"
204
+ local_path: null
205
+ fractalmamba:
206
+ tiny:
207
+ url: ""
208
+ filename: ""
209
+ local_path: null
210
+ small:
211
+ url: ""
212
+ filename: ""
213
+ local_path: null
214
+ base:
215
+ url: ""
216
+ filename: ""
217
+ local_path: null
218
+
219
+ vmamba:
220
+ repo_url: "https://github.com/MzeroMiko/VMamba.git"
221
+ local_dir: third_party/VMamba
222
+ auto_setup: true
223
+ verify_selective_scan: true
224
+
225
+ fractalmamba:
226
+ repo_url: "https://github.com/hkxiao/Fractal-Mamba"
227
+ local_dir: /new-home/buddhiw/mamba-cd/third_party/FractalMambaPlusPlus
228
+ auto_setup: true
229
+ auto_discover_repo: true
230
+ verify_fractal_scan: true
231
+ verify_hilbert_serialization: true
232
+ verify_selective_scan: true
233
+
234
+ hardware:
235
+ device: cuda:0
236
+ gpu_ids: [0]
237
+ mixed_precision: true
238
+
239
+ checkpoint:
240
+ resume: false
241
+ resume_path: null
242
+ test_checkpoint: null
243
+ auto_find_best: true
244
+ save_optimizer: true
245
+ save_scheduler: true
246
+
247
+ logging:
248
+ tensorboard: true
249
+ save_train_log: true
250
+ save_config: true
251
+ save_model_summary: true
252
+ print_params: true
253
+ print_flops: true
254
+ print_fps: true
255
+ print_peak_memory: true
256
+ save_debug_predictions: true
257
+ debug_samples: 8
258
+
259
+ outputs:
260
+ save_tensorboard: false
261
+ save_logs: true
262
+ save_qualitative: true
263
+
264
+ inference:
265
+ input_a: inference/A
266
+ input_b: inference/B
267
+ output: /new-home/buddhiw/mamba-cd/inference/predictions
268
+
269
+ debug:
270
+ check_shapes: true
271
+ overfit_test: false
272
+ overfit_samples: 4
273
+ overfit_iterations: 100
models/dsifn-cd/metrics.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "model": "MambaCD",
4
+ "dataset": "DSIFN-CD",
5
+ "backbone": "VMamba-small",
6
+ "input_size": 256,
7
+ "test_threshold": 0.55,
8
+ "metrics": {"f1": 0.7126559007917083, "miou": 0.7222683885763073, "overall_accuracy": 0.9039402008053767, "recall": 0.7012129627190776, "precision": 0.7244795372275411, "bf1": 0.3705149902215689, "bmiou": null},
9
+ "complexity": {"gflops": null, "gpu_gb": 10.466885632, "gpu_memory_kind": "peak_reserved_training", "fps": null, "parameters_m": 57.2},
10
+ "source_run": "run_20260628_184240_DSIFN-CD_vmamba_baseline",
11
+ "source_checkpoint": "best_iter_060000_F1_0.9330.pth",
12
+ "configuration": "config.yaml",
13
+ "resolved_configuration": "resolved_config.yaml",
14
+ "source_metrics": "source_test_metrics.json",
15
+ "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/dsifn-cd/model.pth?download=true"
16
+ }
models/dsifn-cd/model.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4c45a01e9347b291fed7e24e290eaa21b269867a98c846b0d4daca07ffd230f9
3
+ size 679897412
models/dsifn-cd/resolved_config.yaml ADDED
@@ -0,0 +1,374 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ name: MambaCD
3
+ encoder: vmamba
4
+ encoder_variant: small
5
+ pretrained: true
6
+ freeze_encoder: false
7
+ decoder_channels: 192
8
+ temporal_input:
9
+ region: signed_abs
10
+ boundary: signed_abs
11
+ use_drbi: true
12
+ use_decoder: true
13
+ use_boundary_refinement: true
14
+ drbi:
15
+ type: attentive
16
+ use_depthwise: true
17
+ gate_hidden_ratio: 0.25
18
+ region_gate_min: 0.2
19
+ region_gate_max: 0.8
20
+ boundary_gate_min: 0.0
21
+ boundary_gate_max: 0.4
22
+ share_region_boundary_projection: true
23
+ share_region_boundary_spatial: true
24
+ attentive:
25
+ sources:
26
+ - abs
27
+ - signed
28
+ - pair
29
+ fusion: softmax
30
+ use_channel_attention: false
31
+ use_spatial_attention: true
32
+ region_gate_type: residual
33
+ region_gate_alpha: 0.5
34
+ boundary_gate_type: hybrid_sobel_learned
35
+ boundary_sobel_init_weight: 0.7
36
+ boundary_gate_max: 0.5
37
+ decoder:
38
+ type: weighted_arf_fpn
39
+ use_dilations: true
40
+ dilation_rates:
41
+ - 1
42
+ - 2
43
+ - 3
44
+ fusion: concat
45
+ arf_stages:
46
+ - 2
47
+ - 3
48
+ - 4
49
+ deep_supervision: true
50
+ boundary_refinement:
51
+ type: localized_attentive
52
+ enabled: true
53
+ residual_scale: 0.05
54
+ use_depthwise: true
55
+ zero_init_residual: true
56
+ input_stages:
57
+ - 1
58
+ - 2
59
+ stage_indexing: one_based
60
+ empty_stage_policy: error
61
+ use_stage_attention: true
62
+ use_uncertainty: true
63
+ localize_residual: true
64
+ residual_mask_source: edge_boundary
65
+ boundary_width: 2
66
+ training:
67
+ iterations: 60000
68
+ lr: 8.0e-05
69
+ weight_decay: 0.03
70
+ optimizer: adamw
71
+ scheduler: cosine
72
+ warmup_iters: 1500
73
+ amp: true
74
+ grad_clip: 1.0
75
+ log_every: 50
76
+ val_every: 5000
77
+ save_best_only: true
78
+ best_metric: F1
79
+ higher_is_better: true
80
+ device: cuda:1
81
+ loss:
82
+ bce: 1.0
83
+ dice: 0.5
84
+ aux: 0.1
85
+ boundary: 0.05
86
+ residual_reg: 0.01
87
+ use_pos_weight: false
88
+ pos_weight: auto
89
+ auto_pos_weight: true
90
+ pos_weight_cache: true
91
+ max_pos_weight: 3.0
92
+ boundary_width: 2
93
+ boundary_dice: 0.05
94
+ deep_supervision: 0.1
95
+ deep_weights:
96
+ - 0.4
97
+ - 0.3
98
+ - 0.2
99
+ - 0.1
100
+ eval:
101
+ threshold: 0.5
102
+ sweep_val_threshold: true
103
+ threshold_min: 0.05
104
+ threshold_max: 0.95
105
+ threshold_step: 0.05
106
+ use_val_threshold_for_test: true
107
+ also_report_test_sweep: true
108
+ save_predictions: true
109
+ test_split: test
110
+ allow_val_as_test: false
111
+ boundary_tolerance: 2
112
+ checkpoint_eval:
113
+ enabled: true
114
+ run_dir: latest
115
+ checkpoint_dir: null
116
+ test_split: test
117
+ checkpoint_glob: '*.pth'
118
+ sort_by_iteration: true
119
+ threshold_mode: both
120
+ threshold_min: 0.05
121
+ threshold_max: 0.95
122
+ threshold_step: 0.05
123
+ amp: true
124
+ save_predictions: false
125
+ save_probability_maps: false
126
+ batch_size_override: null
127
+ skip_completed: true
128
+ continue_on_error: true
129
+ output_subdir: checkpoint_test_evaluation
130
+ diagnostic_only: true
131
+ max_checkpoints: null
132
+ ablation_runner:
133
+ official_selection_metric: val_F1
134
+ official_selection_split: val
135
+ prohibit_test_based_selection: true
136
+ enabled: true
137
+ output_dir: /new-home/buddhiw/mamba-cd/ablations_dsifn
138
+ continue_on_error: false
139
+ restore_config_after_each_run: true
140
+ run_train: true
141
+ run_final_test_after_train: true
142
+ test_at_each_val: false
143
+ test_at_each_val_threshold_mode: val_selected
144
+ save_test_at_each_val_predictions: false
145
+ save_val_predictions: false
146
+ save_final_test_predictions: true
147
+ max_variants: null
148
+ start_from_variant: null
149
+ skip_completed: true
150
+ require_clean_dataset_verification: true
151
+ save_intermediate_checkpoints: false
152
+ save_best_only: true
153
+ dry_run: false
154
+ dry_run_print_only: false
155
+ debug_fast: false
156
+ debug_iterations: 200
157
+ debug_val_every: 100
158
+ debug_max_variants: 2
159
+ early_stop:
160
+ enabled: false
161
+ patience_validations: 5
162
+ metric: val_F1
163
+ mode: max
164
+ overfit_diagnostics:
165
+ enabled: true
166
+ track_train_metrics: true
167
+ track_val_metrics: true
168
+ track_test_metrics: true
169
+ compute_generalization_gap: true
170
+ gap_metric: F1
171
+ warning_gap: 0.05
172
+ severe_gap: 0.08
173
+ dataset:
174
+ name: DSIFN-CD
175
+ root: /new-home/buddhiw/mamba-cd/datasets/DSIFN-CD/DSIFN
176
+ raw_root: /new-home/buddhiw/mamba-cd/datasets/DSIFN-CD/DSIFN
177
+ image_size: 256
178
+ batch_size: 8
179
+ num_workers: 8
180
+ binary_threshold: 127
181
+ prepare:
182
+ enabled: false
183
+ source_root: /new-home/buddhiw/mamba-cd/datasets/DSIFN-CD/DSIFN
184
+ target_root: /new-home/buddhiw/mamba-cd/datasets/DSIFN-CD/DSIFN
185
+ mode: existing_folders
186
+ existing_folders:
187
+ train_a: train/t1
188
+ train_b: train/t2
189
+ train_mask: train/mask
190
+ val_a: val/t1
191
+ val_b: val/t2
192
+ val_mask: val/mask
193
+ test_a: test/t1
194
+ test_b: test/t2
195
+ test_mask: test/mask
196
+ split:
197
+ protocol: existing_folders
198
+ source_folders:
199
+ train_a: train/t1
200
+ train_b: train/t2
201
+ train_mask: train/mask
202
+ val_a: val/t1
203
+ val_b: val/t2
204
+ val_mask: val/mask
205
+ test_a: test/t1
206
+ test_b: test/t2
207
+ test_mask: test/mask
208
+ report_dir: /new-home/buddhiw/mamba-cd/dataset_checks
209
+ allow_missing_test: false
210
+ check_filename_overlap: false
211
+ check_hash_overlap: true
212
+ overwrite_splits: true
213
+ verification:
214
+ check_filename_overlap: false
215
+ best_config:
216
+ source_run: outputs/run_20260625_114552_DSIFN-CD_vmamba_baseline
217
+ source_config: outputs/run_20260625_114552_DSIFN-CD_vmamba_baseline/config.yaml
218
+ source_selection: validation_only
219
+ selected_metric:
220
+ val_f1: 0.9312728329993397
221
+ val_iou: 0.8713858854146804
222
+ val_boundary_f1: 0.739704656628777
223
+ official_test_f1: 0.683218099697265
224
+ official_test_iou: 0.5188549894787479
225
+ official_test_boundary_f1: 0.3452506301415701
226
+ notes: Enriched via migration script on 2026-06-27 14:47:10
227
+ project:
228
+ name: mamba-cd
229
+ seed: 42
230
+ storage_root: /new-home/buddhiw/mamba-cd
231
+ output_root: /new-home/buddhiw/mamba-cd/outputs
232
+ run_name: auto
233
+ notes: ''
234
+ cache:
235
+ root: /new-home/buddhiw/mamba-cd/cache
236
+ dataset_config: configs/datasets/dsifn.yaml
237
+ sota_training:
238
+ enabled: true
239
+ registry: configs/sota_models/sota_model_registry.yaml
240
+ output_root: /new-home/buddhiw/mamba-cd/sota_runs
241
+ datasets:
242
+ - configs/datasets/levir.yaml
243
+ - configs/datasets/whu.yaml
244
+ - configs/datasets/sysu.yaml
245
+ - configs/datasets/dsifn.yaml
246
+ - configs/datasets/kate.yaml
247
+ - configs/datasets/wildfire.yaml
248
+ models_to_train:
249
+ - auto
250
+ train_policy:
251
+ use_existing_script_if_available: true
252
+ prefer_wrapper_if_existing_script_incompatible: true
253
+ write_adapter_if_missing: true
254
+ no_command_line_arguments: true
255
+ validation_best_only: true
256
+ run_final_test: true
257
+ run_test_sweep: true
258
+ save_predictions: false
259
+ common_training:
260
+ iterations: 50000
261
+ lr: 0.0001
262
+ weight_decay: 0.01
263
+ optimizer: adamw
264
+ scheduler: cosine
265
+ warmup_iters: 1000
266
+ batch_size: null
267
+ amp: true
268
+ grad_clip: 1.0
269
+ log_every: 50
270
+ val_every: 5000
271
+ eval:
272
+ threshold: 0.5
273
+ sweep_val_threshold: true
274
+ threshold_min: 0.05
275
+ threshold_max: 0.95
276
+ threshold_step: 0.05
277
+ use_val_threshold_for_test: true
278
+ also_report_test_sweep: true
279
+ boundary_tolerance: 2
280
+ execution:
281
+ skip_completed: true
282
+ continue_on_error: true
283
+ max_runs: null
284
+ dry_run: false
285
+ debug_fast: false
286
+ debug_iterations: 200
287
+ debug_val_every: 100
288
+ debug_max_models: 2
289
+ debug_max_datasets: 1
290
+ pretrained:
291
+ dir: /new-home/buddhiw/mamba-cd/pretrained_weights
292
+ auto_download: true
293
+ strict_load: false
294
+ allow_random_init_if_failed: false
295
+ vmamba:
296
+ tiny:
297
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm1_tiny_0230s_ckpt_epoch_264.pth
298
+ filename: vssm1_tiny_0230s_ckpt_epoch_264.pth
299
+ local_path: null
300
+ small:
301
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_small_0229_ckpt_epoch_222.pth
302
+ filename: vssm_small_0229_ckpt_epoch_222.pth
303
+ local_path: null
304
+ base:
305
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_base_0229_ckpt_epoch_237.pth
306
+ filename: vssm_base_0229_ckpt_epoch_237.pth
307
+ local_path: null
308
+ fractalmamba:
309
+ tiny:
310
+ url: ''
311
+ filename: ''
312
+ local_path: null
313
+ small:
314
+ url: ''
315
+ filename: ''
316
+ local_path: null
317
+ base:
318
+ url: ''
319
+ filename: ''
320
+ local_path: null
321
+ vmamba:
322
+ repo_url: https://github.com/MzeroMiko/VMamba.git
323
+ local_dir: third_party/VMamba
324
+ auto_setup: true
325
+ verify_selective_scan: true
326
+ fractalmamba:
327
+ repo_url: https://github.com/hkxiao/Fractal-Mamba
328
+ local_dir: /new-home/buddhiw/mamba-cd/third_party/FractalMambaPlusPlus
329
+ auto_setup: true
330
+ auto_discover_repo: true
331
+ verify_fractal_scan: true
332
+ verify_hilbert_serialization: true
333
+ verify_selective_scan: true
334
+ hardware:
335
+ device: cuda:0
336
+ gpu_ids:
337
+ - 0
338
+ mixed_precision: true
339
+ checkpoint:
340
+ resume: false
341
+ resume_path: null
342
+ test_checkpoint: null
343
+ auto_find_best: true
344
+ save_optimizer: true
345
+ save_scheduler: true
346
+ logging:
347
+ tensorboard: true
348
+ save_train_log: true
349
+ save_config: true
350
+ save_model_summary: true
351
+ print_params: true
352
+ print_flops: true
353
+ print_fps: true
354
+ print_peak_memory: true
355
+ save_debug_predictions: true
356
+ debug_samples: 8
357
+ outputs:
358
+ save_tensorboard: false
359
+ save_logs: true
360
+ save_qualitative: true
361
+ inference:
362
+ input_a: inference/A
363
+ input_b: inference/B
364
+ output: /new-home/buddhiw/mamba-cd/inference/predictions
365
+ debug:
366
+ check_shapes: true
367
+ overfit_test: false
368
+ overfit_samples: 4
369
+ overfit_iterations: 100
370
+ config:
371
+ resolved_from:
372
+ main_config: /new-home/buddhiw/mamba-cd/config.yaml
373
+ dataset_config: configs/datasets/dsifn.yaml
374
+ timestamp: '2026-06-28T18:42:40.475867'
models/dsifn-cd/source_test_metrics.json ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "threshold": 0.55,
3
+ "best_val_threshold": 0.55,
4
+ "test_threshold_used": 0.55,
5
+ "selection_basis": "validation_F1_only",
6
+ "temporal_input": {
7
+ "region": "signed_abs",
8
+ "boundary": "signed_abs"
9
+ },
10
+ "metrics": {
11
+ "f1": 0.7126559007917083,
12
+ "precision": 0.7244795372275411,
13
+ "recall": 0.7012129627190776,
14
+ "iou": 0.5535867821330804,
15
+ "f1_0": 0.9423300717986293,
16
+ "iou_0": 0.8909499950195342,
17
+ "oa": 0.9039402008053767,
18
+ "mf1": 0.8274929862951688,
19
+ "miou": 0.7222683885763073,
20
+ "boundary_f1": 0.3705149902215689,
21
+ "bf1": 0.3705149902215689,
22
+ "pred_positive_ratio": 0.16442362467447916,
23
+ "gt_positive_ratio": 0.16987927754720053,
24
+ "pred_ratio": 0.16442362467447916,
25
+ "gt_ratio": 0.16987927754720053,
26
+ "mean_sigmoid_prob": 0.17070799134671688,
27
+ "mean_prob": 0.17070799134671688,
28
+ "logit_mean": -8.027131716410318,
29
+ "logit_std": 8.014728148162646,
30
+ "prob_min": 1.771795488103578e-10,
31
+ "prob_max": 1.0,
32
+ "tp": 374724.0,
33
+ "fp": 142508.0,
34
+ "fn": 159670.0,
35
+ "tn": 2468826.0,
36
+ "f1_inverted_mask": 0.09110088665599451
37
+ },
38
+ "checkpoint": "/new-home/buddhiw/mamba-cd/outputs/run_20260628_184240_DSIFN-CD_vmamba_baseline/checkpoints/best_iter_060000_F1_0.9330.pth"
39
+ }
models/kate-cd-256/config.yaml ADDED
@@ -0,0 +1,298 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ dataset:
2
+ name: KATE-CD-256
3
+ root: /new-home/buddhiw/Datasets/KATE-CD-256
4
+ raw_root: /new-home/buddhiw/Datasets/KATE-CD-256
5
+ image_size: 256
6
+ batch_size: 8
7
+ num_workers: 8
8
+ binary_threshold: 127
9
+ prepare:
10
+ enabled: false
11
+ source_root: /new-home/buddhiw/Datasets/KATE-CD-256
12
+ target_root: /new-home/buddhiw/Datasets/KATE-CD-256
13
+ mode: existing_folders
14
+ existing_folders:
15
+ train_a: train/A
16
+ train_b: train/B
17
+ train_mask: train/label
18
+ val_a: val/A
19
+ val_b: val/B
20
+ val_mask: val/label
21
+ test_a: test/A
22
+ test_b: test/B
23
+ test_mask: test/label
24
+ split:
25
+ protocol: existing_folders
26
+ source_folders:
27
+ train_a: train/A
28
+ train_b: train/B
29
+ train_mask: train/label
30
+ val_a: val/A
31
+ val_b: val/B
32
+ val_mask: val/label
33
+ test_a: test/A
34
+ test_b: test/B
35
+ test_mask: test/label
36
+ report_dir: /new-home/buddhiw/mamba-cd/dataset_checks
37
+ allow_missing_test: false
38
+ check_filename_overlap: false
39
+ check_hash_overlap: true
40
+ overwrite_splits: true
41
+ binarize_masks: true
42
+ binary_threshold: 127
43
+ verification:
44
+ check_filename_overlap: false
45
+ check_hash_overlap: true
46
+ check_mask_binary: true
47
+ check_size_match: true
48
+ check_empty_masks: true
49
+ allow_empty_masks: true
50
+ allow_all_positive_masks: true
51
+ project:
52
+ name: mamba-cd
53
+ seed: 42
54
+ storage_root: /new-home/buddhiw/mamba-cd
55
+ output_root: /new-home/buddhiw/mamba-cd/outputs
56
+ run_name: auto
57
+ notes: ''
58
+ cache:
59
+ root: /new-home/buddhiw/mamba-cd/cache
60
+ dataset_config: configs/datasets/kate.yaml
61
+ model:
62
+ name: MambaCD
63
+ encoder: vmamba
64
+ encoder_variant: small
65
+ pretrained: true
66
+ freeze_encoder: false
67
+ decoder_channels: 192
68
+ temporal_input:
69
+ region: features_abs
70
+ boundary: features_abs
71
+ use_drbi: true
72
+ use_decoder: true
73
+ use_boundary_refinement: true
74
+ drbi:
75
+ type: attentive
76
+ use_depthwise: true
77
+ gate_hidden_ratio: 0.25
78
+ region_gate_min: 0.2
79
+ region_gate_max: 0.8
80
+ boundary_gate_min: 0.0
81
+ boundary_gate_max: 0.4
82
+ share_region_boundary_projection: true
83
+ share_region_boundary_spatial: true
84
+ attentive:
85
+ sources:
86
+ - abs
87
+ - signed
88
+ - pair
89
+ fusion: softmax
90
+ use_channel_attention: true
91
+ use_spatial_attention: true
92
+ region_gate_type: residual
93
+ region_gate_alpha: 0.5
94
+ boundary_gate_type: hybrid_sobel_learned
95
+ boundary_sobel_init_weight: 0.7
96
+ boundary_gate_max: 0.6
97
+ decoder:
98
+ type: weighted_arf_fpn
99
+ use_dilations: true
100
+ dilation_rates:
101
+ - 1
102
+ - 2
103
+ - 3
104
+ fusion: concat
105
+ arf_stages:
106
+ - 1
107
+ - 2
108
+ - 3
109
+ - 4
110
+ deep_supervision: false
111
+ boundary_refinement:
112
+ type: localized_attentive
113
+ enabled: true
114
+ residual_scale: 0.1
115
+ use_depthwise: true
116
+ zero_init_residual: true
117
+ input_stages:
118
+ - 1
119
+ - 2
120
+ stage_indexing: one_based
121
+ empty_stage_policy: error
122
+ use_stage_attention: true
123
+ use_uncertainty: true
124
+ localize_residual: true
125
+ residual_mask_source: edge_boundary
126
+ boundary_width: 2
127
+ pretrained:
128
+ dir: /new-home/buddhiw/mamba-cd/pretrained_weights
129
+ auto_download: true
130
+ strict_load: false
131
+ allow_random_init_if_failed: false
132
+ vmamba:
133
+ tiny:
134
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm1_tiny_0230s_ckpt_epoch_264.pth
135
+ filename: vssm1_tiny_0230s_ckpt_epoch_264.pth
136
+ local_path: null
137
+ small:
138
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_small_0229_ckpt_epoch_222.pth
139
+ filename: vssm_small_0229_ckpt_epoch_222.pth
140
+ local_path: null
141
+ base:
142
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_base_0229_ckpt_epoch_237.pth
143
+ filename: vssm_base_0229_ckpt_epoch_237.pth
144
+ local_path: null
145
+ fractalmamba:
146
+ tiny:
147
+ url: ''
148
+ filename: ''
149
+ local_path: null
150
+ small:
151
+ url: ''
152
+ filename: ''
153
+ local_path: null
154
+ base:
155
+ url: ''
156
+ filename: ''
157
+ local_path: null
158
+ vmamba:
159
+ repo_url: https://github.com/MzeroMiko/VMamba.git
160
+ local_dir: third_party/VMamba
161
+ auto_setup: true
162
+ verify_selective_scan: true
163
+ fractalmamba:
164
+ repo_url: ''
165
+ local_dir: /new-home/buddhiw/mamba-cd/third_party/FractalMambaPlusPlus
166
+ auto_setup: true
167
+ auto_discover_repo: true
168
+ verify_fractal_scan: true
169
+ verify_hilbert_serialization: true
170
+ verify_selective_scan: true
171
+ hardware:
172
+ device: cuda:0
173
+ gpu_ids:
174
+ - 1
175
+ mixed_precision: true
176
+ training:
177
+ device: cuda:0
178
+ iterations: 50000
179
+ lr: 0.0001
180
+ weight_decay: 0.01
181
+ optimizer: adamw
182
+ scheduler: cosine
183
+ warmup_iters: 5000
184
+ amp: true
185
+ grad_clip: 1.0
186
+ log_every: 50
187
+ val_every: 5000
188
+ save_best_only: true
189
+ best_metric: F1
190
+ higher_is_better: true
191
+ loss:
192
+ bce: 1.0
193
+ dice: 0.5
194
+ aux: 0.2
195
+ boundary: 0.2
196
+ residual_reg: 0.02
197
+ use_pos_weight: true
198
+ pos_weight: auto
199
+ max_pos_weight: 5.0
200
+ boundary_width: 3
201
+ boundary_dice: 0.2
202
+ deep_supervision: 0.2
203
+ deep_weights:
204
+ - 0.4
205
+ - 0.3
206
+ - 0.2
207
+ - 0.1
208
+ _computed_pos_weight: 5.0
209
+ eval:
210
+ threshold: 0.5
211
+ sweep_val_threshold: true
212
+ threshold_min: 0.05
213
+ threshold_max: 0.95
214
+ threshold_step: 0.05
215
+ use_val_threshold_for_test: true
216
+ also_report_test_sweep: true
217
+ save_predictions: true
218
+ test_split: test
219
+ allow_val_as_test: false
220
+ boundary_tolerance: 2
221
+ checkpoint:
222
+ resume: false
223
+ resume_path: null
224
+ test_checkpoint: /new-home/buddhiw/mamba-cd/outputs/run_20260607_230247_LEVIR-CD__vmamba_baseline/checkpoints/best_iter_025000_F1_0.9247.pth
225
+ auto_find_best: true
226
+ save_optimizer: true
227
+ save_scheduler: true
228
+ logging:
229
+ tensorboard: true
230
+ save_train_log: true
231
+ save_config: true
232
+ save_model_summary: true
233
+ print_params: true
234
+ print_flops: true
235
+ print_fps: true
236
+ print_peak_memory: true
237
+ save_debug_predictions: true
238
+ debug_samples: 8
239
+ outputs:
240
+ save_tensorboard: false
241
+ save_logs: true
242
+ save_qualitative: true
243
+ ablation:
244
+ enabled: false
245
+ config_file: configs/ablations/unified_ablation.yaml
246
+ group: none
247
+ id: baseline
248
+ description: baseline
249
+ ablation_runner:
250
+ enabled: true
251
+ output_dir: /new-home/buddhiw/mamba-cd/ablations
252
+ continue_on_error: false
253
+ restore_config_after_each_run: true
254
+ run_train: true
255
+ run_final_test_after_train: true
256
+ test_at_each_val: false
257
+ test_at_each_val_threshold_mode: val_selected
258
+ save_test_at_each_val_predictions: false
259
+ save_val_predictions: false
260
+ save_final_test_predictions: true
261
+ max_variants: null
262
+ start_from_variant: null
263
+ skip_completed: true
264
+ require_clean_dataset_verification: true
265
+ save_intermediate_checkpoints: false
266
+ save_best_only: true
267
+ official_selection_metric: val_F1
268
+ official_selection_split: val
269
+ prohibit_test_based_selection: true
270
+ dry_run: false
271
+ dry_run_print_only: false
272
+ debug_fast: false
273
+ debug_iterations: 200
274
+ debug_val_every: 100
275
+ debug_max_variants: 2
276
+ early_stop:
277
+ enabled: false
278
+ patience_validations: 5
279
+ metric: val_F1
280
+ mode: max
281
+ overfit_diagnostics:
282
+ enabled: true
283
+ track_train_metrics: true
284
+ track_val_metrics: true
285
+ track_test_metrics: true
286
+ compute_generalization_gap: true
287
+ gap_metric: F1
288
+ warning_gap: 0.05
289
+ severe_gap: 0.08
290
+ inference:
291
+ input_a: inference/A
292
+ input_b: inference/B
293
+ output: /new-home/buddhiw/mamba-cd/inference/predictions
294
+ debug:
295
+ check_shapes: true
296
+ overfit_test: false
297
+ overfit_samples: 4
298
+ overfit_iterations: 100
models/kate-cd-256/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "model": "MambaCD",
4
+ "dataset": "KATE-CD-256",
5
+ "backbone": "VMamba-small",
6
+ "input_size": 256,
7
+ "test_threshold": 0.75,
8
+ "metrics": {"f1": 0.6612196007419557, "miou": 0.7375787347664009, "overall_accuracy": 0.9815964949757444, "recall": 0.635695301171737, "precision": 0.6888804177036593, "bf1": 0.6410768758835065, "bmiou": null},
9
+ "complexity": {"gflops": null, "gpu_gb": 10.61158912, "gpu_memory_kind": "peak_reserved_training", "fps": null, "parameters_m": 57.9},
10
+ "source_run": "run_20260610_100835_KATE-CD-256_vmamba_baseline",
11
+ "source_checkpoint": "best_iter_005000_F1_0.5936.pth",
12
+ "configuration": "config.yaml",
13
+ "source_metrics": "source_test_metrics.json",
14
+ "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/kate-cd-256/model.pth?download=true"
15
+ }
models/kate-cd-256/model.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:017e2b98dce9062273613cbec9a9f249e0870b14c918b989d1a62f9e8ad8ebc9
3
+ size 688894312
models/kate-cd-256/source_test_metrics.json ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "threshold": 0.75,
3
+ "best_val_threshold": 0.75,
4
+ "test_threshold_used": 0.75,
5
+ "selection_basis": "validation_F1_only",
6
+ "temporal_input": {
7
+ "region": "features_abs",
8
+ "boundary": "features_abs"
9
+ },
10
+ "metrics": {
11
+ "f1": 0.6612196007419557,
12
+ "precision": 0.6888804177036593,
13
+ "recall": 0.635695301171737,
14
+ "iou": 0.4938975405183601,
15
+ "f1_0": 0.9905408364952892,
16
+ "iou_0": 0.9812599290144416,
17
+ "oa": 0.9815964949757444,
18
+ "mf1": 0.8258802186186225,
19
+ "miou": 0.7375787347664009,
20
+ "boundary_f1": 0.6410768758835065,
21
+ "bf1": 0.6410768758835065,
22
+ "pred_positive_ratio": 0.026070845754523026,
23
+ "gt_positive_ratio": 0.028252049496299343,
24
+ "pred_ratio": 0.026070845754523026,
25
+ "gt_ratio": 0.028252049496299343,
26
+ "mean_sigmoid_prob": 0.031036172540074115,
27
+ "mean_prob": 0.031036172540074115,
28
+ "logit_mean": -6.402840752350657,
29
+ "logit_std": 2.455388540708019,
30
+ "prob_min": 0.00012115620484109968,
31
+ "prob_max": 1.0,
32
+ "tp": 178905.0,
33
+ "fp": 80799.0,
34
+ "fn": 102527.0,
35
+ "tn": 9599241.0,
36
+ "f1_inverted_mask": 0.016257711885191192
37
+ },
38
+ "checkpoint": "/new-home/buddhiw/mamba-cd/outputs/run_20260610_100835_KATE-CD-256_vmamba_baseline/checkpoints/best_iter_005000_F1_0.5936.pth"
39
+ }
models/levir-cd-plus/config.yaml ADDED
@@ -0,0 +1,312 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ dataset:
2
+ name: LEVIR-CD+
3
+ root: /new-home/buddhiw/mamba-cd/datasets/LEVIR-CD-plus-256-test-as-val
4
+ raw_root: /new-home/buddhiw/mamba-cd/datasets/LEVIR-CD-plus/LEVIR-CD+
5
+ image_size: 256
6
+ batch_size: 8
7
+ num_workers: 8
8
+ binary_threshold: 127
9
+ split:
10
+ protocol: test_as_val_literature
11
+ source: official_train_test
12
+ train_source: official_train
13
+ val_source: official_test
14
+ test_source: official_test
15
+ allow_val_test_overlap: true
16
+ allow_train_val_overlap: false
17
+ allow_train_test_overlap: false
18
+ allow_missing_test: false
19
+ binarize_masks: true
20
+ binary_threshold: 127
21
+ prepare:
22
+ enabled: false
23
+ runtime:
24
+ fast_start: true
25
+ profile_startup: true
26
+ verification:
27
+ enabled: false
28
+ verify_dataset_before_train: false
29
+ verify_dataset_before_eval: false
30
+ check_filename_overlap: true
31
+ check_hash_overlap: false
32
+ check_mask_values: false
33
+ check_image_sizes: false
34
+ max_files_to_check: 0
35
+ manifests:
36
+ enabled: true
37
+ reuse_existing: true
38
+ rebuild: false
39
+ loss:
40
+ use_pos_weight: false
41
+ auto_pos_weight: true
42
+ pos_weight_cache: true
43
+ bce: 1.0
44
+ dice: 0.5
45
+ aux: 0.1
46
+ boundary: 0.05
47
+ residual_reg: 0.01
48
+ pos_weight: auto
49
+ max_pos_weight: 3.0
50
+ boundary_width: 2
51
+ boundary_dice: 0.05
52
+ deep_supervision: 0.1
53
+ deep_weights:
54
+ - 0.4
55
+ - 0.3
56
+ - 0.2
57
+ - 0.1
58
+ eval:
59
+ print_model_summary: false
60
+ sweep_val_threshold: true
61
+ sweep_test_threshold: false
62
+ also_report_test_sweep: true
63
+ use_val_threshold_for_test: true
64
+ allow_val_as_test: false
65
+ threshold: 0.5
66
+ threshold_min: 0.05
67
+ threshold_max: 0.95
68
+ threshold_step: 0.05
69
+ save_predictions: true
70
+ test_split: test
71
+ boundary_tolerance: 2
72
+ protocol_note:
73
+ name: LEVIR-CD+ test-as-val literature-comparable protocol
74
+ warning: The official test split is used as validation and test because LEVIR-CD+
75
+ has no released validation split. This protocol is for literature comparability
76
+ only and is not an unbiased clean validation protocol.
77
+ project:
78
+ name: mamba-cd
79
+ seed: 42
80
+ storage_root: /new-home/buddhiw/mamba-cd
81
+ output_root: /new-home/buddhiw/mamba-cd/outputs
82
+ run_name: auto
83
+ cache:
84
+ root: /new-home/buddhiw/mamba-cd/cache
85
+ dataset_config: configs/datasets/levirN.yaml
86
+ model:
87
+ name: MambaCD
88
+ encoder: vmamba
89
+ encoder_variant: base
90
+ pretrained: true
91
+ freeze_encoder: false
92
+ decoder_channels: 192
93
+ temporal_input:
94
+ region: signed_abs
95
+ boundary: signed_abs
96
+ use_drbi: true
97
+ use_decoder: true
98
+ use_boundary_refinement: true
99
+ drbi:
100
+ type: attentive
101
+ use_depthwise: true
102
+ gate_hidden_ratio: 0.25
103
+ region_gate_min: 0.2
104
+ region_gate_max: 0.8
105
+ boundary_gate_min: 0.0
106
+ boundary_gate_max: 0.4
107
+ share_region_boundary_projection: true
108
+ share_region_boundary_spatial: true
109
+ attentive:
110
+ sources:
111
+ - abs
112
+ - signed
113
+ - pair
114
+ fusion: softmax
115
+ use_channel_attention: false
116
+ use_spatial_attention: true
117
+ region_gate_type: residual
118
+ region_gate_alpha: 0.5
119
+ boundary_gate_type: hybrid_sobel_learned
120
+ boundary_sobel_init_weight: 0.7
121
+ boundary_gate_max: 0.5
122
+ decoder:
123
+ type: weighted_arf_fpn
124
+ use_dilations: true
125
+ dilation_rates:
126
+ - 1
127
+ - 2
128
+ - 3
129
+ fusion: concat
130
+ arf_stages:
131
+ - 2
132
+ - 3
133
+ - 4
134
+ deep_supervision: true
135
+ boundary_refinement:
136
+ type: localized_attentive
137
+ enabled: true
138
+ residual_scale: 0.05
139
+ use_depthwise: true
140
+ zero_init_residual: true
141
+ input_stages:
142
+ - 1
143
+ - 2
144
+ stage_indexing: one_based
145
+ empty_stage_policy: error
146
+ use_stage_attention: true
147
+ use_uncertainty: true
148
+ localize_residual: true
149
+ residual_mask_source: edge_boundary
150
+ boundary_width: 2
151
+ pretrained:
152
+ dir: /new-home/buddhiw/mamba-cd/pretrained_weights
153
+ auto_download: true
154
+ strict_load: false
155
+ allow_random_init_if_failed: false
156
+ vmamba:
157
+ tiny:
158
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm1_tiny_0230s_ckpt_epoch_264.pth
159
+ filename: vssm1_tiny_0230s_ckpt_epoch_264.pth
160
+ local_path: null
161
+ small:
162
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_small_0229_ckpt_epoch_222.pth
163
+ filename: vssm_small_0229_ckpt_epoch_222.pth
164
+ local_path: null
165
+ base:
166
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_base_0229_ckpt_epoch_237.pth
167
+ filename: vssm_base_0229_ckpt_epoch_237.pth
168
+ local_path: null
169
+ fractalmamba:
170
+ tiny:
171
+ url: ''
172
+ filename: ''
173
+ local_path: null
174
+ small:
175
+ url: ''
176
+ filename: ''
177
+ local_path: null
178
+ base:
179
+ url: ''
180
+ filename: ''
181
+ local_path: null
182
+ vmamba:
183
+ repo_url: https://github.com/MzeroMiko/VMamba.git
184
+ local_dir: third_party/VMamba
185
+ auto_setup: true
186
+ verify_selective_scan: true
187
+ fractalmamba:
188
+ repo_url: https://github.com/hkxiao/Fractal-Mamba
189
+ local_dir: /new-home/buddhiw/mamba-cd/third_party/FractalMambaPlusPlus
190
+ auto_setup: true
191
+ auto_discover_repo: true
192
+ verify_fractal_scan: true
193
+ verify_hilbert_serialization: true
194
+ verify_selective_scan: true
195
+ hardware:
196
+ device: cuda:0
197
+ gpu_ids:
198
+ - 0
199
+ mixed_precision: true
200
+ training:
201
+ device: cuda:0
202
+ iterations: 60000
203
+ lr: 8.0e-05
204
+ weight_decay: 0.03
205
+ optimizer: adamw
206
+ scheduler: cosine
207
+ warmup_iters: 1500
208
+ amp: true
209
+ grad_clip: 1.0
210
+ log_every: 50
211
+ val_every: 5000
212
+ save_best_only: true
213
+ best_metric: F1
214
+ higher_is_better: true
215
+ checkpoint_eval:
216
+ enabled: true
217
+ run_dir: latest
218
+ checkpoint_dir: null
219
+ test_split: test
220
+ checkpoint_glob: '*.pth'
221
+ sort_by_iteration: true
222
+ threshold_mode: both
223
+ threshold_min: 0.05
224
+ threshold_max: 0.95
225
+ threshold_step: 0.05
226
+ amp: true
227
+ save_predictions: false
228
+ save_probability_maps: false
229
+ batch_size_override: null
230
+ skip_completed: true
231
+ continue_on_error: true
232
+ output_subdir: checkpoint_test_evaluation
233
+ diagnostic_only: true
234
+ max_checkpoints: null
235
+ checkpoint:
236
+ resume: false
237
+ resume_path: null
238
+ test_checkpoint: null
239
+ auto_find_best: true
240
+ save_optimizer: true
241
+ save_scheduler: true
242
+ logging:
243
+ tensorboard: true
244
+ save_train_log: true
245
+ save_config: true
246
+ save_model_summary: true
247
+ print_params: true
248
+ print_flops: true
249
+ print_fps: true
250
+ print_peak_memory: true
251
+ save_debug_predictions: true
252
+ debug_samples: 8
253
+ outputs:
254
+ save_tensorboard: false
255
+ save_logs: true
256
+ save_qualitative: true
257
+ ablation:
258
+ enabled: false
259
+ config_file: configs/ablations/unified_ablation.yaml
260
+ group: none
261
+ id: baseline
262
+ description: baseline
263
+ ablation_runner:
264
+ enabled: true
265
+ output_dir: /new-home/buddhiw/mamba-cd/ablations_dsifn
266
+ continue_on_error: false
267
+ restore_config_after_each_run: true
268
+ run_train: true
269
+ run_final_test_after_train: true
270
+ test_at_each_val: false
271
+ test_at_each_val_threshold_mode: val_selected
272
+ save_test_at_each_val_predictions: false
273
+ save_val_predictions: false
274
+ save_final_test_predictions: true
275
+ max_variants: null
276
+ start_from_variant: null
277
+ skip_completed: true
278
+ require_clean_dataset_verification: true
279
+ save_intermediate_checkpoints: false
280
+ save_best_only: true
281
+ official_selection_metric: val_F1
282
+ official_selection_split: val
283
+ prohibit_test_based_selection: true
284
+ dry_run: false
285
+ dry_run_print_only: false
286
+ debug_fast: false
287
+ debug_iterations: 200
288
+ debug_val_every: 100
289
+ debug_max_variants: 2
290
+ early_stop:
291
+ enabled: false
292
+ patience_validations: 5
293
+ metric: val_F1
294
+ mode: max
295
+ overfit_diagnostics:
296
+ enabled: true
297
+ track_train_metrics: true
298
+ track_val_metrics: true
299
+ track_test_metrics: true
300
+ compute_generalization_gap: true
301
+ gap_metric: F1
302
+ warning_gap: 0.05
303
+ severe_gap: 0.08
304
+ inference:
305
+ input_a: inference/A
306
+ input_b: inference/B
307
+ output: /new-home/buddhiw/mamba-cd/inference/predictions
308
+ debug:
309
+ check_shapes: true
310
+ overfit_test: false
311
+ overfit_samples: 4
312
+ overfit_iterations: 100
models/levir-cd-plus/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "model": "MambaCD",
4
+ "dataset": "LEVIR-CD+",
5
+ "backbone": "VMamba-base",
6
+ "input_size": 256,
7
+ "test_threshold": 0.4,
8
+ "metrics": {"f1": 0.8727663069887339, "miou": 0.8816970489923059, "overall_accuracy": 0.9895284258086079, "recall": 0.8815025103225174, "precision": 0.8642025464630766, "bf1": 0.8321111119940277, "bmiou": null},
9
+ "complexity": {"gflops": null, "gpu_gb": 12.230590464, "gpu_memory_kind": "peak_reserved_training", "fps": null, "parameters_m": 96.0},
10
+ "source_run": "run_20260626_053400_LEVIR-CD__vmamba_baseline",
11
+ "source_checkpoint": "best_iter_045000_F1_0.8728.pth",
12
+ "configuration": "config.yaml",
13
+ "source_metrics": "source_test_metrics.json",
14
+ "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/levir-cd-plus/model.pth?download=true"
15
+ }
models/levir-cd-plus/model.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f7eeea9c7c3e632dbbd8961fa10c06ca26467b34593aac757eccff92f117f2ec
3
+ size 1143187652
models/levir-cd-plus/source_test_metrics.json ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "threshold": 0.4,
3
+ "best_val_threshold": 0.4,
4
+ "test_threshold_used": 0.4,
5
+ "selection_basis": "validation_F1_only",
6
+ "temporal_input": {
7
+ "region": "signed_abs",
8
+ "boundary": "signed_abs"
9
+ },
10
+ "metrics": {
11
+ "f1": 0.8727663069887339,
12
+ "precision": 0.8642025464630766,
13
+ "recall": 0.8815025103225174,
14
+ "iou": 0.7742557727293887,
15
+ "f1_0": 0.9945390075816529,
16
+ "iou_0": 0.9891383252552233,
17
+ "oa": 0.9895284258086079,
18
+ "mf1": 0.9336526572851934,
19
+ "miou": 0.8816970489923059,
20
+ "boundary_f1": 0.8321111119940277,
21
+ "bf1": 0.8321111119940277,
22
+ "pred_positive_ratio": 0.04155891791157339,
23
+ "gt_positive_ratio": 0.040743301654684136,
24
+ "pred_ratio": 0.04155891791157339,
25
+ "gt_ratio": 0.040743301654684136,
26
+ "mean_sigmoid_prob": 0.04107385310865011,
27
+ "mean_prob": 0.04107385310865011,
28
+ "logit_mean": -12.837587639625216,
29
+ "logit_std": 5.509367445324544,
30
+ "prob_min": 1.617488478225515e-12,
31
+ "prob_max": 1.0,
32
+ "tp": 13105661.0,
33
+ "fp": 2059373.0,
34
+ "fn": 1761751.0,
35
+ "tn": 347977663.0,
36
+ "f1_inverted_mask": 0.011277912333834067
37
+ },
38
+ "checkpoint": "/new-home/buddhiw/mamba-cd/outputs/run_20260626_053400_LEVIR-CD__vmamba_baseline/checkpoints/best_iter_045000_F1_0.8728.pth"
39
+ }
models/sysu-cd/config.yaml ADDED
@@ -0,0 +1,317 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ dataset:
2
+ name: SYSU-CD
3
+ root: /new-home/buddhiw/Datasets/SYSU-CD-folders
4
+ raw_root: /new-home/buddhiw/Datasets/SYSU-CD-folders
5
+ image_size: 256
6
+ batch_size: 8
7
+ num_workers: 8
8
+ binary_threshold: 127
9
+ prepare:
10
+ enabled: false
11
+ source_root: /new-home/buddhiw/Datasets/SYSU-CD-folders
12
+ target_root: /new-home/buddhiw/Datasets/SYSU-CD-folders
13
+ mode: existing_folders
14
+ existing_folders:
15
+ train_a: train/A
16
+ train_b: train/B
17
+ train_mask: train/Mask
18
+ val_a: val/A
19
+ val_b: val/B
20
+ val_mask: val/Mask
21
+ test_a: test/A
22
+ test_b: test/B
23
+ test_mask: test/Mask
24
+ split:
25
+ protocol: existing_folders
26
+ source_folders:
27
+ train_a: train/A
28
+ train_b: train/B
29
+ train_mask: train/Mask
30
+ val_a: val/A
31
+ val_b: val/B
32
+ val_mask: val/Mask
33
+ test_a: test/A
34
+ test_b: test/B
35
+ test_mask: test/Mask
36
+ report_dir: /new-home/buddhiw/mamba-cd/dataset_checks
37
+ allow_missing_test: false
38
+ check_filename_overlap: true
39
+ check_hash_overlap: true
40
+ overwrite_splits: true
41
+ binarize_masks: true
42
+ binary_threshold: 127
43
+ verification:
44
+ check_filename_overlap: false
45
+ check_hash_overlap: false
46
+ check_mask_binary: false
47
+ check_size_match: false
48
+ check_empty_masks: false
49
+ allow_empty_masks: true
50
+ allow_all_positive_masks: true
51
+ project:
52
+ name: mamba-cd
53
+ seed: 42
54
+ storage_root: /new-home/buddhiw/mamba-cd
55
+ output_root: /new-home/buddhiw/mamba-cd/outputs
56
+ run_name: auto
57
+ cache:
58
+ root: /new-home/buddhiw/mamba-cd/cache
59
+ dataset_config: configs/datasets/sysu.yaml
60
+ model:
61
+ name: MambaCD
62
+ encoder: vmamba
63
+ encoder_variant: base
64
+ pretrained: true
65
+ freeze_encoder: false
66
+ decoder_channels: 192
67
+ temporal_input:
68
+ region: signed_abs
69
+ boundary: signed_abs
70
+ use_drbi: true
71
+ use_decoder: true
72
+ use_boundary_refinement: true
73
+ drbi:
74
+ type: attentive
75
+ use_depthwise: true
76
+ gate_hidden_ratio: 0.25
77
+ region_gate_min: 0.2
78
+ region_gate_max: 0.8
79
+ boundary_gate_min: 0.0
80
+ boundary_gate_max: 0.4
81
+ share_region_boundary_projection: true
82
+ share_region_boundary_spatial: true
83
+ attentive:
84
+ sources:
85
+ - abs
86
+ - signed
87
+ - pair
88
+ fusion: softmax
89
+ use_channel_attention: false
90
+ use_spatial_attention: true
91
+ region_gate_type: residual
92
+ region_gate_alpha: 0.5
93
+ boundary_gate_type: hybrid_sobel_learned
94
+ boundary_sobel_init_weight: 0.7
95
+ boundary_gate_max: 0.5
96
+ decoder:
97
+ type: weighted_arf_fpn
98
+ use_dilations: true
99
+ dilation_rates:
100
+ - 1
101
+ - 2
102
+ - 3
103
+ fusion: concat
104
+ arf_stages:
105
+ - 2
106
+ - 3
107
+ - 4
108
+ deep_supervision: true
109
+ boundary_refinement:
110
+ type: localized_attentive
111
+ enabled: true
112
+ residual_scale: 0.05
113
+ use_depthwise: true
114
+ zero_init_residual: true
115
+ input_stages:
116
+ - 1
117
+ - 2
118
+ stage_indexing: one_based
119
+ empty_stage_policy: error
120
+ use_stage_attention: true
121
+ use_uncertainty: true
122
+ localize_residual: true
123
+ residual_mask_source: edge_boundary
124
+ boundary_width: 2
125
+ pretrained:
126
+ dir: /new-home/buddhiw/mamba-cd/pretrained_weights
127
+ auto_download: true
128
+ strict_load: false
129
+ allow_random_init_if_failed: false
130
+ vmamba:
131
+ tiny:
132
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm1_tiny_0230s_ckpt_epoch_264.pth
133
+ filename: vssm1_tiny_0230s_ckpt_epoch_264.pth
134
+ local_path: null
135
+ small:
136
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_small_0229_ckpt_epoch_222.pth
137
+ filename: vssm_small_0229_ckpt_epoch_222.pth
138
+ local_path: null
139
+ base:
140
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_base_0229_ckpt_epoch_237.pth
141
+ filename: vssm_base_0229_ckpt_epoch_237.pth
142
+ local_path: null
143
+ fractalmamba:
144
+ tiny:
145
+ url: ''
146
+ filename: ''
147
+ local_path: null
148
+ small:
149
+ url: ''
150
+ filename: ''
151
+ local_path: null
152
+ base:
153
+ url: ''
154
+ filename: ''
155
+ local_path: null
156
+ vmamba:
157
+ repo_url: https://github.com/MzeroMiko/VMamba.git
158
+ local_dir: third_party/VMamba
159
+ auto_setup: true
160
+ verify_selective_scan: true
161
+ fractalmamba:
162
+ repo_url: https://github.com/hkxiao/Fractal-Mamba
163
+ local_dir: /new-home/buddhiw/mamba-cd/third_party/FractalMambaPlusPlus
164
+ auto_setup: true
165
+ auto_discover_repo: true
166
+ verify_fractal_scan: true
167
+ verify_hilbert_serialization: true
168
+ verify_selective_scan: true
169
+ hardware:
170
+ device: cuda:1
171
+ gpu_ids:
172
+ - 1
173
+ mixed_precision: true
174
+ training:
175
+ device: cuda:1
176
+ iterations: 60000
177
+ lr: 8.0e-05
178
+ weight_decay: 0.03
179
+ optimizer: adamw
180
+ scheduler: cosine
181
+ warmup_iters: 1500
182
+ amp: true
183
+ grad_clip: 1.0
184
+ log_every: 50
185
+ val_every: 5000
186
+ save_best_only: true
187
+ best_metric: F1
188
+ higher_is_better: true
189
+ loss:
190
+ bce: 1.0
191
+ dice: 0.5
192
+ aux: 0.1
193
+ boundary: 0.05
194
+ residual_reg: 0.01
195
+ use_pos_weight: false
196
+ pos_weight: auto
197
+ auto_pos_weight: true
198
+ pos_weight_cache: true
199
+ max_pos_weight: 3.0
200
+ boundary_width: 2
201
+ boundary_dice: 0.05
202
+ deep_supervision: 0.1
203
+ deep_weights:
204
+ - 0.4
205
+ - 0.3
206
+ - 0.2
207
+ - 0.1
208
+ eval:
209
+ threshold: 0.5
210
+ sweep_val_threshold: true
211
+ threshold_min: 0.05
212
+ threshold_max: 0.95
213
+ threshold_step: 0.05
214
+ use_val_threshold_for_test: true
215
+ also_report_test_sweep: true
216
+ save_predictions: true
217
+ test_split: test
218
+ allow_val_as_test: false
219
+ boundary_tolerance: 2
220
+ checkpoint_eval:
221
+ enabled: true
222
+ run_dir: latest
223
+ checkpoint_dir: null
224
+ test_split: test
225
+ checkpoint_glob: '*.pth'
226
+ sort_by_iteration: true
227
+ threshold_mode: both
228
+ threshold_min: 0.05
229
+ threshold_max: 0.95
230
+ threshold_step: 0.05
231
+ amp: true
232
+ save_predictions: false
233
+ save_probability_maps: false
234
+ batch_size_override: null
235
+ skip_completed: true
236
+ continue_on_error: true
237
+ output_subdir: checkpoint_test_evaluation
238
+ diagnostic_only: true
239
+ max_checkpoints: null
240
+ checkpoint:
241
+ resume: false
242
+ resume_path: null
243
+ test_checkpoint: null
244
+ auto_find_best: true
245
+ save_optimizer: true
246
+ save_scheduler: true
247
+ logging:
248
+ tensorboard: true
249
+ save_train_log: true
250
+ save_config: true
251
+ save_model_summary: true
252
+ print_params: true
253
+ print_flops: true
254
+ print_fps: true
255
+ print_peak_memory: true
256
+ save_debug_predictions: true
257
+ debug_samples: 8
258
+ outputs:
259
+ save_tensorboard: false
260
+ save_logs: true
261
+ save_qualitative: true
262
+ ablation:
263
+ enabled: false
264
+ config_file: configs/ablations/unified_ablation.yaml
265
+ group: none
266
+ id: baseline
267
+ description: baseline
268
+ ablation_runner:
269
+ enabled: true
270
+ output_dir: /new-home/buddhiw/mamba-cd/ablations_dsifn
271
+ continue_on_error: false
272
+ restore_config_after_each_run: true
273
+ run_train: true
274
+ run_final_test_after_train: true
275
+ test_at_each_val: false
276
+ test_at_each_val_threshold_mode: val_selected
277
+ save_test_at_each_val_predictions: false
278
+ save_val_predictions: false
279
+ save_final_test_predictions: true
280
+ max_variants: null
281
+ start_from_variant: null
282
+ skip_completed: true
283
+ require_clean_dataset_verification: true
284
+ save_intermediate_checkpoints: false
285
+ save_best_only: true
286
+ official_selection_metric: val_F1
287
+ official_selection_split: val
288
+ prohibit_test_based_selection: true
289
+ dry_run: false
290
+ dry_run_print_only: false
291
+ debug_fast: false
292
+ debug_iterations: 200
293
+ debug_val_every: 100
294
+ debug_max_variants: 2
295
+ early_stop:
296
+ enabled: false
297
+ patience_validations: 5
298
+ metric: val_F1
299
+ mode: max
300
+ overfit_diagnostics:
301
+ enabled: true
302
+ track_train_metrics: true
303
+ track_val_metrics: true
304
+ track_test_metrics: true
305
+ compute_generalization_gap: true
306
+ gap_metric: F1
307
+ warning_gap: 0.05
308
+ severe_gap: 0.08
309
+ inference:
310
+ input_a: inference/A
311
+ input_b: inference/B
312
+ output: /new-home/buddhiw/mamba-cd/inference/predictions
313
+ debug:
314
+ check_shapes: true
315
+ overfit_test: false
316
+ overfit_samples: 4
317
+ overfit_iterations: 100
models/sysu-cd/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "model": "MambaCD",
4
+ "dataset": "SYSU-CD",
5
+ "backbone": "VMamba-base",
6
+ "input_size": 256,
7
+ "test_threshold": 0.1,
8
+ "metrics": {"f1": 0.8456742280784304, "miou": 0.8213458712248978, "overall_accuracy": 0.9278532638549769, "recall": 0.8382189478036886, "precision": 0.8532643337993921, "bf1": 0.46001918725884816, "bmiou": null},
9
+ "complexity": {"gflops": null, "gpu_gb": 12.230590464, "gpu_memory_kind": "peak_reserved_training", "fps": null, "parameters_m": 96.0},
10
+ "source_run": "run_20260627_104124_SYSU-CD_vmamba_baseline",
11
+ "source_checkpoint": "best_iter_045000_F1_0.8298.pth",
12
+ "configuration": "config.yaml",
13
+ "source_metrics": "source_test_metrics.json",
14
+ "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/sysu-cd/model.pth?download=true"
15
+ }
models/sysu-cd/model.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:67cc8bf7a8e2d7478e19a854c8f604bb983aa6239b017911a3ace1e3a9a7c622
3
+ size 1143187780
models/sysu-cd/source_test_metrics.json ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "threshold": 0.1,
3
+ "best_val_threshold": 0.1,
4
+ "test_threshold_used": 0.1,
5
+ "selection_basis": "validation_F1_only",
6
+ "temporal_input": {
7
+ "region": "signed_abs",
8
+ "boundary": "signed_abs"
9
+ },
10
+ "metrics": {
11
+ "f1": 0.8456742280784304,
12
+ "precision": 0.8532643337993921,
13
+ "recall": 0.8382189478036886,
14
+ "iou": 0.7326138901919922,
15
+ "f1_0": 0.952921760405053,
16
+ "iou_0": 0.9100778522578035,
17
+ "oa": 0.9278532638549769,
18
+ "mf1": 0.8992979942417417,
19
+ "miou": 0.8213458712248978,
20
+ "boundary_f1": 0.46001918725884816,
21
+ "bf1": 0.46001918725884816,
22
+ "pred_positive_ratio": 0.2316697883605957,
23
+ "gt_positive_ratio": 0.23582808303833008,
24
+ "pred_ratio": 0.2316697883605957,
25
+ "gt_ratio": 0.23582808303833008,
26
+ "mean_sigmoid_prob": 0.2141150074271718,
27
+ "mean_prob": 0.2141150074271718,
28
+ "logit_mean": -7.237031404554844,
29
+ "logit_std": 9.862041876026565,
30
+ "prob_min": 1.2800421313186039e-11,
31
+ "prob_max": 1.0,
32
+ "tp": 51819464.0,
33
+ "fp": 8911381.0,
34
+ "fn": 10001453.0,
35
+ "tn": 191411702.0,
36
+ "f1_inverted_mask": 0.06827198093452998
37
+ },
38
+ "checkpoint": "/new-home/buddhiw/mamba-cd/outputs/run_20260627_104124_SYSU-CD_vmamba_baseline/checkpoints/best_iter_045000_F1_0.8298.pth"
39
+ }
models/whu-cd/config.yaml ADDED
@@ -0,0 +1,309 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ dataset:
2
+ name: WHU-CD
3
+ root: /new-home/buddhiw/mamba-cd/datasets/WHU-CD
4
+ raw_root: /new-home/buddhiw/mamba-cd/datasets/WHU-CD
5
+ image_size: 256
6
+ batch_size: 8
7
+ num_workers: 8
8
+ binary_threshold: 127
9
+ prepare:
10
+ enabled: false
11
+ source_root: /new-home/buddhiw/mamba-cd/datasets/WHU-CD
12
+ target_root: /new-home/buddhiw/mamba-cd/datasets/WHU-CD
13
+ mode: existing_folders
14
+ existing_folders:
15
+ train_a: train/A
16
+ train_b: train/B
17
+ train_mask: train/OUT
18
+ val_a: val/A
19
+ val_b: val/B
20
+ val_mask: val/OUT
21
+ test_a: test/A
22
+ test_b: test/B
23
+ test_mask: test/OUT
24
+ split:
25
+ protocol: existing_folders
26
+ source_folders:
27
+ train_a: train/A
28
+ train_b: train/B
29
+ train_mask: train/OUT
30
+ val_a: val/A
31
+ val_b: val/B
32
+ val_mask: val/OUT
33
+ test_a: test/A
34
+ test_b: test/B
35
+ test_mask: test/OUT
36
+ report_dir: /new-home/buddhiw/mamba-cd/dataset_checks
37
+ allow_missing_test: false
38
+ check_filename_overlap: false
39
+ check_hash_overlap: true
40
+ overwrite_splits: true
41
+ verification:
42
+ check_filename_overlap: false
43
+ project:
44
+ name: mamba-cd
45
+ seed: 42
46
+ storage_root: /new-home/buddhiw/mamba-cd
47
+ output_root: /new-home/buddhiw/mamba-cd/outputs
48
+ run_name: auto
49
+ cache:
50
+ root: /new-home/buddhiw/mamba-cd/cache
51
+ dataset_config: configs/datasets/whu.yaml
52
+ model:
53
+ name: MambaCD
54
+ encoder: vmamba
55
+ encoder_variant: base
56
+ pretrained: true
57
+ freeze_encoder: false
58
+ decoder_channels: 192
59
+ temporal_input:
60
+ region: signed_abs
61
+ boundary: signed_abs
62
+ use_drbi: true
63
+ use_decoder: true
64
+ use_boundary_refinement: true
65
+ drbi:
66
+ type: attentive
67
+ use_depthwise: true
68
+ gate_hidden_ratio: 0.25
69
+ region_gate_min: 0.2
70
+ region_gate_max: 0.8
71
+ boundary_gate_min: 0.0
72
+ boundary_gate_max: 0.4
73
+ share_region_boundary_projection: true
74
+ share_region_boundary_spatial: true
75
+ attentive:
76
+ sources:
77
+ - abs
78
+ - signed
79
+ - pair
80
+ fusion: softmax
81
+ use_channel_attention: false
82
+ use_spatial_attention: true
83
+ region_gate_type: residual
84
+ region_gate_alpha: 0.5
85
+ boundary_gate_type: hybrid_sobel_learned
86
+ boundary_sobel_init_weight: 0.7
87
+ boundary_gate_max: 0.5
88
+ decoder:
89
+ type: weighted_arf_fpn
90
+ use_dilations: true
91
+ dilation_rates:
92
+ - 1
93
+ - 2
94
+ - 3
95
+ fusion: concat
96
+ arf_stages:
97
+ - 2
98
+ - 3
99
+ - 4
100
+ deep_supervision: true
101
+ boundary_refinement:
102
+ type: localized_attentive
103
+ enabled: true
104
+ residual_scale: 0.05
105
+ use_depthwise: true
106
+ zero_init_residual: true
107
+ input_stages:
108
+ - 1
109
+ - 2
110
+ stage_indexing: one_based
111
+ empty_stage_policy: error
112
+ use_stage_attention: true
113
+ use_uncertainty: true
114
+ localize_residual: true
115
+ residual_mask_source: edge_boundary
116
+ boundary_width: 2
117
+ pretrained:
118
+ dir: /new-home/buddhiw/mamba-cd/pretrained_weights
119
+ auto_download: true
120
+ strict_load: false
121
+ allow_random_init_if_failed: false
122
+ vmamba:
123
+ tiny:
124
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm1_tiny_0230s_ckpt_epoch_264.pth
125
+ filename: vssm1_tiny_0230s_ckpt_epoch_264.pth
126
+ local_path: null
127
+ small:
128
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_small_0229_ckpt_epoch_222.pth
129
+ filename: vssm_small_0229_ckpt_epoch_222.pth
130
+ local_path: null
131
+ base:
132
+ url: https://github.com/MzeroMiko/VMamba/releases/download/%23v2cls/vssm_base_0229_ckpt_epoch_237.pth
133
+ filename: vssm_base_0229_ckpt_epoch_237.pth
134
+ local_path: null
135
+ fractalmamba:
136
+ tiny:
137
+ url: ''
138
+ filename: ''
139
+ local_path: null
140
+ small:
141
+ url: ''
142
+ filename: ''
143
+ local_path: null
144
+ base:
145
+ url: ''
146
+ filename: ''
147
+ local_path: null
148
+ vmamba:
149
+ repo_url: https://github.com/MzeroMiko/VMamba.git
150
+ local_dir: third_party/VMamba
151
+ auto_setup: true
152
+ verify_selective_scan: true
153
+ fractalmamba:
154
+ repo_url: https://github.com/hkxiao/Fractal-Mamba
155
+ local_dir: /new-home/buddhiw/mamba-cd/third_party/FractalMambaPlusPlus
156
+ auto_setup: true
157
+ auto_discover_repo: true
158
+ verify_fractal_scan: true
159
+ verify_hilbert_serialization: true
160
+ verify_selective_scan: true
161
+ hardware:
162
+ device: cuda:1
163
+ gpu_ids:
164
+ - 1
165
+ mixed_precision: true
166
+ training:
167
+ device: cuda:1
168
+ iterations: 60000
169
+ lr: 8.0e-05
170
+ weight_decay: 0.03
171
+ optimizer: adamw
172
+ scheduler: cosine
173
+ warmup_iters: 1500
174
+ amp: true
175
+ grad_clip: 1.0
176
+ log_every: 50
177
+ val_every: 5000
178
+ save_best_only: true
179
+ best_metric: F1
180
+ higher_is_better: true
181
+ loss:
182
+ bce: 1.0
183
+ dice: 0.5
184
+ aux: 0.1
185
+ boundary: 0.05
186
+ residual_reg: 0.01
187
+ use_pos_weight: false
188
+ pos_weight: auto
189
+ auto_pos_weight: true
190
+ pos_weight_cache: true
191
+ max_pos_weight: 3.0
192
+ boundary_width: 2
193
+ boundary_dice: 0.05
194
+ deep_supervision: 0.1
195
+ deep_weights:
196
+ - 0.4
197
+ - 0.3
198
+ - 0.2
199
+ - 0.1
200
+ eval:
201
+ threshold: 0.5
202
+ sweep_val_threshold: true
203
+ threshold_min: 0.05
204
+ threshold_max: 0.95
205
+ threshold_step: 0.05
206
+ use_val_threshold_for_test: true
207
+ also_report_test_sweep: true
208
+ save_predictions: true
209
+ test_split: test
210
+ allow_val_as_test: false
211
+ boundary_tolerance: 2
212
+ checkpoint_eval:
213
+ enabled: true
214
+ run_dir: latest
215
+ checkpoint_dir: null
216
+ test_split: test
217
+ checkpoint_glob: '*.pth'
218
+ sort_by_iteration: true
219
+ threshold_mode: both
220
+ threshold_min: 0.05
221
+ threshold_max: 0.95
222
+ threshold_step: 0.05
223
+ amp: true
224
+ save_predictions: false
225
+ save_probability_maps: false
226
+ batch_size_override: null
227
+ skip_completed: true
228
+ continue_on_error: true
229
+ output_subdir: checkpoint_test_evaluation
230
+ diagnostic_only: true
231
+ max_checkpoints: null
232
+ checkpoint:
233
+ resume: false
234
+ resume_path: null
235
+ test_checkpoint: null
236
+ auto_find_best: true
237
+ save_optimizer: true
238
+ save_scheduler: true
239
+ logging:
240
+ tensorboard: true
241
+ save_train_log: true
242
+ save_config: true
243
+ save_model_summary: true
244
+ print_params: true
245
+ print_flops: true
246
+ print_fps: true
247
+ print_peak_memory: true
248
+ save_debug_predictions: true
249
+ debug_samples: 8
250
+ outputs:
251
+ save_tensorboard: false
252
+ save_logs: true
253
+ save_qualitative: true
254
+ ablation:
255
+ enabled: false
256
+ config_file: configs/ablations/unified_ablation.yaml
257
+ group: none
258
+ id: baseline
259
+ description: baseline
260
+ ablation_runner:
261
+ enabled: true
262
+ output_dir: /new-home/buddhiw/mamba-cd/ablations_dsifn
263
+ continue_on_error: false
264
+ restore_config_after_each_run: true
265
+ run_train: true
266
+ run_final_test_after_train: true
267
+ test_at_each_val: false
268
+ test_at_each_val_threshold_mode: val_selected
269
+ save_test_at_each_val_predictions: false
270
+ save_val_predictions: false
271
+ save_final_test_predictions: true
272
+ max_variants: null
273
+ start_from_variant: null
274
+ skip_completed: true
275
+ require_clean_dataset_verification: true
276
+ save_intermediate_checkpoints: false
277
+ save_best_only: true
278
+ official_selection_metric: val_F1
279
+ official_selection_split: val
280
+ prohibit_test_based_selection: true
281
+ dry_run: false
282
+ dry_run_print_only: false
283
+ debug_fast: false
284
+ debug_iterations: 200
285
+ debug_val_every: 100
286
+ debug_max_variants: 2
287
+ early_stop:
288
+ enabled: false
289
+ patience_validations: 5
290
+ metric: val_F1
291
+ mode: max
292
+ overfit_diagnostics:
293
+ enabled: true
294
+ track_train_metrics: true
295
+ track_val_metrics: true
296
+ track_test_metrics: true
297
+ compute_generalization_gap: true
298
+ gap_metric: F1
299
+ warning_gap: 0.05
300
+ severe_gap: 0.08
301
+ inference:
302
+ input_a: inference/A
303
+ input_b: inference/B
304
+ output: /new-home/buddhiw/mamba-cd/inference/predictions
305
+ debug:
306
+ check_shapes: true
307
+ overfit_test: false
308
+ overfit_samples: 4
309
+ overfit_iterations: 100
models/whu-cd/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "model": "MambaCD",
4
+ "dataset": "WHU-CD",
5
+ "backbone": "VMamba-base",
6
+ "input_size": 256,
7
+ "test_threshold": 0.55,
8
+ "metrics": {"f1": 0.961700476753602, "miou": 0.9611873548742476, "overall_accuracy": 0.9963250948687977, "recall": 0.9572860751495865, "precision": 0.9661567891254874, "bf1": 0.9320983315613572, "bmiou": null},
9
+ "complexity": {"gflops": null, "gpu_gb": 12.230590464, "gpu_memory_kind": "peak_reserved_training", "fps": null, "parameters_m": 96.0},
10
+ "source_run": "run_20260626_095006_WHU-CD_vmamba_baseline",
11
+ "source_checkpoint": "best_iter_040000_F1_0.9634.pth",
12
+ "configuration": "config.yaml",
13
+ "source_metrics": "source_test_metrics.json",
14
+ "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/whu-cd/model.pth?download=true"
15
+ }
models/whu-cd/model.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2c34d021aa5de928fc1f3f9909cc24bac459ef213c5197f094d60c737dd5fc98
3
+ size 1143187524
models/whu-cd/source_test_metrics.json ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "threshold": 0.55,
3
+ "best_val_threshold": 0.55,
4
+ "test_threshold_used": 0.55,
5
+ "selection_basis": "validation_F1_only",
6
+ "temporal_input": {
7
+ "region": "signed_abs",
8
+ "boundary": "signed_abs"
9
+ },
10
+ "metrics": {
11
+ "f1": 0.961700476753602,
12
+ "precision": 0.9661567891254874,
13
+ "recall": 0.9572860751495865,
14
+ "iou": 0.9262273730414387,
15
+ "f1_0": 0.9980694504382447,
16
+ "iou_0": 0.9961473367070564,
17
+ "oa": 0.9963250948687977,
18
+ "mf1": 0.9798849635959234,
19
+ "miou": 0.9611873548742476,
20
+ "boundary_f1": 0.9320983315613572,
21
+ "bf1": 0.9320983315613572,
22
+ "pred_positive_ratio": 0.047755223872467605,
23
+ "gt_positive_ratio": 0.04819774877993766,
24
+ "pred_ratio": 0.047755223872467605,
25
+ "gt_ratio": 0.04819774877993766,
26
+ "mean_sigmoid_prob": 0.047904059919522665,
27
+ "mean_prob": 0.047904059919522665,
28
+ "logit_mean": -14.222080956606728,
29
+ "logit_std": 6.705911390110826,
30
+ "prob_min": 2.813519816430299e-13,
31
+ "prob_max": 1.0,
32
+ "tp": 2304111.0,
33
+ "fp": 80710.0,
34
+ "fn": 102809.0,
35
+ "tn": 47450802.0,
36
+ "f1_inverted_mask": 0.0032337202807333804
37
+ },
38
+ "checkpoint": "/new-home/buddhiw/mamba-cd/outputs/run_20260626_095006_WHU-CD_vmamba_baseline/checkpoints/best_iter_040000_F1_0.9634.pth"
39
+ }
models/wildfires2/config.yaml ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ project:
2
+ name: MambaRefine-CD-WildFireS2
3
+ seed: 42
4
+ output_dir: outputs/wildfire_vmamba_tiny_arf256
5
+ dataset:
6
+ name: WildFireS2
7
+ root: /home/buddhiw/SemanticSegmentation/mamba-cd/datasets/WildFireS2
8
+ raw_root: null
9
+ image_size: 256
10
+ batch_size: 8
11
+ num_workers: 4
12
+ split:
13
+ protocol: existing_folders
14
+ prepared_root: null
15
+ allow_missing_test: false
16
+ overwrite_splits: false
17
+ binarize_masks: true
18
+ binary_threshold: 127
19
+ model:
20
+ encoder: vmamba
21
+ encoder_variant: tiny
22
+ pretrained: true
23
+ freeze_encoder: false
24
+ unfreeze_after: null
25
+ temporal_input: signed_abs
26
+ decoder_channels: 256
27
+ decoder: arf_fpn
28
+ use_boundary_refinement: true
29
+ hardware:
30
+ device: cuda
31
+ gpu_ids:
32
+ - 0
33
+ mixed_precision: true
34
+ training:
35
+ iterations: 3600
36
+ lr: 0.0001
37
+ weight_decay: 0.01
38
+ amp: true
39
+ grad_clip: 1.0
40
+ val_every: 360
41
+ log_every: 36
42
+ loss:
43
+ bce: 1.0
44
+ dice: 0.5
45
+ aux: 0.1
46
+ boundary: 0.0
47
+ residual_reg: 0.0
48
+ use_pos_weight: true
49
+ pos_weight: auto
50
+ max_pos_weight: 5.0
51
+ _computed_pos_weight: 0.533700020446671
52
+ eval:
53
+ threshold: 0.5
54
+ sweep_val_threshold: true
55
+ threshold_min: 0.05
56
+ threshold_max: 0.95
57
+ threshold_step: 0.05
58
+ use_val_threshold_for_test: true
59
+ also_report_test_sweep: true
60
+ save_predictions: true
61
+ test_split: test
62
+ allow_val_as_test: false
63
+ checkpoint:
64
+ resume: false
65
+ resume_path: null
66
+ test_checkpoint: null
67
+ auto_find_best: true
68
+ pretrained:
69
+ dir: pretrained_weights
70
+ auto_download: false
71
+ strict_load: false
72
+ allow_random_init_if_failed: false
73
+ vmamba:
74
+ tiny:
75
+ filename: vssm1_tiny_0230s_ckpt_epoch_264.pth
models/wildfires2/metrics.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "model": "MambaCD",
4
+ "dataset": "WildFireS2",
5
+ "backbone": "VMamba-tiny",
6
+ "input_size": 256,
7
+ "test_threshold": 0.4,
8
+ "metrics": {"f1": 0.8644930939862182, "miou": 0.647695165019166, "overall_accuracy": 0.8125961303706805, "recall": 0.9079930346206913, "precision": 0.8249714845238529, "bf1": null, "bmiou": null},
9
+ "complexity": {"gflops": null, "gpu_gb": 8.684306432, "gpu_memory_kind": "peak_reserved_training", "fps": null, "parameters_m": 44.0},
10
+ "notes": ["The source run predates the corrected symmetric boundary evaluator; its recorded BF1=0.0 is not treated as a valid measurement."],
11
+ "source_run": "wildfire_vmamba_tiny_arf256/run_20260601_111638",
12
+ "source_checkpoint": "best_iter_003600_F1_0.9492.pth",
13
+ "configuration": "config.yaml",
14
+ "source_metrics": "source_test_metrics.json",
15
+ "download_url": "https://huggingface.co/dineth18/ORBIT-Mamba/resolve/main/models/wildfires2/model.pth?download=true"
16
+ }
models/wildfires2/model.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:de722be97bbaafc28865dd1cf8fb6b390c3bc55e5140e64f8884390318f6ff42
3
+ size 522062752
models/wildfires2/source_test_metrics.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "threshold": 0.4,
3
+ "best_val_threshold": 0.4,
4
+ "test_threshold_used": 0.4,
5
+ "temporal_input": "signed_abs",
6
+ "metrics": {
7
+ "f1": 0.8644930939862182,
8
+ "precision": 0.8249714845238529,
9
+ "recall": 0.9079930346206913,
10
+ "iou": 0.7613286788930388,
11
+ "f1_0": 0.6962709375508812,
12
+ "iou_0": 0.5340616511452934,
13
+ "oa": 0.8125961303706805,
14
+ "mf1": 0.7803820157685497,
15
+ "miou": 0.647695165019166,
16
+ "boundary_f1": 0.0,
17
+ "bf1": 0.0,
18
+ "pred_positive_ratio": 0.7246220906575521,
19
+ "gt_positive_ratio": 0.6583669026692708,
20
+ "pred_ratio": 0.7246220906575521,
21
+ "gt_ratio": 0.6583669026692708,
22
+ "mean_sigmoid_prob": 0.6822513580322266,
23
+ "mean_prob": 0.6822513580322266,
24
+ "logit_mean": 1.6429937442143758,
25
+ "logit_std": 4.063245998697801,
26
+ "prob_min": 3.901368117453785e-09,
27
+ "prob_max": 0.9999992847442627,
28
+ "tp": 1175308.0,
29
+ "fp": 249357.0,
30
+ "fn": 119094.0,
31
+ "tn": 422321.0,
32
+ "f1_inverted_mask": 0.2378967025256246
33
+ },
34
+ "checkpoint": "/home/buddhiw/SemanticSegmentation/mamba-cd/outputs/wildfire_vmamba_tiny_arf256/run_20260601_111638/checkpoints/best_iter_003600_F1_0.9492.pth"
35
+ }