Release FLAME pretrained weights (flame_starcop, flame_emit)

README.md

@@ -0,0 +1,61 @@
+---
+license: apache-2.0
+library_name: pytorch
+tags:
+- methane-detection
+- hyperspectral
+- semantic-segmentation
+- remote-sensing
+- fourier-neural-operator
+- aviris-ng
+- emit
+---
+
+# FLAME — pretrained weights
+
+Pretrained checkpoints for **FLAME: Physics-Guided Neural Operators for
+Onboard Satellite Methane Detection in Hyperspectral Imagery**.
+
+Code, configs, and usage instructions: https://github.com/ROKMC1250/FLAME
+
+## Files
+
+| File | Trained on | SWIR input | Seed | Config |
+|---|---|---|---|---|
+| `flame_starcop.pt` | STARCOP (AVIRIS-NG) | 72 bands, 2122–2488 nm | 42 | `flame_starcop.yaml` |
+| `flame_emit.pt` | OxHyperSyntheticCH4 (EMIT) | 64 bands, 2004–2478 nm | 42 | `flame_emit.yaml` |
+| `seeds/flame_starcop_seed43.pt` | STARCOP | — | 43 | `flame_starcop.yaml` |
+| `seeds/flame_emit_seed43.pt` | EMIT | — | 43 | `flame_emit.yaml` |
+| `seeds/flame_emit_seed44.pt` | EMIT | — | 44 | `flame_emit.yaml` |
+
+Each checkpoint is a `torch.save` dict with keys `model` (state dict),
+`epoch`, and `metric`. The matching training configuration is included next to
+the weights and is also shipped in the code repository under `configs/`.
+
+## Usage
+
+```bash
+pip install huggingface_hub
+hf download hjh1037/FLAME flame_starcop.pt --local-dir .
+hf download hjh1037/FLAME flame_emit.pt --local-dir .
+```
+
+Then follow the *Pretrained weights* section of the
+[code repository README](https://github.com/ROKMC1250/FLAME#pretrained-weights)
+to place them under `logs/<uid>/seed_<N>/weights/best.pt` for evaluation and
+visualization.
+
+```python
+import torch, yaml
+from flame.model import build_model  # from the code repository
+
+cfg = yaml.safe_load(open('configs/flame_starcop.yaml'))
+model = build_model(cfg['model']).eval()
+state = torch.load('flame_starcop.pt', map_location='cpu', weights_only=False)
+model.load_state_dict(state['model'])
+```
+
+## License
+
+Apache-2.0. The STARCOP and OxHyperSyntheticCH4 datasets are distributed under
+their own licenses.

flame_emit.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d423830004bba3f5c20d0840d9263a0fbad28ffd7ad20d56b51ba262546f1de
+size 27111928

flame_emit.yaml

@@ -0,0 +1,79 @@
+# FLAME on OxHyperSyntheticCH4 (EMIT) — reference training configuration.
+#
+# Protocol (aligned to the HyperspectralViTs training regime, arXiv:2410.17248):
+# deterministic 64x64 / stride-32 grid windows from
+# train_filtered_v2_tiled_64_32.csv, plume windows balanced to ~50% of every
+# epoch by a with-replacement weighted sampler, BCE with pos_weight capped at
+# pw_max, 50 epochs, bf16 autocast. Validation reproduces the test protocol
+# (window pooling, logits >= 0, no morphological filtering).
+#
+# This configuration trains the self-contained physics variant
+# (use_mag_in_seg: false): the segmentation head consumes the model's own
+# parameter-free physics score, and mag1c (B_magic30_tile.tif / score_divisor,
+# clipped to [0, 2]) acts only as a training-time auxiliary target for the
+# score head, co-trained at a weak weight that fades linearly over the whole
+# run (aux_weight 0.15, decay_frac 1.0). aux_bg_weight adds mild background
+# suppression (score -> 0 where mag1c is silent) and pw_max: 3 provides a mild
+# recall bootstrap — both guard against the all-negative collapse.
+#
+# Alternative (mag-in-seg) variant: set use_mag_in_seg: true so the seg head
+# consumes mag1c directly at train and inference time; then set aux_weight: 0,
+# aux_bg_weight: 0, pw_max: 1.
+
+dataset: emit
+
+model:
+  in_channels: 64                # SWIR 2004-2478 nm (EMIT)
+  width: 32
+  modes1: 16
+  modes2: 16
+  n_fno_layers: 3
+  n_ufno_layers: 3
+  dropout_rate: 0.0
+  seg_channels: [32, 16]
+  seg_kernel_size: 3
+  score_divisor: 1750.0          # shared by score normalisation and the mag1c input
+  rgb_divisor: 20.0
+  norm_type: group
+  score_clamp: [-10.0, 10.0]
+  use_mag_in_seg: false
+  spectrum_path: resources/emit/ch4_spectrum.npy
+  centers_path: resources/emit/band_centers.npy
+  baseline_stats_path: resources/emit/baseline_stats.pt
+
+train:
+  uid: flame_emit
+  lr: 5.0e-4
+  epochs: 50
+  batch_size: 18                 # must be divisible by the number of GPUs
+  n_workers: 2
+  log_dir: logs
+  seed: 42
+  patience: 25
+  ckpt_interval: 10
+  prefetch_factor: 2
+  vis_interval: 10
+  pretrain_epochs: 0
+  aux_weight: 0.15
+  decay_frac: 1.0
+  aux_bg_weight: 0.1
+  grad_clip_norm: 5.0
+  seg_loss: bce
+  pw_max: 3
+  focal_gamma: 0.0
+  val_threshold: 0.5
+  amp: bf16
+  val_window_batch: 256
+  find_unused_parameters: true
+
+data:
+  root_dir: datasets/oxhyper_synthetic_ch4
+  store_dir: datasets/emit_ram64   # flat fp16 memmap store, built automatically
+  npy_cache_dir: ~                 # optional (scripts/build_emit_npy_cache.py)
+  windows_csv: train_filtered_v2_tiled_64_32.csv
+  wv_window: [2004.0, 2478.0]
+  patch_size: 64
+  stride: 32
+  augment: true
+  max_train_windows: ~
+  max_val_tiles: ~

flame_starcop.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d65bf264ed94d8eb2a9f5eebf70c47d2b9e932dc6b41b759722ddb2af168e9a3
+size 3163354

flame_starcop.yaml

@@ -0,0 +1,63 @@
+# FLAME on STARCOP (AVIRIS-NG) — reference training configuration.
+#
+# Protocol: full-tile (512x512) training with the two-phase curriculum —
+# aux-only score pretrain (pretrain_epochs), then end-to-end segmentation with
+# the auxiliary mag1c-sas L1 loss decaying over decay_frac of the remaining
+# epochs. Validation: pixel F1 at sigmoid > 0.5 with 3x3 morphological opening.
+
+dataset: starcop
+
+model:
+  in_channels: 72                # SWIR 2122-2488 nm (AVIRIS-NG)
+  width: 14
+  modes1: 12
+  modes2: 12
+  n_fno_layers: 3
+  n_ufno_layers: 3
+  dropout_rate: 0.0
+  seg_channels: [28, 14]
+  seg_kernel_size: 3
+  score_divisor: 1750.0
+  rgb_divisor: 60.0
+  norm_type: batch
+  score_clamp: [0.0, 2.0]
+  wv_range: [2122, 2488]
+  spectrum_path: resources/starcop/ch4_spectrum.npy
+  centers_path: resources/starcop/band_centers.npy
+  baseline_stats_path: resources/starcop/baseline_stats.pt
+
+train:
+  uid: flame_starcop
+  lr: 2.0e-3
+  epochs: 80
+  batch_size: 24                 # must be divisible by the number of GPUs
+  n_workers: 4
+  log_dir: logs
+  seed: 42
+  patience: 25
+  ckpt_interval: 10
+  prefetch_factor: 1
+  vis_interval: 10
+  pretrain_epochs: 10
+  aux_weight: 1.0
+  decay_frac: 0.5
+  seg_loss: dice_bce
+  warmup_epochs: 3
+  warmup_start_factor: 0.1
+
+data:
+  train_data_path:
+    - datasets/starcop
+  val_data_path: datasets/starcop/STARCOP_allbands_Eval
+  tile_size: 512
+  npy_dir: ~                     # optional SWIR npy cache (scripts/build_starcop_npy_cache.py)
+  val_max_tiles: ~
+  wv_range: [2122, 2488]
+  load_rgb: true
+  augment: true
+
+vis:
+  n_val_plume: 15
+  n_val_noplume: 5
+  n_train: 10
+  seed: 42

seeds/flame_emit_seed43.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:441741721d9b2a783a9c8ca3fb42cf25662117286a70cb3a22219fc331473c0b
+size 27111928

seeds/flame_emit_seed44.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b140bfe53bb1641b7f22bea6ba18144793a0bccda5ef46dc98dba03648937f98
+size 27111928

seeds/flame_starcop_seed43.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c23e8f5bf8bdf2f29e2746811b78421f16dd66e507d3ec850e27ac5bc2ecf29e
+size 3163354