Upload folder using huggingface_hub

README.md

@@ -0,0 +1,46 @@
+# SAII-CLDM LDM Checkpoints
+
+This repository hosts the raw LDM-backend checkpoints for SAII-CLDM.
+For the Diffusers-format release with bundled inference code, see
+[mally-2000/saii-cldm-synthetic](https://huggingface.co/mally-2000/saii-cldm-synthetic).
+
+## Files
+
+| File | Description |
+| --- | --- |
+| `stage1_vqgan.ckpt` | Stage 1 VQGAN checkpoint (first stage autoencoder). |
+| `stage2_ldm.ckpt` | Stage 2 latent diffusion / CLDM checkpoint. |
+| `ldm_backend_vqgan_marmousi.yaml` | Stage 1 training config. |
+| `ldm_backend_a101_train_hwd.yaml` | Stage 2 training config. |
+| `ldm_backend_lightning.yaml` | Lightning trainer, logger, and Overthrust eval callback config. |
+
+## Usage
+
+Clone the training code repository and copy these checkpoints/configs into it
+(or set the corresponding `.env` paths):
+
+```bash
+cp stage1_vqgan.ckpt stage2_ldm.ckpt <your_repo>/models/
+cp *.yaml <your_repo>/configs/
+```
+
+Then set in `.env`:
+
+```bash
+FIRST_STAGE_CKPT=./models/stage1_vqgan.ckpt
+```
+
+And run the Stage 2 training resuming from `stage2_ldm.ckpt`, or evaluate it
+with:
+
+```bash
+python eval_overthrust.py CLDM \
+  --ckpt ./models/stage2_ldm.ckpt \
+  --output runs/eval_cldm \
+  --device cuda --steps 30
+```
+
+## Paper
+
+SAII-CLDM: Seismic Acoustic Impedance Inversion through Conditional Latent Diffusion Model.
+arXiv: [2506.13529](https://arxiv.org/html/2506.13529v1)

ldm_backend_a101_train_hwd.yaml

@@ -0,0 +1,104 @@
+model:
+  base_learning_rate: 2.0e-06
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.0015
+    linear_end: 0.0195
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    cond_drop: 1.0
+    first_stage_key: image
+    use_first_stage_cond:
+    - dipin
+    cond_stage_key:
+    - dipin
+    - record
+    cond_stage_trainable: true
+    conditioning_key: concat
+    concat_mode: true
+    image_size: 64
+    channels: 1
+    monitor: train/loss_simple_epoch
+    monitor_mode: min
+    monitor_top_k: 3
+    dynamic_size: false
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        in_channels: 6
+        out_channels: 1
+        model_channels: 224
+        attention_resolutions:
+        - 8
+        - 4
+        - 2
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 3
+        - 4
+        num_head_channels: 32
+    first_stage_config:
+      target: ldm.models.autoencoder.VQModelInterface
+      params:
+        ckpt_path: ${oc.env:FIRST_STAGE_CKPT,./models/pretrained/vqgan.ckpt}
+        embed_dim: 1
+        n_embed: 32768
+        ddconfig:
+          double_z: false
+          z_channels: 3
+          resolution: 256
+          in_channels: 1
+          out_ch: 1
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0
+        lossconfig:
+          target: torch.nn.Identity
+    cond_stage_config:
+      target: ldm.modules.encoders.C2f_CGuidedBlock.HWD_down4
+      params:
+        in_c: 1
+        channels: 4
+        out_c: 4
+        wave: haar
+        mode: reflect
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 24
+    num_workers: 0
+    wrap: false
+    train:
+      target: ldm.data.seisimic.Marmousi3_256
+      params:
+        dt_path: ${oc.env:MARMousi_NPZ,./data/dtA89-1.npz}
+        size: 256
+        normalize: minmax
+        ricks:
+        - 25
+        - 30
+        - 35
+        noise_snr:
+        - 20
+        dipins:
+        - 0.006
+        - 0.012
+        - 0.024
+        - 0.036
+        ricks_phase:
+        - 0
+        train_keys:
+        - image
+        - dipin
+        - record
+        noise_type: guassian_band
+        zhengyan_type: nonlinear
+        record_noraml: true

ldm_backend_lightning.yaml

@@ -0,0 +1,25 @@
+lightning:
+  callbacks:
+    overthrust_eval:
+      params:
+        dipin_v: 0.012
+        every_n_epochs: 10
+        f0: 30
+        f0_phase: 0
+        log_wandb: true
+        noise_snr: 15
+        num_timesteps: 1000
+        output_dir: runs/eval
+      target: ldm.callbacks.overthrust_eval.OverthrustEvalCallback
+  logger:
+    params:
+      entity: 601882280-xi-an-jiaotong-university-
+      name: ldm-backend-a101-train-hwd
+      offline: false
+      project: saii-cldm-reproduce
+    target: pytorch_lightning.loggers.WandbLogger
+  trainer:
+    accelerator: gpu
+    benchmark: true
+    gpus: 1
+    max_epochs: 500

ldm_backend_vqgan_marmousi.yaml

@@ -0,0 +1,83 @@
+model:
+  base_learning_rate: 4.5e-05
+  target: ldm.models.autoencoder.VQModel
+  params:
+    embed_dim: 1
+    n_embed: 32768
+    image_key: image
+    monitor: train/rec_loss
+    monitor_mode: min
+    monitor_top_k: 3
+    ddconfig:
+      double_z: false
+      z_channels: 3
+      resolution: 256
+      in_channels: 1
+      out_ch: 1
+      ch: 128
+      ch_mult:
+      - 1
+      - 2
+      - 4
+      num_res_blocks: 2
+      attn_resolutions: []
+      dropout: 0.0
+    lossconfig:
+      target: ldm.modules.losses.vqperceptual.VQLPIPSWithDiscriminator
+      params:
+        disc_conditional: false
+        disc_in_channels: 1
+        disc_start: 20000
+        disc_weight: 0.5
+        perceptual_weight: 0.5
+        codebook_weight: 1.0
+
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 5
+    num_workers: 0
+    wrap: false
+    train:
+      target: ldm.data.seisimic.Marmousi3_256
+      params:
+        dt_path: ${oc.env:MARMousi_NPZ,./data/dtA89-1.npz}
+        size: 256
+        normalize: minmax
+        ricks:
+        - 30
+        noise_snr:
+        - 0
+        dipins:
+        - 0.012
+        ricks_phase:
+        - 0
+        train_keys:
+        - image
+        noise_type: guassian_band
+        zhengyan_type: nonlinear
+        record_noraml: true
+
+lightning:
+  callbacks:
+    vqgan_reconstruction:
+      target: runtime.components.VQGanReconstructionCallback
+      params:
+        every_n_epochs: 10
+        log_first_epoch: true
+        sample_indices:
+        - 0
+        - 1
+        - 2
+        log_wandb: true
+  logger:
+    target: pytorch_lightning.loggers.WandbLogger
+    params:
+      entity: 601882280-xi-an-jiaotong-university-
+      project: saii-cldm-reproduce
+      name: ldm-backend-vqgan-marmousi
+      offline: false
+  trainer:
+    benchmark: true
+    gpus: 1
+    max_epochs: 1000

stage1_vqgan.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5657a113bb8f265a23c154f92f6a8ac077d2255a08403a4cbd8147ba4a1a72ee
+size 756178255

stage2_ldm.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f4af7c914246c9ee86c9dbb501eeed54b3b057d0674dc9a23a5fa1e61007fb3
+size 4606897830