Upload folder using huggingface_hub

README.md

@@ -9,9 +9,10 @@ For the Diffusers-format release with bundled inference code, see
 | File | Description |
 | --- | --- |
 | `stage1_vqgan.ckpt` | Stage 1 VQGAN checkpoint (first stage autoencoder). |
-| `stage2_ldm.ckpt` | Stage 2 latent diffusion / CLDM checkpoint. |
+| `stage2_ldm.ckpt` | Stage 2 latent diffusion / CLDM checkpoint (original A101 run, epoch 212 / step 13,991). |
+| `stage2_ldm_config.yaml` | Exact Stage 2 training config used for `stage2_ldm.ckpt`, with env-based paths. |
 | `ldm_backend_vqgan_marmousi.yaml` | Stage 1 training config. |
-| `ldm_backend_a101_train_hwd.yaml` | Stage 2 training config. |
+| `ldm_backend_a101_train_hwd.yaml` | Stage 2 training config from the refactored repo. |
 | `ldm_backend_lightning.yaml` | Lightning trainer, logger, and Overthrust eval callback config. |
 
 ## Usage
@@ -21,7 +22,7 @@ Clone the training code repository and copy these checkpoints/configs into it
 
 ```bash
 cp stage1_vqgan.ckpt stage2_ldm.ckpt <your_repo>/models/
-cp *.yaml <your_repo>/configs/
+cp stage2_ldm_config.yaml <your_repo>/configs/
 ```
 
 Then set in `.env`:

stage2_ldm.ckpt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8f4af7c914246c9ee86c9dbb501eeed54b3b057d0674dc9a23a5fa1e61007fb3
-size 4606897830
+oid sha256:5ed163c70c1950ee448248410cc2a5a80c7fe9790b29d372cf467ef06ac8ad74
+size 4606892027

stage2_ldm_config.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: false
+    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/stage1_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: 1
+    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