Upload folder using huggingface_hub

README.md

@@ -10,10 +10,9 @@ For the Diffusers-format release with bundled inference code, see
 | --- | --- |
 | `stage1_vqgan.ckpt` | Stage 1 VQGAN checkpoint (first stage autoencoder), 721 MB. |
 | `stage2_ldm.ckpt` | Stage 2 latent diffusion / CLDM checkpoint (original A101 run, epoch 212 / step 13,991), 4.3 GB. |
-| `stage2_ldm_config.yaml` | Exact Stage 2 training config used for `stage2_ldm.ckpt`, with env-based relative paths. |
-| `ldm_backend_vqgan_marmousi.yaml` | Stage 1 training config from the refactored repo. |
-| `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. |
+| `ldm_backend_vqgan_marmousi.yaml` | Stage 1 VQGAN training config. |
+| `ldm_backend_a101_train_hwd.yaml` | Stage 2 CLDM training config, aligned with the released `stage2_ldm.ckpt`. |
+| `ldm_backend_lightning.yaml` | Lightning trainer / logger / Overthrust eval callback config (used for Stage 2). |
 
 ## Download the checkpoints
 
@@ -45,10 +44,10 @@ Clone the training code repository and copy the checkpoints/configs into it:
 ```bash
 git clone <SAII-CLDM-training-repo>
 cd SAII-CLDM
-mkdir -p models
+mkdir -p models configs
 cp saii-cldm-ldm-checkpoints/stage1_vqgan.ckpt models/
 cp saii-cldm-ldm-checkpoints/stage2_ldm.ckpt models/
-cp saii-cldm-ldm-checkpoints/stage2_ldm_config.yaml configs/
+cp saii-cldm-ldm-checkpoints/*.yaml configs/
 ```
 
 Set the required paths in `.env` (copy from `.env.example`):
@@ -58,12 +57,11 @@ FIRST_STAGE_CKPT=./models/stage1_vqgan.ckpt
 OVERTHRUST_DATA_DIR=./data/overthrust
 ```
 
-Then run CLDM inference on the Overthrust benchmark:
+Then run CLDM inference on the Overthrust benchmark. `eval_overthrust.py` defaults to `configs/ldm_backend_a101_train_hwd.yaml`, which is now aligned with the released Stage 2 checkpoint:
 
 ```bash
 CUDA_VISIBLE_DEVICES=0 python eval_overthrust.py CLDM \
   --ckpt ./models/stage2_ldm.ckpt \
-  --config ./configs/stage2_ldm_config.yaml \
   --output runs/eval_cldm \
   --device cuda --steps 30
 ```
@@ -73,7 +71,6 @@ If GPU 0 is occupied or you hit OOM, switch to another GPU:
 ```bash
 CUDA_VISIBLE_DEVICES=1 python eval_overthrust.py CLDM \
   --ckpt ./models/stage2_ldm.ckpt \
-  --config ./configs/stage2_ldm_config.yaml \
   --output runs/eval_cldm \
   --device cuda --steps 30
 ```
@@ -82,12 +79,19 @@ The script will save `metrics_summary.json`, numpy arrays, and a comparison figu
 
 ## Resume training
 
-To resume Stage 2 training from this checkpoint:
+Stage 2 training / resume from the released checkpoint:
 
 ```bash
 CUDA_VISIBLE_DEVICES=0 python train_ldm_backend.py -t \
   --resume ./models/stage2_ldm.ckpt \
-  --base ./configs/stage2_ldm_config.yaml ./configs/ldm_backend_lightning.yaml
+  --base ./configs/ldm_backend_a101_train_hwd.yaml ./configs/ldm_backend_lightning.yaml
+```
+
+Stage 1 VQGAN training from scratch:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python train_ldm_backend.py -t \
+  --base ./configs/ldm_backend_vqgan_marmousi.yaml
 ```
 
 ## Paper

ldm_backend_a101_train_hwd.yaml

@@ -14,7 +14,7 @@ model:
     cond_stage_key:
     - dipin
     - record
-    cond_stage_trainable: true
+    cond_stage_trainable: false
     conditioning_key: concat
     concat_mode: true
     image_size: 64
@@ -43,7 +43,7 @@ model:
     first_stage_config:
       target: ldm.models.autoencoder.VQModelInterface
       params:
-        ckpt_path: ${oc.env:FIRST_STAGE_CKPT,./models/pretrained/vqgan.ckpt}
+        ckpt_path: ${oc.env:FIRST_STAGE_CKPT,./models/stage1_vqgan.ckpt}
         embed_dim: 1
         n_embed: 32768
         ddconfig:
@@ -74,7 +74,7 @@ data:
   target: main.DataModuleFromConfig
   params:
     batch_size: 24
-    num_workers: 0
+    num_workers: 1
     wrap: false
     train:
       target: ldm.data.seisimic.Marmousi3_256

ldm_backend_lightning.yaml

@@ -15,7 +15,7 @@ lightning:
     params:
       entity: 601882280-xi-an-jiaotong-university-
       name: ldm-backend-a101-train-hwd
-      offline: false
+      offline: true
       project: saii-cldm-reproduce
     target: pytorch_lightning.loggers.WandbLogger
   trainer:

ldm_backend_vqgan_marmousi.yaml

@@ -76,7 +76,7 @@ lightning:
       entity: 601882280-xi-an-jiaotong-university-
       project: saii-cldm-reproduce
       name: ldm-backend-vqgan-marmousi
-      offline: false
+      offline: true
   trainer:
     benchmark: true
     gpus: 1