Inference model and adopted (pre-trained) geo-encoders

configs/eval.yaml

@@ -1,18 +1,43 @@
 # @package _global_
 
+# specify here default configuration
+# order of defaults determines the order in which configs override each other
 defaults:
   - _self_
-  - data: mnist # choose datamodule with `test_dataloader()` for evaluation
-  - model: mnist
-  - logger: null
-  - trainer: default
-  - paths: default
+  - data: butterfly_coords
+  - model: predictive_geoclip
+  - callbacks: default
+  - logger: ${oc.env:LOGGER,wandb}
+  - trainer: ${oc.env:TRAINER_PROFILE,default}
+  - paths: ${oc.env:STORAGE_MODE,local}
   - extras: default
   - hydra: default
+  - metrics: butterfly_predictive
 
+  # experiment configs allow for version control of specific hyperparameters
+  # e.g. best hyperparameters for given model and datamodule
+  - experiment: null
+
+  # config for hyperparameter optimization
+  - hparams_search: null
+
+  # optional local config for machine/user specific settings
+  # it's optional since it doesn't need to exist and is excluded from version control
+  - optional local: default
+
+  # debugging config (enable through command line, e.g. `python train.py debug=default)
+  - debug: null
+
+# task name, determines output directory path
 task_name: "eval"
 
+# tags to help you identify your experiments
+# you can overwrite this in experiment configs
+# overwrite from command line with `python train.py tags="[first_tag, second_tag]"`
 tags: ["dev"]
 
+# seed for random number generators in pytorch, numpy and python.random
+seed: 12345
+
 # passing checkpoint path is necessary for evaluation
-ckpt_path: ???
+ckpt_path: aether/logs/train/runs/2026-03-15_13-15-08/checkpoints/epoch_098.ckpt

configs/inference.yaml

@@ -0,0 +1,25 @@
+defaults:
+  - _self_
+  - paths: ${oc.env:STORAGE_MODE,local}
+  - extras: default
+  - hydra: default
+
+# task name, determines output directory path
+task_name: "inference"
+tags: ["dev"]
+
+# If set, inference.py will load this merged checkpoint directly.
+#inference_ckpt_path: ${paths.log_dir}${task_name}/2026-04-07_12-56-16/inference_model.ckpt
+
+# If `inference_ckpt_path` is not set, stitch the inference model from:
+# - predictive ckpt (provides prediction_head weights)
+# - alignment ckpt (provides text_encoder weights + geo_encoder)
+predictive_ckpt_path: ${paths.log_dir}train/runs/2026-04-02_15-54-53/checkpoints/epoch_000.ckpt
+alignment_ckpt_path: ${paths.log_dir}train/runs/2026-04-02_15-40-03/checkpoints/epoch_000.ckpt
+
+# If set, inference.py will save a merged inference checkpoint you can reload with
+# `inference_ckpt_path`.
+save_inference_ckpt_path: ${paths.log_dir}${task_name}/${now:%Y-%m-%d}_${now:%H-%M-%S}/inference_model.ckpt
+#save_inference_ckpt_path: null
+
+training_order: ["alignment_model", "prediction_model"]

configs/paths/shared.yaml

@@ -9,7 +9,7 @@ data_dir: ${oc.env:DATA_DIR,${oc.env:SHARED_ROOT,${paths.root_dir}}/data}
 cache_dir: ${oc.env:SHARED_CACHE,${paths.data_dir}/cache}
 
 # path to logging directory
-log_dir: ${oc.env:SHARED_ROOT}/logs/
+log_dir: ${oc.env:SHARED_ROOT,${paths.root_dir}}/logs/
 
 # path to output directory, created dynamically by hydra
 # path generation pattern is specified in `configs/hydra/local.yaml`
@@ -21,4 +21,4 @@ work_dir: ${hydra:runtime.cwd}
 
 # huggingface cache directory
 # can be overridden via HF_HOME environment variable
-huggingface_cache: ${oc.env:HF_HOME,oc.env:SHARED_CACHE/huggingface}
+huggingface_cache: ${oc.env:HF_HOME,${oc.env:SHARED_CACHE}/huggingface}