Simplify Hydra dirname scheme + add config options for ordinal mode i…

…n experiment configs
creatis-myriad · Nov 1, 2023 · 501c61e · 501c61e
1 parent 18fe180
commit 501c61e
Show file tree

Hide file tree

Showing 5 changed files with 17 additions and 10 deletions.
diff --git a/didactic/config/experiment/cardinal/multimodal-xformer-finetune.yaml b/didactic/config/experiment/cardinal/multimodal-xformer-finetune.yaml
@@ -12,6 +12,7 @@ excluded_clinical_attrs: ${oc.dict.keys:task.predict_losses}
 
 task:
   predict_losses: ???
+  ordinal_mode: True
   contrastive_loss:
     _target_: vital.metrics.train.metric.NTXent
   contrastive_loss_weight: 0
@@ -28,7 +29,7 @@ strict: False # Only load weights where they match the defined network, to only
 
 hydra:
   run:
-    dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/finetune/${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/${hydra.job.override_dirname}
+    dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/finetune/${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/ordinal_mode=${task.ordinal_mode},distribution=${oc.select:task.model.ordinal_head.distribution,null},tau_mode=${oc.select:task.model.ordinal_head.tau_mode,null}/${hydra.job.override_dirname}
   sweep:
     dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/finetune
-    subdir: ${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/${hydra.job.override_dirname}
+    subdir: ${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/ordinal_mode=${task.ordinal_mode},distribution=${oc.select:task.model.ordinal_head.distribution,null},tau_mode=${oc.select:task.model.ordinal_head.tau_mode,null}/${hydra.job.override_dirname}
diff --git a/didactic/config/experiment/cardinal/multimodal-xformer-head.yaml b/didactic/config/experiment/cardinal/multimodal-xformer-head.yaml
@@ -12,6 +12,7 @@ excluded_clinical_attrs: ${oc.dict.keys:task.predict_losses}
 
 task:
   predict_losses: ???
+  ordinal_mode: True
 
 callbacks:
   transformer_encoder_freeze:
@@ -24,7 +25,7 @@ strict: False # Only load weights where they match the defined network, to only
 
 hydra:
   run:
-    dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/head/${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/${hydra.job.override_dirname}
+    dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/head/${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/ordinal_mode=${task.ordinal_mode},distribution=${oc.select:task.model.ordinal_head.distribution,null},tau_mode=${oc.select:task.model.ordinal_head.tau_mode,null}/${hydra.job.override_dirname}
   sweep:
     dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/head
-    subdir: ${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/${hydra.job.override_dirname}
+    subdir: ${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/ordinal_mode=${task.ordinal_mode},distribution=${oc.select:task.model.ordinal_head.distribution,null},tau_mode=${oc.select:task.model.ordinal_head.tau_mode,null}/${hydra.job.override_dirname}
diff --git a/didactic/config/experiment/cardinal/multimodal-xformer-scratch.yaml b/didactic/config/experiment/cardinal/multimodal-xformer-scratch.yaml
@@ -10,14 +10,15 @@ excluded_clinical_attrs: ${oc.dict.keys:task.predict_losses}
 
 task:
   predict_losses: ???
+  ordinal_mode: True
   contrastive_loss:
     _target_: vital.metrics.train.metric.NTXent
   contrastive_loss_weight: 0
   mtr_p: [ 0.3, 0 ]
 
 hydra:
   run:
-    dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/scratch/${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/${hydra.job.override_dirname}
+    dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/scratch/${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/ordinal_mode=${task.ordinal_mode},distribution=${oc.select:task.model.ordinal_head.distribution,null},tau_mode=${oc.select:task.model.ordinal_head.tau_mode,null}/${hydra.job.override_dirname}
   sweep:
     dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/scratch
-    subdir: ${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/${hydra.job.override_dirname}
+    subdir: ${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/ordinal_mode=${task.ordinal_mode},distribution=${oc.select:task.model.ordinal_head.distribution,null},tau_mode=${oc.select:task.model.ordinal_head.tau_mode,null}/${hydra.job.override_dirname}
diff --git a/didactic/config/experiment/cardinal/multimodal-xformer.yaml b/didactic/config/experiment/cardinal/multimodal-xformer.yaml
@@ -122,7 +122,7 @@ callbacks:
     _target_: pytorch_lightning.callbacks.LearningRateFinder
 
 
-experiment_dirname: encoder=${hydra:runtime.choices.task/model/encoder}/img_tokenizer=${hydra:runtime.choices.task/img_tokenizer/model}/n_clinical_attrs=${n_clinical_attrs},n_img_attrs=${n_img_attrs}/contrastive=${oc.select:task.contrastive_loss_weight,0}/embed_dim=${task.embed_dim},depth=${task.model.encoder.num_layers},nhead=${task.model.encoder.encoder_layer.nhead},dropout=${task.model.encoder.encoder_layer.dropout}/mtr_p=${task.mtr_p},mt_by_attr=${task.mt_by_attr}
+experiment_dirname: img_tokenizer=${hydra:runtime.choices.task/img_tokenizer/model}/n_clinical_attrs=${n_clinical_attrs},n_img_attrs=${n_img_attrs}/contrastive=${oc.select:task.contrastive_loss_weight,0}
 hydra:
   job:
     config:
@@ -162,3 +162,7 @@ hydra:
           - task.model.encoder.encoder_layer.nhead
           - task.model.encoder.encoder_layer.dim_feedforward
           - task.model.encoder.encoder_layer.dropout
+
+          - task.ordinal_mode
+          - task.model.ordinal_head.backbone_distribution
+          - task.model.ordinal_head.tau_mode
diff --git a/didactic/config/experiment/cardinal/xtab-finetune.yaml b/didactic/config/experiment/cardinal/xtab-finetune.yaml
@@ -11,6 +11,7 @@ excluded_clinical_attrs: ${oc.dict.keys:task.predict_losses}
 
 task:
   predict_losses: ???
+  ordinal_mode: True
   contrastive_loss:
     _target_: vital.metrics.train.metric.NTXent
   contrastive_loss_weight: 0
@@ -34,10 +35,9 @@ ckpt: ??? # Make it mandatory to provide a checkpoint
 weights_only: True  # Only load the weights and ignore the hyperparameters
 strict: False # Only load weights where they match the defined network, to only some changes (e.g. heads, etc.)
 
-experiment_dirname: encoder=${hydra:runtime.choices.task/model/encoder}/img_tokenizer=${hydra:runtime.choices.task/img_tokenizer/model}/n_clinical_attrs=${n_clinical_attrs},n_img_attrs=${n_img_attrs}/contrastive=${oc.select:task.contrastive_loss_weight,0}/embed_dim=${task.embed_dim},depth=${task.model.encoder.n_blocks},nhead=${task.model.encoder.attention_n_heads},dropout=${task.model.encoder.attention_dropout},${task.model.encoder.ffn_dropout},${task.model.encoder.residual_dropout}/mtr_p=${task.mtr_p},mt_by_attr=${task.mt_by_attr}
 hydra:
   run:
-    dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/xtab-finetune/${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/${hydra.job.override_dirname}
+    dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/xtab-finetune/${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/ordinal_mode=${task.ordinal_mode},distribution=${oc.select:task.model.ordinal_head.distribution,null},tau_mode=${oc.select:task.model.ordinal_head.tau_mode,null}/${hydra.job.override_dirname}
   sweep:
     dir: ${oc.env:CARDIAC_MULTIMODAL_REPR_PATH}/xtab-finetune
-    subdir: ${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/${hydra.job.override_dirname}
+    subdir: ${experiment_dirname}/targets=${oc.dict.keys:task.predict_losses}/ordinal_mode=${task.ordinal_mode},distribution=${oc.select:task.model.ordinal_head.distribution,null},tau_mode=${oc.select:task.model.ordinal_head.tau_mode,null}/${hydra.job.override_dirname}