⚗️ Style Transfer: add additional ds experiments

lib/style-transfer/configs/dstream/baseline_icd_train.yaml → lib/style-transfer/configs/ds_stream/ablation_icd/20.yaml

@@ -1,35 +1,36 @@
 # @package _global_
 defaults:
-  - default
   - _self_
+  - default
 
 dataset:
-  _target_: datasets.Dataset
-  name: 0.04-0
-  topk: 20
-  percentile: 10
-  precision: true
+  name: 0.06-2-ofzh3aqu
   random_sampling: true
+  percentile: 0
+  precision: true
+  topk: 100
 
 training_args:
   _target_: transformers.TrainingArguments
   per_device_train_batch_size: 4
   gradient_accumulation_steps: 4
   logging_steps: 20
+  warmup_steps: 50
   eval_steps: 100
   evaluation_strategy: "steps"
   remove_unused_columns: true
   save_strategy: "no"
   output_dir: "models/icd"
-  lr_scheduler_type: "constant"
-  num_train_epochs: 8
+  num_train_epochs: 10
   learning_rate: 2e-5
 
-wandb_project: style-transfer-baseline-icd
+wandb_project: style-transfer-icd-seed
 model: microsoft/deberta-v3-base
+seed: 0
 hydra:
   sweeper:
     params:
-      dataset.topk: 20,50,100,400
-      dataset.precision: true
-      model: microsoft/deberta-v3-base
+      dataset.name: 0.06-2-ofzh3aqu
+      dataset.random_sampling: true, false
+      dataset.percentile: 20
+      seed: 0, 1, 2, 3, 4

lib/style-transfer/configs/ds_stream/ablation_icd/40.yaml

@@ -0,0 +1,35 @@
+# @package _global_
+defaults:
+  - _self_
+
+dataset:
+  name: 0.06-2-ofzh3aqu
+  random_sampling: true
+  percentile: 0
+  precision: true
+  topk: 100
+
+training_args:
+  _target_: transformers.TrainingArguments
+  per_device_train_batch_size: 4
+  gradient_accumulation_steps: 4
+  logging_steps: 20
+  warmup_steps: 50
+  eval_steps: 100
+  evaluation_strategy: "steps"
+  remove_unused_columns: true
+  save_strategy: "no"
+  output_dir: "models/icd"
+  num_train_epochs: 10
+  learning_rate: 2e-5
+
+wandb_project: style-transfer-icd-seed
+model: microsoft/deberta-v3-base
+seed: 0
+hydra:
+  sweeper:
+    params:
+      dataset.name: 0.06-2-ofzh3aqu
+      dataset.random_sampling: true, false
+      dataset.percentile: 40
+      seed: 0, 1, 2, 3, 4

lib/style-transfer/configs/ds_stream/ablation_icd/60.yaml

@@ -0,0 +1,35 @@
+# @package _global_
+defaults:
+  - _self_
+
+dataset:
+  name: 0.06-2-ofzh3aqu
+  random_sampling: true
+  percentile: 0
+  precision: true
+  topk: 100
+
+training_args:
+  _target_: transformers.TrainingArguments
+  per_device_train_batch_size: 4
+  gradient_accumulation_steps: 4
+  logging_steps: 20
+  warmup_steps: 50
+  eval_steps: 100
+  evaluation_strategy: "steps"
+  remove_unused_columns: true
+  save_strategy: "no"
+  output_dir: "models/icd"
+  num_train_epochs: 10
+  learning_rate: 2e-5
+
+wandb_project: style-transfer-icd-seed
+model: microsoft/deberta-v3-base
+seed: 0
+hydra:
+  sweeper:
+    params:
+      dataset.name: 0.06-2-ofzh3aqu
+      dataset.random_sampling: true, false
+      dataset.percentile: 60
+      seed: 0, 1, 2, 3, 4

lib/style-transfer/configs/ds_stream/ablation_icd/80.yaml

@@ -0,0 +1,35 @@
+# @package _global_
+defaults:
+  - _self_
+
+dataset:
+  name: 0.06-2-ofzh3aqu
+  random_sampling: true
+  percentile: 0
+  precision: true
+  topk: 100
+
+training_args:
+  _target_: transformers.TrainingArguments
+  per_device_train_batch_size: 4
+  gradient_accumulation_steps: 4
+  logging_steps: 20
+  warmup_steps: 50
+  eval_steps: 100
+  evaluation_strategy: "steps"
+  remove_unused_columns: true
+  save_strategy: "no"
+  output_dir: "models/icd"
+  num_train_epochs: 10
+  learning_rate: 2e-5
+
+wandb_project: style-transfer-icd-seed
+model: microsoft/deberta-v3-base
+seed: 0
+hydra:
+  sweeper:
+    params:
+      dataset.name: 0.06-2-ofzh3aqu
+      dataset.random_sampling: true, false
+      dataset.percentile: 80
+      seed: 0, 1, 2, 3, 4

lib/style-transfer/configs/ds_stream/ablation_icd/default.yaml

@@ -0,0 +1,32 @@
+# @package _global_
+dataset:
+  name: 0.06-2-ofzh3aqu
+  random_sampling: true
+  percentile: 0
+  precision: true
+  topk: 100
+
+training_args:
+  _target_: transformers.TrainingArguments
+  per_device_train_batch_size: 4
+  gradient_accumulation_steps: 4
+  logging_steps: 20
+  warmup_steps: 50
+  eval_steps: 100
+  evaluation_strategy: "steps"
+  remove_unused_columns: true
+  save_strategy: "no"
+  output_dir: "models/icd"
+  num_train_epochs: 10
+  learning_rate: 2e-5
+
+wandb_project: style-transfer-icd-seed
+model: microsoft/deberta-v3-base
+seed: 0
+hydra:
+  sweeper:
+    params:
+      dataset.name: 0.06-2-ofzh3aqu
+      dataset.random_sampling: true, false
+      dataset.percentile: ???
+      seed: 0, 1, 2, 3, 4

lib/style-transfer/configs/dstream/ablation_all.yaml → lib/style-transfer/configs/ds_stream/ablation_ner/20.yaml

@@ -1,11 +1,10 @@
 # @package _global_
 defaults:
-  - default
   - _self_
 
 dataset:
   _target_: datasets.Dataset
-  name: all
+  name: 0.04-0
   topk: 20
   percentile: 0
   precision: true
@@ -16,18 +15,22 @@ training_args:
   per_device_train_batch_size: 4
   gradient_accumulation_steps: 4
   logging_steps: 20
+  warmup_steps: 50
   eval_steps: 100
   evaluation_strategy: "steps"
   remove_unused_columns: true
   save_strategy: "no"
-  output_dir: "models/icd"
-  lr_scheduler_type: "constant"
-  num_train_epochs: 8
+  output_dir: "models/ner"
+  num_train_epochs: 10
   learning_rate: 2e-5
 
-wandb_project: style-transfer-ablation-all-percentile-1
+wandb_project: style-transfer-ner-seed
 model: microsoft/deberta-v3-base
+seed: 0
 hydra:
   sweeper:
     params:
-      dataset.topk: 20,50,100,400
+      dataset.name: 0.06-2-ofzh3aqu
+      dataset.random_sampling: true, false
+      dataset.percentile: 20
+      seed: 0, 1, 2, 3, 4

lib/style-transfer/configs/ds_stream/ablation_ner/40.yaml

@@ -0,0 +1,36 @@
+# @package _global_
+defaults:
+  - _self_
+
+dataset:
+  _target_: datasets.Dataset
+  name: 0.04-0
+  topk: 20
+  percentile: 0
+  precision: true
+  random_sampling: false
+
+training_args:
+  _target_: transformers.TrainingArguments
+  per_device_train_batch_size: 4
+  gradient_accumulation_steps: 4
+  logging_steps: 20
+  warmup_steps: 50
+  eval_steps: 100
+  evaluation_strategy: "steps"
+  remove_unused_columns: true
+  save_strategy: "no"
+  output_dir: "models/ner"
+  num_train_epochs: 10
+  learning_rate: 2e-5
+
+wandb_project: style-transfer-ner-seed
+model: microsoft/deberta-v3-base
+seed: 0
+hydra:
+  sweeper:
+    params:
+      dataset.name: 0.06-2-ofzh3aqu
+      dataset.random_sampling: true, false
+      dataset.percentile: 20
+      seed: 0, 1, 2, 3, 4

lib/style-transfer/configs/ds_stream/ablation_ner/60.yaml

@@ -0,0 +1,36 @@
+# @package _global_
+defaults:
+  - _self_
+
+dataset:
+  _target_: datasets.Dataset
+  name: 0.04-0
+  topk: 20
+  percentile: 0
+  precision: true
+  random_sampling: false
+
+training_args:
+  _target_: transformers.TrainingArguments
+  per_device_train_batch_size: 4
+  gradient_accumulation_steps: 4
+  logging_steps: 20
+  warmup_steps: 50
+  eval_steps: 100
+  evaluation_strategy: "steps"
+  remove_unused_columns: true
+  save_strategy: "no"
+  output_dir: "models/ner"
+  num_train_epochs: 10
+  learning_rate: 2e-5
+
+wandb_project: style-transfer-ner-seed
+model: microsoft/deberta-v3-base
+seed: 0
+hydra:
+  sweeper:
+    params:
+      dataset.name: 0.06-2-ofzh3aqu
+      dataset.random_sampling: true, false
+      dataset.percentile: 20
+      seed: 0, 1, 2, 3, 4

lib/style-transfer/configs/ds_stream/ablation_ner/80.yaml

@@ -0,0 +1,36 @@
+# @package _global_
+defaults:
+  - _self_
+
+dataset:
+  _target_: datasets.Dataset
+  name: 0.04-0
+  topk: 20
+  percentile: 0
+  precision: true
+  random_sampling: false
+
+training_args:
+  _target_: transformers.TrainingArguments
+  per_device_train_batch_size: 4
+  gradient_accumulation_steps: 4
+  logging_steps: 20
+  warmup_steps: 50
+  eval_steps: 100
+  evaluation_strategy: "steps"
+  remove_unused_columns: true
+  save_strategy: "no"
+  output_dir: "models/ner"
+  num_train_epochs: 10
+  learning_rate: 2e-5
+
+wandb_project: style-transfer-ner-seed
+model: microsoft/deberta-v3-base
+seed: 0
+hydra:
+  sweeper:
+    params:
+      dataset.name: 0.06-2-ofzh3aqu
+      dataset.random_sampling: true, false
+      dataset.percentile: 20
+      seed: 0, 1, 2, 3, 4

lib/style-transfer/configs/dstream/icd_train-100.yaml → lib/style-transfer/configs/ds_stream/combined_icd/100.yaml

@@ -1,8 +1,4 @@
 # @package _global_
-defaults:
-  - default
-  - _self_
-
 dataset:
   _target_: datasets.Dataset
   name: 0.04-0
@@ -27,11 +23,13 @@ training_args:
   num_train_epochs: 8
   learning_rate: 2e-5
 
-wandb_project: style-transfer-icd-5
+wandb_project: style-transfer-icd-seed
 model: microsoft/deberta-v3-base
+seed: 0
 hydra:
   sweeper:
     params:
       dataset.precision: true
-      dataset.name: gold, 0.06-0, 0.06-1-ofzh3aqu, 0.06-2-ofzh3aqu, 0.04-1-mru97w7c, 0.04-2-mru97w7c
+      dataset.name: combined
       model: microsoft/deberta-v3-base
+      seed: 0, 1, 2, 3, 4

lib/style-transfer/configs/dstream/icd_train-20.yaml → lib/style-transfer/configs/ds_stream/combined_icd/20.yaml

@@ -1,8 +1,4 @@
 # @package _global_
-defaults:
-  - default
-  - _self_
-
 dataset:
   _target_: datasets.Dataset
   name: 0.04-0
@@ -27,11 +23,13 @@ training_args:
   num_train_epochs: 8
   learning_rate: 2e-5
 
-wandb_project: style-transfer-icd-5
+wandb_project: style-transfer-icd-seed
 model: microsoft/deberta-v3-base
+seed: 0
 hydra:
   sweeper:
     params:
       dataset.precision: true
-      dataset.name: gold, 0.06-0, 0.06-1-ofzh3aqu, 0.06-2-ofzh3aqu, 0.04-1-mru97w7c, 0.04-2-mru97w7c
+      dataset.name: combined
       model: microsoft/deberta-v3-base
+      seed: 0, 1, 2, 3, 4