chore: experiment setting

downstream/Clone-detection-CodeNet/run.sh

@@ -21,7 +21,7 @@ python ./code/run.py \
     --epoch 2 \
     --block_size 400 \
     --train_batch_size 8 \
-    --eval_batch_size 16 \
+    --eval_batch_size 64 \
     --learning_rate 2e-5 \
     --max_grad_norm 1.0 \
     --evaluate_during_training \

downstream/Clone-detection-CodeNet/run_aug_test.sh

@@ -20,7 +20,7 @@ python ./code/run.py \
     --test_predictions_file=aug_predictions.jsonl \
     --epoch 2 \
     --block_size 400 \
-    --eval_batch_size 16 \
+    --eval_batch_size 64 \
     --learning_rate 2e-5 \
     --max_grad_norm 1.0 \
     --seed 123456 2>&1| tee $output_dir/aug_train.log

downstream/Clone-detection-POJ104/run.sh

@@ -17,7 +17,7 @@ python ./code/run.py \
     --epoch 2 \
     --block_size 400 \
     --train_batch_size 8 \
-    --eval_batch_size 16 \
+    --eval_batch_size 64 \
     --learning_rate 2e-5 \
     --max_grad_norm 1.0 \
     --evaluate_during_training \

downstream/Clone-detection-POJ104/run_aug_test.sh

@@ -19,7 +19,7 @@ python ./code/run.py \
     --test_predictions_file=aug_predictions.jsonl \
     --epoch 2 \
     --block_size 400 \
-    --eval_batch_size 16 \
+    --eval_batch_size 64 \
     --learning_rate 2e-5 \
     --max_grad_norm 1.0 \
     --seed 123456 2>&1| tee $output_dir/aug_train.log

downstream/Code-classification-CodeNet/run_aug_test.sh

@@ -19,7 +19,7 @@ python ./code/run.py \
     --num_train_epochs 5 \
     --block_size 256 \
     --train_batch_size 8 \
-    --eval_batch_size 16 \
+    --eval_batch_size 64 \
     --learning_rate 2e-5 \
     --max_grad_norm 1.0 \
     --seed 123456  2>&1 | tee $output_dir/aug_test.log

downstream/Code-classification-POJ104/run_aug_test.sh

@@ -19,7 +19,7 @@ python ./code/run.py \
     --num_train_epochs 5 \
     --block_size 256 \
     --train_batch_size 8 \
-    --eval_batch_size 16 \
+    --eval_batch_size 64 \
     --learning_rate 2e-5 \
     --max_grad_norm 1.0 \
     --seed 123456  2>&1 | tee $output_dir/aug_test.log

experiments/grouped/codebert.yaml

@@ -0,0 +1,22 @@
+# InvPT pre-training: Grouped — CodeBERT
+# Effective batch size: 16 * 16 = 256
+# Usage: python modeling/cli.py run experiments/grouped/codebert.yaml
+
+dataset_path: "data/aug_csn.jsonl"
+model_name: "microsoft/codebert-base"
+
+batch_size: 16
+num_epochs: 3
+gradient_accumulation_steps: 16
+learning_rate: 2.0e-5
+
+seed: 0
+run_name: "InvCodeBERT-grouped"
+
+alpha: 1.0
+temperature: 0.1
+max_seq_length: 512
+sample_rate: 0.2
+
+contra_mode: "grouped"
+max_num_augs: 6

experiments/grouped/contrabert_c.yaml

@@ -0,0 +1,23 @@
+# InvPT pre-training: Grouped — ContraBERT_C
+# Effective batch size: 16 * 16 = 256
+# Usage: python modeling/cli.py run experiments/grouped/contrabert_c.yaml
+
+dataset_path: "data/aug_csn.jsonl"
+model_name: "./saved_models/ContraBERT_C"
+tokenizer_name: "microsoft/codebert-base"
+
+batch_size: 16
+num_epochs: 3
+gradient_accumulation_steps: 16
+learning_rate: 2.0e-5
+
+seed: 0
+run_name: "InvContraBERT_C-grouped"
+
+alpha: 1.0
+temperature: 0.1
+max_seq_length: 512
+sample_rate: 0.2
+
+contra_mode: "grouped"
+max_num_augs: 6

experiments/grouped/contrabert_g.yaml

@@ -0,0 +1,23 @@
+# InvPT pre-training: Grouped — ContraBERT_G
+# Effective batch size: 16 * 16 = 256
+# Usage: python modeling/cli.py run experiments/grouped/contrabert_g.yaml
+
+dataset_path: "data/aug_csn.jsonl"
+model_name: "./saved_models/ContraBERT_G"
+tokenizer_name: "microsoft/graphcodebert-base"
+
+batch_size: 16
+num_epochs: 3
+gradient_accumulation_steps: 16
+learning_rate: 2.0e-5
+
+seed: 0
+run_name: "InvContraBERT_G-grouped"
+
+alpha: 1.0
+temperature: 0.1
+max_seq_length: 512
+sample_rate: 0.2
+
+contra_mode: "grouped"
+max_num_augs: 6

experiments/grouped/graphcodebert.yaml

@@ -0,0 +1,22 @@
+# InvPT pre-training: Grouped — GraphCodeBERT
+# Effective batch size: 16 * 16 = 256
+# Usage: python modeling/cli.py run experiments/grouped/graphcodebert.yaml
+
+dataset_path: "data/aug_csn.jsonl"
+model_name: "microsoft/graphcodebert-base"
+
+batch_size: 16
+num_epochs: 3
+gradient_accumulation_steps: 16
+learning_rate: 2.0e-5
+
+seed: 0
+run_name: "InvGraphCodeBERT-grouped"
+
+alpha: 1.0
+temperature: 0.1
+max_seq_length: 512
+sample_rate: 0.2
+
+contra_mode: "grouped"
+max_num_augs: 6

experiments/supcon/codebert.yaml

@@ -0,0 +1,21 @@
+# InvPT pre-training: SupCon — CodeBERT
+# Effective batch size: 64 * 4 = 256
+# Usage: python modeling/cli.py run experiments/supcon/codebert.yaml
+
+dataset_path: "data/aug_csn.jsonl"
+model_name: "microsoft/codebert-base"
+
+batch_size: 64
+num_epochs: 3
+gradient_accumulation_steps: 4
+learning_rate: 2.0e-5
+
+seed: 0
+run_name: "InvCodeBERT-supcon"
+
+alpha: 1.0
+temperature: 0.1
+max_seq_length: 512
+sample_rate: 0.2
+
+contra_mode: "supcon"

experiments/supcon/contrabert_c.yaml

@@ -0,0 +1,22 @@
+# InvPT pre-training: SupCon — ContraBERT_C
+# Effective batch size: 64 * 4 = 256
+# Usage: python modeling/cli.py run experiments/supcon/contrabert_c.yaml
+
+dataset_path: "data/aug_csn.jsonl"
+model_name: "./saved_models/ContraBERT_C"
+tokenizer_name: "microsoft/codebert-base"
+
+batch_size: 64
+num_epochs: 3
+gradient_accumulation_steps: 4
+learning_rate: 2.0e-5
+
+seed: 0
+run_name: "InvContraBERT_C-supcon"
+
+alpha: 1.0
+temperature: 0.1
+max_seq_length: 512
+sample_rate: 0.2
+
+contra_mode: "supcon"

experiments/base.yaml → experiments/supcon/contrabert_g.yaml

@@ -1,5 +1,6 @@
-# InvPT pre-training experiment configuration
-# Usage: python -m modeling.pretrain --config experiments/base.yaml
+# InvPT pre-training: SupCon — ContraBERT_G
+# Effective batch size: 64 * 4 = 256
+# Usage: python modeling/cli.py run experiments/supcon/contrabert_g.yaml
 
 dataset_path: "data/aug_csn.jsonl"
 model_name: "./saved_models/ContraBERT_G"
@@ -11,12 +12,11 @@ gradient_accumulation_steps: 4
 learning_rate: 2.0e-5
 
 seed: 0
-run_name: "InvContraBERT_G-aug02-supcon"
+run_name: "InvContraBERT_G-supcon"
 
 alpha: 1.0
 temperature: 0.1
 max_seq_length: 512
 sample_rate: 0.2
 
 contra_mode: "supcon"
-# max_num_augs: 6  # only used with contra_mode: "grouped"

experiments/supcon/graphcodebert.yaml

@@ -0,0 +1,21 @@
+# InvPT pre-training: SupCon — GraphCodeBERT
+# Effective batch size: 64 * 4 = 256
+# Usage: python modeling/cli.py run experiments/supcon/graphcodebert.yaml
+
+dataset_path: "data/aug_csn.jsonl"
+model_name: "microsoft/graphcodebert-base"
+
+batch_size: 64
+num_epochs: 3
+gradient_accumulation_steps: 4
+learning_rate: 2.0e-5
+
+seed: 0
+run_name: "InvGraphCodeBERT-supcon"
+
+alpha: 1.0
+temperature: 0.1
+max_seq_length: 512
+sample_rate: 0.2
+
+contra_mode: "supcon"