Final touch

bert.py

@@ -1,7 +1,5 @@
-import math
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
+from base_bert import *
+from everything import *
 
 
 class BertSelfAttention(nn.Module):

classifier.py

@@ -184,13 +184,13 @@ def classifier_run(args, custom_bert=None):
     
     print(f'Training Sentiment Classifier on {args.dataset}...')
     config = SimpleNamespace(
-        filepath=f'/kaggle/working/{args.dataset}-classifier.pt',
+        filepath=f'{args.dataset}-classifier.pt',
         lr=args.lr,
         batch_size=args.batch_size,
         fine_tune_mode=args.fine_tune_mode,
         train=args.train, dev=args.dev, test=args.test,
-        dev_out  = f'/kaggle/working/predictions/{args.fine_tune_mode}-{args.dataset}-dev-out.csv',
-        test_out = f'/kaggle/working/predictions/{args.fine_tune_mode}-{args.dataset}-test-out.csv'
+        dev_out  = f'/predictions/{args.fine_tune_mode}-{args.dataset}-dev-out.csv',
+        test_out = f'/predictions/{args.fine_tune_mode}-{args.dataset}-test-out.csv'
     )
 
     train(config, custom_bert)

constants.py

@@ -28,8 +28,8 @@ STSB_DEV = os.path.join(DATA_DIR, 'stsb-dev.parquet')
 # Training-specific constants
 SEED=11711
 NUM_CPU_CORES=4
-EPOCHS=10
-USE_GPU=True
+EPOCHS=1
+USE_GPU=False
 BATCH_SIZE_CSE=8
 BATCH_SIZE_SST=64
 BATCH_SIZE_CFIMDB=8

everything.py

@@ -1,4 +1,5 @@
 import csv
+import math
 import torch
 import random
 import numpy as np
@@ -13,6 +14,7 @@ from scipy.stats import spearmanr
 from torch.utils.data import Dataset, DataLoader
 from sklearn.metrics import f1_score, accuracy_score
 
+from utils import *
 from constants import *
 
 import random, numpy as np, argparse

{finetune-scripts → finetune_scripts}/sup.py

@@ -4,7 +4,7 @@ from finetuning import finetune_bert
 
 ARGUMENTS = SimpleNamespace(
     mode='sup',
-    filepath='/minbert-model/sup-cse-bert.pth',
+    filepath='sup-cse-bert.pth',
     batch_size_train=12,
     batch_size_dev=64,
     temp=0.05, lr=1e-5,

{finetune-scripts → finetune_scripts}/unsup.py

@@ -4,7 +4,7 @@ from finetuning import finetune_bert
 
 ARGUMENTS = SimpleNamespace(
     mode='unsup',
-    filepath='/minbert-model/unsup-cse-bert.pth',
+    filepath='unsup-cse-bert.pth',
     batch_size_train=8,
     batch_size_dev=64,
     temp=0.05, lr=1e-5,

requirements.txt

@@ -6,4 +6,6 @@ filelock==3.16.1
 sklearn==0.0
 tokenizers==0.15
 explainaboard_client==0.0.7
-pandas==2.2.3
+pandas==2.2.3
+pyarrow==18.1.0
+fastparquet==2024.11.0

run.py

@@ -0,0 +1,46 @@
+import argparse
+
+
+parser = argparse.ArgumentParser()
+
+parser.add_argument("--task", type=str, choices=('finetune', 'train'))
+parser.add_argument("--model", type=str, choices=('base', 'sup', 'unsup'))
+parser.add_argument("--dataset", type=str, choices=('sst', 'cfimdb'), default='sst')
+parser.add_argument("--train-mode", type=str, choices=('last-linear', 'full-model'), default='last-linear')
+
+
+args = parser.parse_args()
+
+if args.task == "finetune":
+    if args.model == "sup":
+        import finetune_scripts.sup
+    elif args.model == "unsup":
+        import finetune_scripts.unsup
+    else:
+        raise Exception("Invalid task for finetuning.")
+else:
+    match (args.model, args.dataset, args.train_mode):
+        case ('base', 'sst', 'last-linear'):
+            import train_scripts.base_sst_ll
+        case ('base', 'sst', 'full-model'):
+            import train_scripts.base_sst_fm
+        case ('base', 'cfimdb', 'last-linear'):
+            import train_scripts.base_cfimdb_ll
+        case ('base', 'cfimdb', 'full-model'):
+            import train_scripts.base_cfimdb_fm
+        case ('unsup', 'sst', 'last-linear'):
+            import train_scripts.unsup_sst_ll
+        case ('unsup', 'sst', 'full-model'):
+            import train_scripts.unsup_sst_fm
+        case ('unsup', 'cfimdb', 'last-linear'):
+            import train_scripts.unsup_cfimdb_ll
+        case ('unsup', 'cfimdb', 'full-model'):
+            import train_scripts.unsup_cfimdb_fm
+        case ('sup', 'sst', 'last-linear'):
+            import train_scripts.sup_sst_ll
+        case ('sup', 'sst', 'full-model'):
+            import train_scripts.sup_sst_fm
+        case ('sup', 'cfimdb', 'last-linear'):
+            import train_scripts.sup_cfimdb_ll
+        case ('sup', 'cfimdb', 'full-model'):
+            import train_scripts.sup_cfimdb_fm