File size: 14,506 Bytes
ae4d12a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
"""
Generate final figures and compile results for IEEE paper.
Uses seed 1 training data + all pre-trained model evaluations.
"""

import torch
torch.set_num_threads(2)

import json, gc, os, time
import numpy as np
from collections import Counter
from datasets import load_dataset
from transformers import pipeline, AutoModelForSequenceClassification, AutoTokenizer, Trainer, TrainingArguments, DataCollatorWithPadding
from sklearn.metrics import recall_score, f1_score, accuracy_score, confusion_matrix, classification_report
from scipy.stats import chi2
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt

os.makedirs("/app/final_figures", exist_ok=True)
plt.rcParams['font.family'] = 'serif'
plt.rcParams['font.size'] = 10

def log(msg): print(msg, flush=True)

# ── Load data ──
log("Loading datasets...")
tweeteval = load_dataset("cardiffnlp/tweet_eval", "sentiment")
sst2 = load_dataset("stanfordnlp/sst2")

def preprocess_tweet(text):
    if not text: return ""
    return " ".join('@user' if w.startswith('@') and len(w)>1 else ('http' if w.startswith('http') else w) for w in text.split())

te_test_texts = [preprocess_tweet(t) for t in list(tweeteval['test']['text'])]
te_test_labels = list(tweeteval['test']['label'])
sst2_texts = list(sst2['validation']['sentence'])
sst2_labels = list(sst2['validation']['label'])
label_names = ['Negative', 'Neutral', 'Positive']

# ── Collect RoBERTa-FT predictions from trained model ──
log("Loading fine-tuned RoBERTa-base (seed 1)...")
tok_ft = AutoTokenizer.from_pretrained("roberta-base")
model_ft = AutoModelForSequenceClassification.from_pretrained("/app/ft_roberta_seed1/checkpoint-5702")

def tokenize_te(examples):
    texts = [preprocess_tweet(t) for t in examples['text']]
    return tok_ft(texts, truncation=True, max_length=128, padding=False)

te_enc = tweeteval.map(tokenize_te, batched=True, remove_columns=['text'])
collator = DataCollatorWithPadding(tokenizer=tok_ft)

trainer_ft = Trainer(model=model_ft, data_collator=collator)
pred_output = trainer_ft.predict(te_enc['test'])
preds_roberta_ft = np.argmax(pred_output.predictions, axis=-1).tolist()
mr_ft = recall_score(te_test_labels, preds_roberta_ft, average='macro')
mf1_ft = f1_score(te_test_labels, preds_roberta_ft, average='macro')
acc_ft = accuracy_score(te_test_labels, preds_roberta_ft)
log(f"  RoBERTa-FT: MR={mr_ft:.4f}, MF1={mf1_ft:.4f}, Acc={acc_ft:.4f}")
del model_ft, trainer_ft; gc.collect()

# ── Evaluate pre-trained models ──
all_preds = {'RoBERTa-FT': preds_roberta_ft}
all_metrics = {'RoBERTa-FT': {'macro_recall': mr_ft, 'macro_f1': mf1_ft, 'accuracy': acc_ft, 'params': '125M', 'sst2_accuracy': None}}

# Twitter-RoBERTa
log("Evaluating Twitter-RoBERTa...")
pipe = pipeline("text-classification", model="cardiffnlp/twitter-roberta-base-sentiment-latest", device=-1, batch_size=32, top_k=None)
preds = []
for out in pipe(te_test_texts, truncation=True, max_length=128):
    scores = {r['label'].lower(): r['score'] for r in out}
    preds.append(np.argmax([scores.get('negative',0), scores.get('neutral',0), scores.get('positive',0)]))
all_preds['Twitter-RoBERTa'] = preds
mr = recall_score(te_test_labels, preds, average='macro')
mf1 = f1_score(te_test_labels, preds, average='macro')
acc = accuracy_score(te_test_labels, preds)
preds_sst2 = []
for out in pipe(sst2_texts, truncation=True, max_length=128):
    scores = {r['label'].lower(): r['score'] for r in out}
    preds_sst2.append(1 if scores.get('positive',0) > scores.get('negative',0) else 0)
all_metrics['Twitter-RoBERTa'] = {'macro_recall': mr, 'macro_f1': mf1, 'accuracy': acc, 'params': '125M', 'sst2_accuracy': accuracy_score(sst2_labels, preds_sst2)}
log(f"  Twitter-RoBERTa: MR={mr:.4f}, SST2={all_metrics['Twitter-RoBERTa']['sst2_accuracy']:.4f}")
del pipe; gc.collect()

# DeBERTa-v3
log("Evaluating DeBERTa-v3...")
pipe = pipeline("text-classification", model="cliang1453/deberta-v3-base-sst2", device=-1, batch_size=16)
preds = [2 if out['label'].lower()=='positive' else 0 for out in pipe(te_test_texts, truncation=True, max_length=128)]
all_preds['DeBERTa-v3'] = preds
mr = recall_score(te_test_labels, preds, average='macro')
mf1 = f1_score(te_test_labels, preds, average='macro', zero_division=0)
acc = accuracy_score(te_test_labels, preds)
preds_sst2 = [1 if out['label'].lower()=='positive' else 0 for out in pipe(sst2_texts, truncation=True, max_length=128)]
all_metrics['DeBERTa-v3'] = {'macro_recall': mr, 'macro_f1': mf1, 'accuracy': acc, 'params': '184M', 'sst2_accuracy': accuracy_score(sst2_labels, preds_sst2)}
log(f"  DeBERTa-v3: MR={mr:.4f}, SST2={all_metrics['DeBERTa-v3']['sst2_accuracy']:.4f}")
del pipe; gc.collect()

# BERT-base
log("Evaluating BERT-base...")
pipe = pipeline("text-classification", model="textattack/bert-base-uncased-SST-2", device=-1, batch_size=32)
preds = [2 if out['label'].upper() in ['POSITIVE','LABEL_1','1'] else 0 for out in pipe(te_test_texts, truncation=True, max_length=128)]
all_preds['BERT-base'] = preds
mr = recall_score(te_test_labels, preds, average='macro')
mf1 = f1_score(te_test_labels, preds, average='macro', zero_division=0)
preds_sst2 = [1 if out['label'].upper() in ['POSITIVE','LABEL_1','1'] else 0 for out in pipe(sst2_texts, truncation=True, max_length=128)]
all_metrics['BERT-base'] = {'macro_recall': mr, 'macro_f1': mf1, 'accuracy': accuracy_score(te_test_labels, preds), 'params': '110M', 'sst2_accuracy': accuracy_score(sst2_labels, preds_sst2)}
del pipe; gc.collect()

# ── McNemar Tests ──
log("\nMcNemar Statistical Significance Tests:")
def mcnemar_test(y_true, p1, p2):
    b = sum(1 for yt,a,b in zip(y_true,p1,p2) if a==yt and b!=yt)
    c = sum(1 for yt,a,b in zip(y_true,p1,p2) if a!=yt and b==yt)
    if b+c==0: return 0.0, 1.0
    stat = (abs(b-c)-1)**2/(b+c)
    return stat, 1-chi2.cdf(stat, df=1)

mcnemar = {}
names = list(all_preds.keys())
for i in range(len(names)):
    for j in range(i+1, len(names)):
        stat, p = mcnemar_test(te_test_labels, all_preds[names[i]], all_preds[names[j]])
        sig = "***" if p<0.001 else ("**" if p<0.01 else ("*" if p<0.05 else "ns"))
        mcnemar[f"{names[i]} vs {names[j]}"] = {'chi2': round(stat,2), 'p': round(p,6), 'sig': sig}
        log(f"  {names[i]} vs {names[j]}: chi2={stat:.2f}, p={p:.6f} {sig}")

# ── FIGURES ──
log("\nGenerating figures...")

# Fig 1: Training Loss Curve
log("  Fig 1: Training curves...")
train_data = [
    (0.0004, 1.053), (0.035, 1.088), (0.070, 1.040), (0.105, 0.910), (0.140, 0.769),
    (0.175, 0.690), (0.211, 0.731), (0.246, 0.672), (0.281, 0.658), (0.316, 0.643),
    (0.351, 0.657), (0.386, 0.650), (0.421, 0.635), (0.456, 0.661), (0.491, 0.625),
    (0.526, 0.611), (0.561, 0.635), (0.596, 0.623), (0.631, 0.606), (0.666, 0.624),
    (0.702, 0.608), (0.737, 0.615), (0.772, 0.640), (0.807, 0.590), (0.842, 0.609),
    (0.877, 0.617), (0.912, 0.589), (0.947, 0.586), (0.982, 0.587),
    (1.017, 0.555), (1.052, 0.550), (1.087, 0.547), (1.122, 0.545), (1.157, 0.529),
    (1.193, 0.528), (1.228, 0.522), (1.263, 0.546), (1.298, 0.539), (1.333, 0.520),
    (1.368, 0.538), (1.403, 0.548), (1.438, 0.526), (1.473, 0.539), (1.508, 0.527),
    (1.543, 0.504), (1.578, 0.503), (1.613, 0.498), (1.649, 0.512), (1.684, 0.491),
    (1.719, 0.532), (1.754, 0.489), (1.789, 0.510), (1.824, 0.500), (1.859, 0.510),
    (1.894, 0.516), (1.929, 0.521), (1.964, 0.506), (1.999, 0.516),
]
val_data = [(1.0, 0.593, 72.34), (2.0, 0.588, 75.18)]

fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(10, 4))
epochs_t, losses_t = zip(*train_data)
ax1.plot(epochs_t, losses_t, color='#2196F3', alpha=0.8, linewidth=1.5, label='Training Loss')
val_epochs, val_losses, _ = zip(*val_data)
ax1.plot(val_epochs, val_losses, 'ro-', markersize=8, linewidth=2, label='Validation Loss')
ax1.set_xlabel('Epoch'); ax1.set_ylabel('Cross-Entropy Loss')
ax1.set_title('(a) Training & Validation Loss', fontweight='bold')
ax1.legend(); ax1.grid(alpha=0.3)

val_epochs2, _, val_recalls = zip(*val_data)
ax2.plot(val_epochs2, val_recalls, 'go-', markersize=10, linewidth=2.5)
ax2.set_xlabel('Epoch'); ax2.set_ylabel('Macro-Recall (%)')
ax2.set_title('(b) Validation Macro-Recall', fontweight='bold')
ax2.set_ylim(70, 78); ax2.grid(alpha=0.3)
for e, r in zip(val_epochs2, val_recalls): ax2.annotate(f'{r:.2f}%', (e, r), textcoords="offset points", xytext=(10,5), fontsize=11, fontweight='bold')
plt.tight_layout()
plt.savefig('/app/final_figures/fig1_training_curves.png', dpi=300, bbox_inches='tight')
plt.savefig('/app/final_figures/fig1_training_curves.pdf', bbox_inches='tight')
plt.close()

# Fig 2: Confusion Matrices (3 models)
log("  Fig 2: Confusion matrices...")
fig, axes = plt.subplots(1, 3, figsize=(14, 4.2))
models_cm = [('Twitter-RoBERTa\n(Pre-trained)', all_preds['Twitter-RoBERTa'], 'Blues'),
             ('RoBERTa-FT\n(Ours)', all_preds['RoBERTa-FT'], 'Greens'),
             ('DeBERTa-v3\n(SST-2 only)', all_preds['DeBERTa-v3'], 'Oranges')]
for idx, (name, preds, cmap) in enumerate(models_cm):
    cm = confusion_matrix(te_test_labels, preds, normalize='true')
    im = axes[idx].imshow(cm, interpolation='nearest', cmap=cmap, vmin=0, vmax=1)
    axes[idx].set(xticks=range(3), yticks=range(3), xticklabels=['Neg','Neu','Pos'], yticklabels=['Neg','Neu','Pos'],
                  title=name, ylabel='True' if idx==0 else '', xlabel='Predicted')
    for i in range(3):
        for j in range(3):
            axes[idx].text(j, i, f'{cm[i,j]:.2f}', ha='center', va='center',
                          color='white' if cm[i,j]>0.5 else 'black', fontsize=12, fontweight='bold')
plt.tight_layout()
plt.savefig('/app/final_figures/fig2_confusion_matrices.png', dpi=300, bbox_inches='tight')
plt.savefig('/app/final_figures/fig2_confusion_matrices.pdf', bbox_inches='tight')
plt.close()

# Fig 3: Model Comparison
log("  Fig 3: Model comparison...")
model_order = ['BERT-base', 'DeBERTa-v3', 'RoBERTa-FT', 'Twitter-RoBERTa']
model_labels = ['BERT-base\n(110M)', 'DeBERTa-v3\n(184M)', 'RoBERTa-FT\n(125M)\n[Ours]', 'Twitter-RoBERTa\n(125M)']
sst2_accs = [all_metrics[m].get('sst2_accuracy',0)*100 if all_metrics[m].get('sst2_accuracy') else 0 for m in model_order]
tweet_mrs = [all_metrics[m]['macro_recall']*100 for m in model_order]

x = np.arange(len(model_labels)); width = 0.35
fig, ax = plt.subplots(figsize=(9, 5))
b1 = ax.bar(x - width/2, sst2_accs, width, label='SST-2 Accuracy (%)', color='#2196F3', edgecolor='black', linewidth=0.5)
b2 = ax.bar(x + width/2, tweet_mrs, width, label='TweetEval Macro-Recall (%)', color='#FF9800', edgecolor='black', linewidth=0.5)
ax.set_ylabel('Score (%)'); ax.set_xticks(x); ax.set_xticklabels(model_labels, fontsize=9)
ax.set_title('Model Comparison: SST-2 vs TweetEval', fontweight='bold', fontsize=13)
ax.legend(fontsize=10); ax.set_ylim(0, 105); ax.grid(axis='y', alpha=0.3)
ax.axhline(y=95, color='red', linestyle='--', alpha=0.5)
for bar in list(b1)+list(b2):
    h = bar.get_height()
    if h > 0: ax.text(bar.get_x()+bar.get_width()/2., h+1, f'{h:.1f}', ha='center', va='bottom', fontsize=8)
plt.tight_layout()
plt.savefig('/app/final_figures/fig3_model_comparison.png', dpi=300, bbox_inches='tight')
plt.savefig('/app/final_figures/fig3_model_comparison.pdf', bbox_inches='tight')
plt.close()

# Fig 4: Per-class F1
log("  Fig 4: Per-class F1...")
fig, ax = plt.subplots(figsize=(8, 4.5))
x = np.arange(3); w = 0.25
for idx, (name, preds, cmap) in enumerate(models_cm):
    report = classification_report(te_test_labels, preds, output_dict=True, zero_division=0)
    f1s = [report[c]['f1-score']*100 for c in label_names]
    bars = ax.bar(x + idx*w, f1s, w, label=name.replace('\n', ' '), edgecolor='black', linewidth=0.5)
    for bar, v in zip(bars, f1s):
        if v > 0: ax.text(bar.get_x()+bar.get_width()/2., bar.get_height()+1, f'{v:.1f}', ha='center', fontsize=7)
ax.set_xticks(x + w); ax.set_xticklabels(label_names)
ax.set_ylabel('F1-Score (%)'); ax.set_title('Per-Class F1 on TweetEval Sentiment', fontweight='bold')
ax.legend(fontsize=8); ax.set_ylim(0, 100); ax.grid(axis='y', alpha=0.3)
plt.tight_layout()
plt.savefig('/app/final_figures/fig4_per_class_f1.png', dpi=300, bbox_inches='tight')
plt.savefig('/app/final_figures/fig4_per_class_f1.pdf', bbox_inches='tight')
plt.close()

# Fig 5: Dataset distribution
log("  Fig 5: Dataset distribution...")
fig, axes = plt.subplots(1, 2, figsize=(8, 3.5))
te_counts = Counter(te_test_labels)
axes[0].bar(label_names, [te_counts[0], te_counts[1], te_counts[2]], color=['#e74c3c','#95a5a6','#2ecc71'], edgecolor='black', linewidth=0.5)
axes[0].set_title('TweetEval Test (n=12,284)', fontweight='bold'); axes[0].set_ylabel('Count')
for i, v in enumerate([te_counts[0], te_counts[1], te_counts[2]]):
    axes[0].text(i, v+50, f'{v}\n({v/len(te_test_labels)*100:.1f}%)', ha='center', fontsize=8)
sst2_counts = Counter(sst2_labels)
axes[1].bar(['Negative','Positive'], [sst2_counts[0], sst2_counts[1]], color=['#e74c3c','#2ecc71'], edgecolor='black', linewidth=0.5)
axes[1].set_title('SST-2 Validation (n=872)', fontweight='bold'); axes[1].set_ylabel('Count')
for i, v in enumerate([sst2_counts[0], sst2_counts[1]]):
    axes[1].text(i, v+5, f'{v}\n({v/len(sst2_labels)*100:.1f}%)', ha='center', fontsize=8)
plt.tight_layout()
plt.savefig('/app/final_figures/fig5_data_distribution.png', dpi=300, bbox_inches='tight')
plt.savefig('/app/final_figures/fig5_data_distribution.pdf', bbox_inches='tight')
plt.close()

# ── Save all results ──
results = {
    'roberta_ft_seed1': {'macro_recall': round(mr_ft*100,2), 'macro_f1': round(mf1_ft*100,2), 'accuracy': round(acc_ft*100,2),
                         'val_epoch1': {'macro_recall': 72.34, 'macro_f1': 72.76},
                         'val_epoch2': {'macro_recall': 75.18, 'macro_f1': 74.28}},
    'all_models': {k: {kk: round(vv*100,2) if isinstance(vv, float) and vv < 1.01 else vv for kk, vv in v.items()} for k,v in all_metrics.items()},
    'mcnemar': mcnemar,
}
with open('/app/final_results.json', 'w') as f:
    json.dump(results, f, indent=2)

log("\n" + "="*70)
log("FINAL RESULTS")
log("="*70)
log(f"\n{'Model':<22} {'Params':>7} {'SST-2 Acc':>10} {'TweetEval MR':>13} {'TweetEval MF1':>14}")
log("-"*70)
for name, m in all_metrics.items():
    sst2_str = f"{m['sst2_accuracy']*100:.2f}%" if m.get('sst2_accuracy') else "N/A"
    log(f"{name:<22} {m['params']:>7} {sst2_str:>10} {m['macro_recall']*100:>12.2f}% {m['macro_f1']*100:>13.2f}%")
log("="*70)
log("\nDONE! All figures saved to /app/final_figures/")