File size: 5,775 Bytes
9333fd3 | 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 | """
models/text_encoder.py
----------------------
RoBERTa-base text encoder for sentiment + emotion classification.
Produces a 768-d [CLS] embedding + sentiment score + 7-class emotion logits.
Pre-trained on:
- SST-2 sentiment (via HuggingFace SentimentIntensityAnalyzer)
- GoEmotions 27-class → mapped to 7 core emotions
"""
from __future__ import annotations
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import RobertaModel, RobertaTokenizerFast
EMOTION_LABELS = ["Angry", "Disgust", "Fear", "Happy", "Sad", "Surprise", "Neutral"]
NUM_EMOTIONS = len(EMOTION_LABELS)
TEXT_EMBED_DIM = 768
MODEL_NAME = "roberta-base"
# Map GoEmotions 28 labels → our 7 core emotions
GOEMOTIONS_TO_CORE = {
# Angry
"anger": "Angry", "annoyance": "Angry", "disapproval": "Angry",
# Disgust
"disgust": "Disgust", "embarrassment": "Disgust",
# Fear
"fear": "Fear", "nervousness": "Fear",
# Happy
"joy": "Happy", "amusement": "Happy", "excitement": "Happy",
"gratitude": "Happy", "love": "Happy", "optimism": "Happy",
"pride": "Happy", "relief": "Happy",
# Sad
"sadness": "Sad", "grief": "Sad", "disappointment": "Sad",
"remorse": "Sad",
# Surprise
"surprise": "Surprise", "realization": "Surprise",
"confusion": "Surprise",
# Neutral
"neutral": "Neutral", "approval": "Neutral", "caring": "Neutral",
"curiosity": "Neutral", "desire": "Neutral", "admiration": "Neutral",
}
class TextEmotionEncoder(nn.Module):
"""
RoBERTa-base with two task heads:
1. Emotion classifier (7-class)
2. Sentiment regressor (0 = negative, 1 = positive)
"""
def __init__(
self,
model_name: str = MODEL_NAME,
freeze_base_layers: int = 8, # freeze bottom N of 12 transformer layers
):
super().__init__()
self.roberta = RobertaModel.from_pretrained(model_name)
self.tokenizer = RobertaTokenizerFast.from_pretrained(model_name)
# Freeze lower layers to preserve pre-trained features
for i, layer in enumerate(self.roberta.encoder.layer):
if i < freeze_base_layers:
for p in layer.parameters():
p.requires_grad = False
# Emotion head
self.emotion_head = nn.Sequential(
nn.Linear(TEXT_EMBED_DIM, 256),
nn.GELU(),
nn.Dropout(0.2),
nn.Linear(256, NUM_EMOTIONS),
)
# Sentiment head (scalar regression)
self.sentiment_head = nn.Sequential(
nn.Linear(TEXT_EMBED_DIM, 64),
nn.GELU(),
nn.Linear(64, 1),
nn.Sigmoid(), # → [0, 1]
)
def forward(
self,
input_ids: torch.Tensor,
attention_mask: torch.Tensor,
) -> dict[str, torch.Tensor]:
"""
Args:
input_ids: (B, seq_len) — tokenized text
attention_mask: (B, seq_len) — padding mask
Returns:
{
'embedding': (B, 768) — [CLS] token representation,
'logits': (B, 7) — emotion logits,
'probs': (B, 7) — emotion probabilities,
'sentiment': (B, 1) — sentiment score [0, 1],
}
"""
outputs = self.roberta(input_ids=input_ids, attention_mask=attention_mask)
cls_embedding = outputs.last_hidden_state[:, 0, :] # [CLS] token: (B, 768)
logits = self.emotion_head(cls_embedding) # (B, 7)
sentiment = self.sentiment_head(cls_embedding) # (B, 1)
return {
"embedding": cls_embedding,
"logits": logits,
"probs": F.softmax(logits, dim=-1),
"sentiment": sentiment,
}
def encode_text(self, text: str | list[str], device: str = "cpu", max_length: int = 128):
"""Tokenize and run a forward pass. Convenience wrapper for inference."""
if isinstance(text, str):
text = [text]
enc = self.tokenizer(
text,
padding=True,
truncation=True,
max_length=max_length,
return_tensors="pt",
)
enc = {k: v.to(device) for k, v in enc.items()}
with torch.no_grad():
return self.forward(enc["input_ids"], enc["attention_mask"])
def get_embedding(self, text: str | list[str], device: str = "cpu") -> torch.Tensor:
"""Returns only the [CLS] embedding — used by the fusion layer."""
return self.encode_text(text, device)["embedding"]
# ── Quick test ───────────────────────────────────────────────────────────────
if __name__ == "__main__":
print("Loading RoBERTa-base (first run downloads ~500MB)...")
model = TextEmotionEncoder()
model.eval()
samples = [
"This is exactly what I needed. Very impressed!",
"I've been waiting 45 minutes. Completely unacceptable.",
"I'm not sure if my account is secure.",
"I would like to update my shipping address.",
]
results = model.encode_text(samples)
print("\nText Encoder — forward pass:")
print(f" embedding shape : {results['embedding'].shape}") # (4, 768)
print(f" logits shape : {results['logits'].shape}") # (4, 7)
print(f" sentiments : {results['sentiment'].squeeze().tolist()}")
print(f"\nPredicted emotions:")
for text, probs in zip(samples, results["probs"]):
top = EMOTION_LABELS[probs.argmax().item()]
conf = probs.max().item()
print(f" [{top:>8s} {conf:.0%}] {text[:60]}")
print("\n✓ Text encoder OK")
|