Spaces:
Running on Zero
Running on Zero
File size: 8,517 Bytes
3d9ba5b | 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 | import os
import random
import torch
import torch.nn as nn
from typing import List, Dict, Tuple, Optional
from nlp_core.vocabulary import Vocabulary
from .encoder import Seq2SeqEncoder
from .decoder import Seq2SeqDecoder
class Seq2SeqSummarizer(nn.Module):
"""
End-to-End Sequence-to-Sequence Abstractive Summarization Model
with Bahdanau Attention, Greedy Decoding, and Beam Search.
"""
def __init__(
self,
src_vocab: Vocabulary,
trg_vocab: Vocabulary,
emb_dim: int = 256,
enc_hid_dim: int = 512,
dec_hid_dim: int = 512,
dropout: float = 0.3,
device: str = "cpu"
):
super().__init__()
self.src_vocab = src_vocab
self.trg_vocab = trg_vocab
self.device = torch.device(device if torch.cuda.is_available() and device == "cuda" else "cpu")
self.encoder = Seq2SeqEncoder(
input_dim=len(src_vocab),
emb_dim=emb_dim,
enc_hid_dim=enc_hid_dim,
dec_hid_dim=dec_hid_dim,
dropout=dropout,
pad_idx=src_vocab.PAD_IDX
)
self.decoder = Seq2SeqDecoder(
output_dim=len(trg_vocab),
emb_dim=emb_dim,
enc_hid_dim=enc_hid_dim,
dec_hid_dim=dec_hid_dim,
dropout=dropout,
pad_idx=trg_vocab.PAD_IDX
)
self.to(self.device)
def forward(
self,
src: torch.Tensor,
trg: torch.Tensor,
teacher_forcing_ratio: float = 0.5
) -> torch.Tensor:
"""
Forward pass during model training.
Args:
src: [batch_size, src_len]
trg: [batch_size, trg_len]
teacher_forcing_ratio: Probability of using ground-truth token as next input
Returns:
outputs: [batch_size, trg_len, trg_vocab_size]
"""
batch_size = src.shape[0]
trg_len = trg.shape[1]
trg_vocab_size = self.decoder.output_dim
# Tensor to store decoder outputs
outputs = torch.zeros(batch_size, trg_len, trg_vocab_size, device=self.device)
# Source mask (1 for valid token, 0 for pad)
mask = (src != self.src_vocab.PAD_IDX).to(self.device)
# Encode source sequence
encoder_outputs, hidden = self.encoder(src)
# First input to the decoder is the <sos> token
input_token = trg[:, 0]
for t in range(1, trg_len):
output, hidden, _ = self.decoder(input_token, hidden, encoder_outputs, mask=mask)
outputs[:, t, :] = output
# Decide whether to use teacher forcing
teacher_force = random.random() < teacher_forcing_ratio
top1 = output.argmax(1)
input_token = trg[:, t] if teacher_force else top1
return outputs
def summarize_greedy(
self,
src_tokens: List[str],
max_len: int = 50
) -> Tuple[List[str], torch.Tensor]:
"""
Generates summary tokens using Greedy Search decoding.
"""
self.eval()
with torch.no_grad():
src_indices = self.src_vocab.encode(src_tokens, add_sos=True, add_eos=True)
src_tensor = torch.tensor([src_indices], dtype=torch.long, device=self.device)
mask = (src_tensor != self.src_vocab.PAD_IDX).to(self.device)
encoder_outputs, hidden = self.encoder(src_tensor)
trg_indices = [self.trg_vocab.SOS_IDX]
attentions = []
for _ in range(max_len):
input_token = torch.tensor([trg_indices[-1]], dtype=torch.long, device=self.device)
output, hidden, attention = self.decoder(input_token, hidden, encoder_outputs, mask=mask)
attentions.append(attention.squeeze(0))
predicted_idx = output.argmax(1).item()
if predicted_idx == self.trg_vocab.EOS_IDX:
break
trg_indices.append(predicted_idx)
summary_tokens = self.trg_vocab.decode(trg_indices[1:], remove_special_tokens=True)
attn_tensor = torch.stack(attentions) if attentions else torch.empty(0)
return summary_tokens, attn_tensor
def summarize_beam(
self,
src_tokens: List[str],
beam_width: int = 3,
max_len: int = 50,
length_penalty: float = 0.7
) -> List[str]:
"""
Generates summary tokens using Beam Search decoding with length penalty.
"""
self.eval()
with torch.no_grad():
src_indices = self.src_vocab.encode(src_tokens, add_sos=True, add_eos=True)
src_tensor = torch.tensor([src_indices], dtype=torch.long, device=self.device)
mask = (src_tensor != self.src_vocab.PAD_IDX).to(self.device)
encoder_outputs, hidden = self.encoder(src_tensor)
# Beam state: (log_prob, [token_indices], hidden_state)
beams = [(0.0, [self.trg_vocab.SOS_IDX], hidden)]
completed_beams = []
for _ in range(max_len):
new_beams = []
for score, seq, h in beams:
if seq[-1] == self.trg_vocab.EOS_IDX:
completed_beams.append((score, seq))
continue
input_token = torch.tensor([seq[-1]], dtype=torch.long, device=self.device)
output, new_h, _ = self.decoder(input_token, h, encoder_outputs, mask=mask)
# Log softmax over output
log_probs = torch.log_softmax(output, dim=1).squeeze(0)
topk_log_probs, topk_indices = torch.topk(log_probs, beam_width)
for k in range(beam_width):
next_token = topk_indices[k].item()
next_score = score + topk_log_probs[k].item()
new_beams.append((next_score, seq + [next_token], new_h))
if not new_beams:
break
# Sort and keep top beam_width
# Apply length penalty to ranking score: score / (length ** length_penalty)
def rank_key(b):
seq_len = len(b[1])
lp = (seq_len ** length_penalty) if seq_len > 0 else 1.0
return b[0] / lp
new_beams.sort(key=rank_key, reverse=True)
beams = new_beams[:beam_width]
if len(completed_beams) >= beam_width:
break
if not completed_beams:
completed_beams = [(score, seq) for score, seq, _ in beams]
# Choose best completed beam
best_seq = max(completed_beams, key=lambda x: x[0] / (len(x[1]) ** length_penalty))[1]
summary_tokens = self.trg_vocab.decode(best_seq[1:], remove_special_tokens=True)
return summary_tokens
def save_checkpoint(self, filepath: str):
"""Saves model weights and vocabulary configuration."""
os.makedirs(os.path.dirname(filepath), exist_ok=True)
checkpoint = {
"state_dict": self.state_dict(),
"src_vocab": self.src_vocab.word2idx,
"trg_vocab": self.trg_vocab.word2idx,
"config": {
"emb_dim": self.encoder.emb_dim,
"enc_hid_dim": self.encoder.enc_hid_dim,
"dec_hid_dim": self.decoder.dec_hid_dim,
}
}
torch.save(checkpoint, filepath)
@classmethod
def load_checkpoint(cls, filepath: str, device: str = "cpu") -> 'Seq2SeqSummarizer':
"""Loads model weights and configuration from a saved checkpoint."""
checkpoint = torch.load(filepath, map_location=device)
src_vocab = Vocabulary()
src_vocab.word2idx = checkpoint["src_vocab"]
src_vocab.idx2word = {int(idx): w for w, idx in src_vocab.word2idx.items()}
trg_vocab = Vocabulary()
trg_vocab.word2idx = checkpoint["trg_vocab"]
trg_vocab.idx2word = {int(idx): w for w, idx in trg_vocab.word2idx.items()}
config = checkpoint.get("config", {})
model = cls(
src_vocab=src_vocab,
trg_vocab=trg_vocab,
emb_dim=config.get("emb_dim", 256),
enc_hid_dim=config.get("enc_hid_dim", 512),
dec_hid_dim=config.get("dec_hid_dim", 512),
device=device
)
model.load_state_dict(checkpoint["state_dict"])
return model
|