File size: 8,410 Bytes
2441092 | 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 | from random import randint
from string import printable
import numpy as np
import torch
from rapidfuzz.distance.Levenshtein import distance as ldistance
from torch.optim import AdamW
from models import EditDistanceModel
def pad_with_null(string: str, target_length: int):
null_char = "\0"
padding_needed = max(0, target_length - len(string))
return (string + (null_char * padding_needed))[:target_length]
def string_to_tensor(string: str, length: int) -> torch.Tensor:
"""Converts a string to a tensor of character indices."""
padded = pad_with_null(string, length)
# Use ord() to get integer representation, clamp to vocab size
indices = [min(ord(c), 127) for c in padded]
return torch.tensor(indices, dtype=torch.long)
def random_char() -> str:
pos = randint(0, len(printable) - 1)
return printable[pos]
def random_str(length: int) -> str:
return "".join([random_char() for _ in range(length)])
def mangle_string(source: str, d: int) -> str:
"""
Efficiently mangles a string to approximately the target distance
Uses list operations for better performance
"""
if d <= 0:
return source
mangled = list(source)
edits_made = 0
max_attempts = d * 3 # Prevent infinite loops
attempts = 0
while edits_made < d and attempts < max_attempts:
attempts += 1
if len(mangled) == 0:
position = 0
edit = "insert"
else:
position = randint(0, len(mangled) - 1)
edit = ["insert", "delete", "modify"][randint(0, 2)]
if edit == "insert":
mangled.insert(position, random_char())
edits_made += 1
elif edit == "modify" and len(mangled) > 0:
old_char = mangled[position]
new_char = random_char()
if old_char != new_char: # Only count as edit if actually different
mangled[position] = new_char
edits_made += 1
elif edit == "delete" and len(mangled) > 0:
mangled.pop(position)
edits_made += 1
return "".join(mangled)
def get_random_edit_distance(
minimum: int, maximum: int, mean: float, dev: float
) -> int:
sample = np.random.normal(loc=mean, scale=dev)
sample = int(sample)
return min(max(sample, minimum), maximum)
def get_homologous_pair(
source: str, length: int
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
# Use more reasonable distance distribution
distance = get_random_edit_distance(1, min(length//4, 10), 3, 2)
mangled = mangle_string(source, distance)
# Verify actual distance and use it for training
actual_distance = ldistance(source, mangled)
return (
string_to_tensor(source, length),
string_to_tensor(mangled, length),
torch.tensor(float(actual_distance), dtype=torch.float),
)
def get_non_homologous_pair(
length: int,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
source = random_str(length)
other = random_str(length)
# Ensure strings are actually different
max_attempts = 5
attempt = 0
while source == other and attempt < max_attempts:
other = random_str(length)
attempt += 1
distance = ldistance(source, other)
return (
string_to_tensor(source, length),
string_to_tensor(other, length),
torch.tensor(float(distance), dtype=torch.float),
)
def squared_euclidean_distance(v1: torch.Tensor, v2: torch.Tensor) -> torch.Tensor:
return torch.sum((v1 - v2) ** 2, dim=1)
def get_batch(
size: int, batch_size: int
) -> list[tuple[torch.Tensor, torch.Tensor, torch.Tensor]]:
half_b = int(batch_size / 2)
# Generate diverse source strings for homologous pairs
h_pairs = []
for _ in range(half_b):
source = random_str(size)
h_pairs.append(get_homologous_pair(source, size))
non_h_pairs = [get_non_homologous_pair(size) for _ in range(half_b)]
# Shuffle the batch to prevent learning batch order patterns
all_pairs = h_pairs + non_h_pairs
np.random.shuffle(all_pairs)
return all_pairs
def estimate_M(length: int, num_samples: int = 1000) -> float:
"""Estimates M, the average Levenshtein distance for non-homologous pairs."""
total_distance = 0.0
for _ in range(num_samples):
_, _, dist_tensor = get_non_homologous_pair(length)
total_distance += dist_tensor.item()
return total_distance / num_samples
def get_distances(
batch: list[tuple[torch.Tensor, torch.Tensor, torch.Tensor]],
model: torch.nn.Module,
M: float | None = None,
embedding_dim: int | None = None
):
first: torch.Tensor = torch.stack([b[0] for b in batch])
first = model(first)
second: torch.Tensor = torch.stack([b[1] for b in batch])
second = model(second)
ds = torch.stack([b[2] for b in batch])
d_hats = squared_euclidean_distance(first, second)
if M is not None and embedding_dim is not None:
# r(n) = sqrt(M / (2n)) from paper Eq. 6
# We need r(n)^2 * d_hats, so (M / (2n)) * d_hats
scaling_factor_squared = M / (2 * embedding_dim)
d_hats = d_hats * scaling_factor_squared
return (d_hats, ds)
def approximation_error(d_hat: torch.Tensor, d: torch.Tensor):
return torch.mean(torch.abs(d - d_hat))
def get_loss(d_hat: torch.Tensor, d: torch.Tensor) -> torch.Tensor:
"""
Wei et al. Poisson regression loss function
"""
# Wei et al. Poisson regression with improved numerical stability
# PNLL(d̂, d) = d̂ - d * ln(d̂) with better handling of edge cases
epsilon = 1e-8
d_hat_stable = torch.clamp(d_hat, min=epsilon)
return torch.mean(d_hat_stable - d * torch.log(d_hat_stable))
def validate_training_data(batch: list[tuple[torch.Tensor, torch.Tensor, torch.Tensor]]) -> dict:
"""Validate and analyze training batch quality"""
distances = [b[2].item() for b in batch]
stats = {
'min_distance': min(distances),
'max_distance': max(distances),
'mean_distance': np.mean(distances),
'std_distance': np.std(distances),
'zero_distance_count': sum(1 for d in distances if d == 0),
'high_distance_count': sum(1 for d in distances if d > 15)
}
return stats
def run_experiment(
embedding_dim: int,
model: torch.nn.Module,
learning_rate: float,
num_steps: int,
size: int,
batch_size: int,
use_gradient_clipping: bool = True,
max_grad_norm: float = 1.0,
distance_metric: str = "euclidean"
):
"""
Runs a training experiment with the given parameters and improved loss functions.
"""
optimizer = AdamW(model.parameters(), lr=learning_rate, weight_decay=1e-5)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=200, gamma=0.5)
final_loss = 0.0
final_approx_error = 0.0
# Estimate M once at the beginning of the experiment
M_estimate = estimate_M(size)
print(f"Estimated M (average non-homologous distance): {M_estimate:.2f}")
for x in range(num_steps):
batch = get_batch(size, batch_size)
distances = get_distances(batch, model, distance_metric, M=M_estimate, embedding_dim=embedding_dim)
loss = get_loss(distances[0], distances[1])
if x % 10 == 0:
print(
f"step: {x}, loss: {loss.item()}, approx_error: {approximation_error(distances[0], distances[1]).item()}"
)
loss.backward()
optimizer.step()
scheduler.step()
final_loss = loss.item()
final_approx_error = approximation_error(distances[0], distances[1]).item()
return final_loss, final_approx_error
if __name__ == "__main__":
embedding_dim = 140
model = EditDistanceModel(embedding_dim=embedding_dim)
final_loss, final_approx_error = run_experiment(
embedding_dim=embedding_dim,
model=model,
learning_rate=0.000817,
num_steps=1000,
size=80,
batch_size=32,
use_gradient_clipping=True,
max_grad_norm=2.463,
distance_metric="euclidean",
)
print(f"Final loss: {final_loss:.4f}")
print(f"Final approximation error: {final_approx_error:.4f}")
# Save the trained model
model_path = "megashtein_trained_model.pth"
torch.save(model.state_dict(), model_path)
print(f"\n model saved to: {model_path}")
|