File size: 8,863 Bytes
3c68cf6 | 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 | """
Text-as-Image — Universal encoding via byte rasterization
============================================================
Text → raw UTF-8 bytes → pack into (3, H, W) → PatchSVAE → reconstruct → unpack → text
The model never knows it's looking at text. Bytes 0-255 are pixel values.
Every character, every language, every encoding. Just numbers in a grid.
For 64×64: 3 × 64 × 64 = 12,288 bytes ≈ 12KB of text per sample
For 128×128: 3 × 128 × 128 = 49,152 bytes ≈ 48KB of text per sample
Usage:
# Test existing Fresnel/Johanna on text (zero-shot)
python text_svae.py --mode test --model AbstractPhil/svae-fresnel-128
# Train a text-specific variant
python text_svae.py --mode train --size 128
# Evaluate byte-level accuracy
python text_svae.py --mode eval --checkpoint /content/checkpoints/best.pt
"""
import os
import torch
import torch.nn.functional as F
import numpy as np
from datasets import load_dataset
# ── Text ↔ Tensor Conversion ────────────────────────────────────
def text_to_tensor(text: str, H: int = 128, W: int = 128, C: int = 3) -> torch.Tensor:
"""Pack UTF-8 bytes into a (C, H, W) tensor normalized to [-1, 1].
Args:
text: Raw string
H, W: Spatial dimensions
C: Channels (3 for compatibility with existing models)
Returns:
(C, H, W) float tensor in [-1, 1]
int: actual byte count before padding
"""
raw = text.encode('utf-8')
n_bytes = C * H * W
actual_len = min(len(raw), n_bytes)
# Pad or truncate
if len(raw) < n_bytes:
raw = raw + b'\x00' * (n_bytes - len(raw))
else:
raw = raw[:n_bytes]
# To tensor: bytes 0-255 → normalized [-1, 1]
arr = np.frombuffer(raw, dtype=np.uint8).copy()
tensor = torch.from_numpy(arr).float()
tensor = (tensor / 127.5) - 1.0 # [0,255] → [-1, 1]
tensor = tensor.reshape(C, H, W)
return tensor, actual_len
def tensor_to_text(tensor: torch.Tensor, byte_count: int = None) -> str:
"""Unpack a (C, H, W) tensor back to UTF-8 text.
Args:
tensor: (C, H, W) float tensor in [-1, 1]
byte_count: if known, truncate to actual text length
Returns:
Decoded string (lossy characters replaced with ?)
"""
# Denormalize: [-1, 1] → [0, 255]
arr = ((tensor.flatten() + 1.0) * 127.5).round().clamp(0, 255).byte().numpy()
if byte_count is not None:
arr = arr[:byte_count]
# Decode with error handling
return arr.tobytes().rstrip(b'\x00').decode('utf-8', errors='replace')
# ── Text Dataset ─────────────────────────────────────────────────
class TextAsImageDataset(torch.utils.data.Dataset):
"""Stream text from HuggingFace, pack as images.
Each sample is a chunk of text rasterized to (3, H, W).
The model sees it as a regular image.
"""
def __init__(self, size=100000, img_size=128, text_field='text',
dataset_name='wikitext', dataset_config='wikitext-103-raw-v1',
split='train'):
self.size = size
self.img_size = img_size
self.n_bytes = 3 * img_size * img_size
self.text_field = text_field
# Load and concatenate all text into one big buffer
print(f" Loading text corpus: {dataset_name}/{dataset_config}...")
ds = load_dataset(dataset_name, dataset_config, split=split)
all_text = '\n'.join([x[text_field] for x in ds if x[text_field].strip()])
self.raw_bytes = all_text.encode('utf-8')
print(f" Corpus: {len(self.raw_bytes):,} bytes ({len(self.raw_bytes)/1024/1024:.1f}MB)")
def __len__(self):
return self.size
def __getitem__(self, idx):
# Random chunk from corpus
max_start = len(self.raw_bytes) - self.n_bytes
if max_start <= 0:
start = 0
else:
start = torch.randint(0, max_start, (1,)).item()
chunk = self.raw_bytes[start:start + self.n_bytes]
if len(chunk) < self.n_bytes:
chunk = chunk + b'\x00' * (self.n_bytes - len(chunk))
arr = np.frombuffer(chunk, dtype=np.uint8).copy()
tensor = torch.from_numpy(arr).float()
tensor = (tensor / 127.5) - 1.0
tensor = tensor.reshape(3, self.img_size, self.img_size)
return tensor, 0 # label placeholder
def get_text_loaders(batch_size=128, img_size=128,
train_size=100000, val_size=5000):
"""Text-as-image data loaders from WikiText-103."""
train_ds = TextAsImageDataset(size=train_size, img_size=img_size,
split='train')
val_ds = TextAsImageDataset(size=val_size, img_size=img_size,
split='validation')
train_loader = torch.utils.data.DataLoader(
train_ds, batch_size=batch_size, shuffle=True,
num_workers=2, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
val_ds, batch_size=batch_size, shuffle=False,
num_workers=2, pin_memory=True)
return train_loader, val_loader
# ── Zero-shot Test (existing model on text) ──────────────────────
def test_zero_shot(model_repo="AbstractPhil/svae-fresnel-128", device='cuda'):
"""Test an existing Fresnel/Johanna model on text without any text training."""
from transformers import AutoModel
print(f"\n{'='*70}")
print(f"ZERO-SHOT TEXT TEST: {model_repo}")
print(f"{'='*70}")
model = AutoModel.from_pretrained(model_repo, trust_remote_code=True).to(device).eval()
test_texts = [
"Hello, world! This is a test of the Fresnel geometric encoder.",
"The quick brown fox jumps over the lazy dog. 0123456789",
"import torch\nmodel = AutoModel.from_pretrained('test')\noutput = model(x)",
"To be, or not to be, that is the question.",
"E = mc² — Albert Einstein, 1905",
]
img_size = model.config.image_size
print(f" Image size: {img_size}×{img_size}")
print(f" Bytes per sample: {3 * img_size * img_size}")
for text in test_texts:
tensor, actual_len = text_to_tensor(text, img_size, img_size)
tensor = tensor.unsqueeze(0).to(device)
with torch.no_grad():
output = model(tensor)
recon = output["recon"]
mse = F.mse_loss(recon, tensor).item()
recovered = tensor_to_text(recon.squeeze(0).cpu(), actual_len)
# Byte accuracy
orig_bytes = ((tensor.squeeze(0).flatten() + 1.0) * 127.5).round().clamp(0, 255).byte()
recon_bytes = ((recon.squeeze(0).cpu().flatten() + 1.0) * 127.5).round().clamp(0, 255).byte()
exact_match = (orig_bytes[:actual_len] == recon_bytes[:actual_len]).float().mean().item()
print(f"\n Input: '{text[:80]}{'...' if len(text) > 80 else ''}'")
print(f" Output: '{recovered[:80]}{'...' if len(recovered) > 80 else ''}'")
print(f" MSE: {mse:.6f}")
print(f" Byte acc: {exact_match*100:.1f}%")
print(f" Exact: {'YES' if recovered.rstrip('\x00') == text else 'NO'}")
# ── Evaluation ───────────────────────────────────────────────────
def eval_text_reconstruction(model, test_loader, device='cuda', n_batches=10):
"""Evaluate byte-level reconstruction accuracy on text."""
model.eval()
total_mse, total_acc, n = 0, 0, 0
with torch.no_grad():
for i, (images, _) in enumerate(test_loader):
if i >= n_batches:
break
images = images.to(device)
out = model(images)
recon = out['recon']
mse = F.mse_loss(recon, images).item()
# Byte accuracy
orig_bytes = ((images.flatten(1) + 1.0) * 127.5).round().clamp(0, 255).long()
recon_bytes = ((recon.flatten(1) + 1.0) * 127.5).round().clamp(0, 255).long()
acc = (orig_bytes == recon_bytes).float().mean().item()
total_mse += mse * len(images)
total_acc += acc * len(images)
n += len(images)
return {
'mse': total_mse / n,
'byte_accuracy': total_acc / n,
'n_samples': n,
}
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--mode', default='test', choices=['test', 'eval'])
parser.add_argument('--model', default='AbstractPhil/svae-fresnel-128')
parser.add_argument('--device', default='cuda')
args = parser.parse_args()
if args.mode == 'test':
test_zero_shot(args.model, args.device) |