Upload __init__.py with huggingface_hub
Browse files- __init__.py +130 -0
__init__.py
ADDED
|
@@ -0,0 +1,130 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
SocrateX - Custom OCR library based on Transformer architecture.
|
| 3 |
+
|
| 4 |
+
This library was developed to facilitate training, inference,
|
| 5 |
+
and experimentation with SOCRATE models. Everything is modular and customizable.
|
| 6 |
+
|
| 7 |
+
Quick Start:
|
| 8 |
+
-----------------------------------
|
| 9 |
+
import SocrateX as sx
|
| 10 |
+
|
| 11 |
+
# 1. Build a custom model architecture using sx.Config:
|
| 12 |
+
config = sx.Config(
|
| 13 |
+
d_model=640,
|
| 14 |
+
nhead=10,
|
| 15 |
+
num_layers=12,
|
| 16 |
+
dim_feedforward=2560,
|
| 17 |
+
pool_height=4 # nn.AdaptiveMaxPool2d((pool_height, None))
|
| 18 |
+
)
|
| 19 |
+
tokenizer = sx.load_tokenizer("ocr_bpe_tokenizer.json")
|
| 20 |
+
model = sx.init(config=config, tokenizer=tokenizer)
|
| 21 |
+
|
| 22 |
+
# 2. Build a dataset (height and max_length go here, not in Config):
|
| 23 |
+
train_set = model.make_dataset(images, labels, height=32, max_length=64)
|
| 24 |
+
|
| 25 |
+
# 3. Train:
|
| 26 |
+
model.fit(dataloader, optimizer, criterion, epochs=50)
|
| 27 |
+
|
| 28 |
+
# 4. Predict (inference params go here):
|
| 29 |
+
results = model.predict(
|
| 30 |
+
image_paths=["document.jpg"],
|
| 31 |
+
function="generate",
|
| 32 |
+
temp=0.5,
|
| 33 |
+
max_iter=64,
|
| 34 |
+
penalty=1.15,
|
| 35 |
+
top_k=5
|
| 36 |
+
)
|
| 37 |
+
"""
|
| 38 |
+
|
| 39 |
+
from .configuration_socrate import SocrateConfig
|
| 40 |
+
from .config import Config
|
| 41 |
+
from .model import SOCRATE, cat, rat, mice, ResidualBlock, PositionalEncoding, SocratePool
|
| 42 |
+
from .dataset import Makeset, SmartBatchSampler, load_dataset
|
| 43 |
+
from .trainer import train, Trainer
|
| 44 |
+
from .inference import predict, generate, generate_fast, beam_search, extract_crops_from_image
|
| 45 |
+
from .tokenizer import init_tokenizer, SocrateXTokenizer
|
| 46 |
+
from .synthetic import generate_silly_training_set, generate_silly_testing_set
|
| 47 |
+
|
| 48 |
+
__all__ = [
|
| 49 |
+
"Config",
|
| 50 |
+
"SOCRATE",
|
| 51 |
+
"SocrateConfig",
|
| 52 |
+
"cat",
|
| 53 |
+
"rat",
|
| 54 |
+
"mice",
|
| 55 |
+
"ResidualBlock",
|
| 56 |
+
"PositionalEncoding",
|
| 57 |
+
"SocratePool",
|
| 58 |
+
"Makeset",
|
| 59 |
+
"SmartBatchSampler",
|
| 60 |
+
"train",
|
| 61 |
+
"Trainer",
|
| 62 |
+
"predict",
|
| 63 |
+
"generate",
|
| 64 |
+
"generate_fast",
|
| 65 |
+
"beam_search",
|
| 66 |
+
"init_tokenizer",
|
| 67 |
+
"load_tokenizer",
|
| 68 |
+
"SocrateXTokenizer",
|
| 69 |
+
"generate_silly_training_set",
|
| 70 |
+
"generate_silly_testing_set",
|
| 71 |
+
"load_dataset",
|
| 72 |
+
]
|
| 73 |
+
|
| 74 |
+
def load_tokenizer(path="ocr_bpe_tokenizer.json"):
|
| 75 |
+
"""
|
| 76 |
+
Alias to easily load a tokenizer from a JSON file.
|
| 77 |
+
"""
|
| 78 |
+
return SocrateXTokenizer.from_file(path)
|
| 79 |
+
|
| 80 |
+
def load(model_type="cat", weights=None, tokenizer_path="ocr_bpe_tokenizer.json", device="cuda"):
|
| 81 |
+
"""
|
| 82 |
+
Automatically loads the desired model and tokenizer.
|
| 83 |
+
Returns (model, tokenizer).
|
| 84 |
+
"""
|
| 85 |
+
from tokenizers import Tokenizer
|
| 86 |
+
tokenizer = Tokenizer.from_file(tokenizer_path)
|
| 87 |
+
|
| 88 |
+
if model_type == "cat":
|
| 89 |
+
model = cat(tokenizer=tokenizer, weights=weights if weights else cat.pretrained, device=device)
|
| 90 |
+
elif model_type == "rat":
|
| 91 |
+
model = rat(tokenizer=tokenizer, weights=weights if weights else rat.pretrained, device=device)
|
| 92 |
+
elif model_type == "mice":
|
| 93 |
+
model = mice(tokenizer=tokenizer, weights=weights if weights else mice.pretrained, device=device)
|
| 94 |
+
else:
|
| 95 |
+
raise ValueError(f"Unknown model type: {model_type}")
|
| 96 |
+
|
| 97 |
+
return model, tokenizer
|
| 98 |
+
|
| 99 |
+
def init(tokenizer=None, config=None, device="cuda"):
|
| 100 |
+
"""
|
| 101 |
+
Initializes a SOCRATE model from scratch.
|
| 102 |
+
|
| 103 |
+
Pass an sx.Config() object to fully control the architecture:
|
| 104 |
+
config = sx.Config(d_model=256, nhead=4, num_layers=3, pool_height=4)
|
| 105 |
+
model = sx.init(config=config, tokenizer=tokenizer)
|
| 106 |
+
|
| 107 |
+
If config is None, defaults to the cat (158M) architecture.
|
| 108 |
+
"""
|
| 109 |
+
if tokenizer is None:
|
| 110 |
+
raise ValueError("You must provide a tokenizer (sx.init_tokenizer() or sx.load_tokenizer()).")
|
| 111 |
+
|
| 112 |
+
if config is None:
|
| 113 |
+
# Default: cat architecture
|
| 114 |
+
config = Config()
|
| 115 |
+
|
| 116 |
+
hf_config = SocrateConfig(
|
| 117 |
+
d_model=config.d_model,
|
| 118 |
+
max_len=config.max_len,
|
| 119 |
+
nhead=config.nhead,
|
| 120 |
+
dim_feedforward=config.dim_feedforward,
|
| 121 |
+
activation=config.activation,
|
| 122 |
+
norm_first=config.norm_first,
|
| 123 |
+
num_layers=config.num_layers,
|
| 124 |
+
vocab_size=tokenizer.get_vocab_size(),
|
| 125 |
+
pad_id=tokenizer.token_to_id("<pad>"),
|
| 126 |
+
bos_id=tokenizer.token_to_id("<bos>"),
|
| 127 |
+
eos_id=tokenizer.token_to_id("<eos>"),
|
| 128 |
+
)
|
| 129 |
+
model = SOCRATE(hf_config, tokenizer=tokenizer, sx_config=config).to(device)
|
| 130 |
+
return model
|