tridis_v2_HTR_historical_manuscripts

TrOCR model adapted to Handwritting Text Recognition on medieval manuscripts (12th-16th centuries)

TRIDIS (Tria Digita Scribunt) is a Handwriting Text Recognition model trained on semi-diplomatic transcriptions from medieval and Early Modern Manuscripts. It is suitable for work on documentary manuscripts, that is, manuscripts arising from legal, administrative, and memorial practices such as registers, feudal books, charters, proceedings, comptability more commonly from the Late Middle Ages (13th century and onwards). It can also show good performance on documents from other domains, such as literature books, scholarly treatises and cartularies providing a versatile tool for historians and philologists in transforming and analyzing historical texts.

A paper presenting the first version of the model is available here: Sergio Torres Aguilar, Vincent Jolivet. Handwritten Text Recognition for Documentary Medieval Manuscripts. Journal of Data Mining and Digital Humanities. 2023. https://hal.science/hal-03892163

A paper presenting the second version of the model (tris one) is available here: Sergio Torres Aguilar. Handwritten Text Recognition for Historical Documents using Visual Language Models and GANs. 2023. https://hal.science/hal-04716654

The training corpora can be download on Huggingface: https://doi.org/10.57967/hf/5001

Rules of transcription :

Main factor of semi-diplomatic edition is that abbreviations have been resolved:

• both those by suspension (<mark>facimꝰ</mark> ---> <mark>facimus</mark>) and by contraction (<mark>dñi</mark> --> <mark>domini</mark>).
• Likewise, those using conventional signs (<mark>⁊</mark> --> <mark>et</mark> ; <mark>ꝓ</mark> --> <mark>pro</mark>) have been resolved. 
• The named entities (names of persons, places and institutions) have been capitalized.
• The beginning of a block of text as well as the original capitals used by the scribe are also capitalized.
• The consonantal <mark>i</mark> and <mark>u</mark> characters have been transcribed as <mark>j</mark> and <mark>v</mark> in both French and Latin.
• The punctuation marks used in the manuscript like: <mark>.</mark> or <mark>/</mark> or <mark>|</mark> have not been systematically transcribed as the transcription has been standardized with modern punctuation.
• Corrections and words that appear cancelled in the manuscript have been transcribed surrounded by the sign <mark>$</mark> at the beginning and at the end.

Corpora

The model was trained on documents from the Late Medieval period (11th-16th centuries).

The training and evaluation ground-truth datasets involved 2950 pages, 245k lines of text, and almost 2.3M tokens, conducted using several freely available ground-truth corpora:

• The Alcar-HOME database: https://zenodo.org/record/5600884
• The e-NDP corpus: https://zenodo.org/record/7575693
• The Himanis project: https://zenodo.org/record/5535306
• Königsfelden Abbey corpus: https://zenodo.org/record/5179361
• CODEA
• Monumenta Luxemburgensia.

Addionally 400k synthetic lines were used to reinforce the pre-training phase of the encoder-decoder. These lines were generated using a GAN system (https://github.com/ganji15/HiGANplus) trained on medieval manuscripts pages.

Accuracy

TRIDIS was trained using a encode-decoder architecture based on a fine-tuned version of the TrOCR-large handwritten (microsoft/trocr-large-handwritten) and a RoBERTa modelized on medieval texts (magistermilitum/Roberta_Historical).

This final model operates in a multilingual environment (Latin, Old French, and Old Spanish) and is capable of recognizing several Latin script families (mostly Textualis and Cursiva) in documents produced circa 11th - 16th centuries.

During evaluation, the model showed an accuracy of 96.8% on the validation set and a CER (Character Error Ratio) of about 0.05 to 0.10 on three external unseen datasets and a WER of about 0.13 to 0.24 respectively, which is about 30% lower compared to CRNN+CTC solutions trained on the same corpora.

Other formats

A CRNN+CTC version of this model trained on Kraken 4.0 (https://github.com/mittagessen/kraken) using the same gold-standard and synsthetic annotation is available in Zenodo:

Torres Aguilar, S. (2024). TRIDIS v2 : HTR model for Multilingual Medieval and Early Modern Documentary Manuscripts (11th-16th) (Version 2). Zenodo. https://doi.org/10.5281/zenodo.13862096

Testing the Model

The following snippets can be used to get model inferences on manuscript lines.

1. Clone the model using: git lfs clone https://huggingface.co/magistermilitum/tridis_v2_HTR_historical_manuscripts

2. Here is how to test the model on one single image:

from transformers import TrOCRProcessor, AutoTokenizer, VisionEncoderDecoderModel
from safetensors.torch import load_file
import torch.nn as nn

from PIL import Image

# load image from the IAM database
path="/path/to/image/file.png"
image = Image.open(path).convert("RGB")

processor = TrOCRProcessor.from_pretrained("./tridis_v2_HTR_historical_manuscripts")
model = VisionEncoderDecoderModel.from_pretrained('microsoft/trocr-large-handwritten')

# Load the weights of this model
safetensors_path = "./tridis_v2_HTR_historical_manuscripts/model.safetensors" #load the weights from the downloaded model
state_dict = load_file(safetensors_path)

# Load the trocr model
model = VisionEncoderDecoderModel.from_pretrained("microsoft/trocr-large-handwritten")

#Modify the embeddings size and vocab
model.config.decoder.vocab_size = processor.tokenizer.vocab_size
model.config.vocab_size = model.config.decoder.vocab_size
model.decoder.output_projection = nn.Linear(1024, processor.tokenizer.vocab_size)
#model.decoder.model.decoder.embed_tokens = nn.Embedding(processor.tokenizer.vocab_size, 1024, padding_idx=1)
model.decoder.embed_tokens = nn.Embedding(processor.tokenizer.vocab_size, 1024, padding_idx=1)

# set beam search parameters
model.config.eos_token_id = processor.tokenizer.sep_token_id
model.config.max_length = 160
model.config.early_stopping = True
model.config.no_repeat_ngram_size = 3
model.config.length_penalty = 2.0
model.config.num_beams = 3

model.load_state_dict(state_dict)

pixel_values = processor(images=image, return_tensors="pt").pixel_values

generated_ids = model.generate(pixel_values)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(generated_text)

1. Here is how test the model on a dataset. Ideally the test dataset must be passed to the model on the form of a json list redirecting to the images:

for ex (graphical_line_path, line_text_content):

[   ["liber_eSc_line_b9f83857", "Et pour ces deniers que je ai ressus de"],   ["liber_eSc_line_8da10559", "lui , sui je ses hons et serai tant con je vive-"], etc. ]

import glob
import json, random
import multiprocessing
from tqdm import tqdm
import torchvision.transforms as transforms
from huggingface_hub import hf_hub_download

import string
import unicodedata
import editdistance
import numpy as np
import pandas as pd

def ocr_metrics(predicts, ground_truth, norm_accentuation=True, norm_punctuation=False):
    """Calculate Character Error Rate (CER), Word Error Rate (WER) and Sequence Error Rate (SER)"""

    if len(predicts) == 0 or len(ground_truth) == 0:
        return (1, 1, 1)

    cer, wer, ser = [], [], []

    for (pd, gt) in zip(predicts, ground_truth):
        pd, gt = pd.lower(), gt.lower()

        if norm_accentuation:
            pd = unicodedata.normalize("NFKD", pd).encode("ASCII", "ignore").decode("ASCII")
            gt = unicodedata.normalize("NFKD", gt).encode("ASCII", "ignore").decode("ASCII")
        if norm_punctuation:
            pd = pd.translate(str.maketrans("", "", string.punctuation))
            gt = gt.translate(str.maketrans("", "", string.punctuation))

        pd_cer, gt_cer = list(pd), list(gt)
        dist = editdistance.eval(pd_cer, gt_cer)
        cer.append(dist / (max(len(pd_cer), len(gt_cer))))

        pd_wer, gt_wer = pd.split(), gt.split()
        dist = editdistance.eval(pd_wer, gt_wer)
        wer.append(dist / (max(len(pd_wer), len(gt_wer))))

        pd_ser, gt_ser = [pd], [gt]
        dist = editdistance.eval(pd_ser, gt_ser)
        ser.append(dist / (max(len(pd_ser), len(gt_ser))))

    metrics = [cer, wer, ser]
    metrics = np.mean(metrics, axis=1)
    return metrics

def cleaning_output(text):
    import re
    clean_output = re.sub(r"[,.;]", "", text) #remove punctuation
    clean_output = re.sub(r"\s+", " ", clean_output) #remove extra spaces
    return clean_output

import torch
from torch.utils.data import Dataset
from PIL import Image

# Define the dataset class
class IAMDataset(Dataset): 
    def __init__(self, root_dir, df, processor, max_target_length=160):
        self.root_dir = root_dir
        self.df = df
        self.processor = processor
        self.max_target_length = max_target_length

    def __len__(self):
        return len(self.df)

    def __getitem__(self, idx):
        # get file name + text
        file_name = self.df['file_name'][idx]
        text = self.df['text'][idx]
        # prepare image (i.e. resize + normalize)
        image = Image.open(self.root_dir + file_name).convert("RGB")
        pixel_values = self.processor(image, return_tensors="pt").pixel_values
        # add labels (input_ids) by encoding the text
        labels = self.processor.tokenizer(text,
                                          padding="max_length",
                                          max_length=self.max_target_length).input_ids
        # important: make sure that PAD tokens are ignored by the loss function
        labels = [label if label != self.processor.tokenizer.pad_token_id else -100 for label in labels]
        # Include `file_name` to the results dict
        encoding = {"pixel_values": pixel_values.squeeze(), "labels": torch.tensor(labels), "file_name": file_name}
        return encoding

# Load the dataset
from transformers import TrOCRProcessor, AutoTokenizer

#Load the processor from the model
processor = TrOCRProcessor.from_pretrained("./tridis_v2_HTR_historical_manuscripts") #load the processor from the downloaded model

# Define the dataset
#Open the file with text lines
with open('/your/lines/file.json', encoding='utf-8') as fh:
    transcriptions = json.load(fh)
    random.shuffle(transcriptions)
    transcriptions=list(filter(lambda x: x is not None, transcriptions))
    transcriptions = [[x[0]+".png", x[1]] for x in transcriptions if (len(x[1])>3 and len(x[1])<201 and type(x[1])==str)] #filter by length (optional) with *.png by default
    print(len(transcriptions))
    df = pd.DataFrame(transcriptions, columns=["file_name", "text"])
    print(df.head())
    print(sum([len(x[1]) for x in transcriptions]))

# Open the file with the images lines
test_dataset = IAMDataset(root_dir='/your/images/folder/',
                           df=df,
                           processor=processor)
print("Number of test examples:", len(test_dataset))


# Load the test dataloader
from torch.utils.data import DataLoader
import torch.nn as nn

test_dataloader = DataLoader(test_dataset, batch_size=16) #adapt batch size to your GPU
batch = next(iter(test_dataloader))
labels = batch["labels"]
labels[labels == -100] = processor.tokenizer.pad_token_id
label_str = processor.batch_decode(labels, skip_special_tokens=True)
label_str

# Load the model
from transformers import VisionEncoderDecoderModel, AutoModelForCausalLM
import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

from safetensors.torch import load_file
# Load the weights of this model
safetensors_path = "./tridis_v2_HTR_historical_manuscripts/model.safetensors" #load the weights from the downloaded model
state_dict = load_file(safetensors_path)

# Load the trocr model
model = VisionEncoderDecoderModel.from_pretrained("microsoft/trocr-large-handwritten")

# set special tokens used for creating the decoder_input_ids from the labels
model.config.decoder_start_token_id = processor.tokenizer.cls_token_id
model.config.pad_token_id = processor.tokenizer.pad_token_id
# make sure vocab size is set correctly
model.config.vocab_size = model.config.decoder.vocab_size

#Configure the embeddings size
model.config.decoder.vocab_size = processor.tokenizer.vocab_size
model.config.vocab_size = model.config.decoder.vocab_size
model.decoder.output_projection = nn.Linear(1024, processor.tokenizer.vocab_size)
#model.decoder.model.decoder.embed_tokens = nn.Embedding(processor.tokenizer.vocab_size, 1024, padding_idx=1)
model.decoder.embed_tokens = nn.Embedding(processor.tokenizer.vocab_size, 1024, padding_idx=1)

#Useful Hyper-parameters (optional)
model.config.decoder.activation_function="gelu"
model.config.decoder.layernorm_embedding=True
model.config.decoder.max_position_embeddings=514
model.config.decoder.scale_embedding=False
model.config.decoder.use_learned_position_embeddings=True

# set beam search parameters
model.config.eos_token_id = processor.tokenizer.sep_token_id
model.config.max_length = 160
model.config.early_stopping = True
model.config.no_repeat_ngram_size = 3
model.config.length_penalty = 2.0
model.config.num_beams = 3

# update the model weights
model.load_state_dict(state_dict)
model.to(device)

# Load the metrics
from datasets import load_metric
bert= load_metric("bertscore")

# Evaluate the model
print("Running evaluation...")

dictionary=[]
for batch in tqdm(test_dataloader):
    pixel_values = batch["pixel_values"].to(device)
    outputs = model.generate(pixel_values)
    # Decoding predictions and references
    pred_str = processor.batch_decode(outputs, skip_special_tokens=True)
    labels = batch["labels"]
    labels[labels == -100] = processor.tokenizer.pad_token_id
    label_str = processor.batch_decode(labels, skip_special_tokens=True)
    file_names = batch["file_name"]  # Assert that DataLoader includes `file_name` in each batch
    dictionary.extend([[file_name, pred, ref] for file_name, pred, ref in zip(file_names, pred_str, label_str)])

# Save results as a dictionary
with open("/your/save/path/dictionary_of_results.json", "w", encoding='utf-8') as jsonfile:
    json.dump(dictionary, jsonfile, ensure_ascii=False, indent=1)

#compute the BERT score
bert_score=bert.compute(references=[x[1] for x in dictionary], predictions=[x[2] for x in dictionary], model_type="bert-base-multilingual-cased")
bert_score_mean=np.mean(bert_score["f1"])
bert_score_std=np.std(bert_score["f1"])

# Print the results according to the metrics
print("BERT_SCORE_MEAN : ", bert_score_mean, "BERT_SCORE_STD : ", bert_score_std )
print("RAW   metrics : ", ocr_metrics([x[1] for x in dictionary], [x[2] for x in dictionary]))
print("CLEAN metrics : ", ocr_metrics([cleaning_output(x[1]) for x in dictionary], [cleaning_output(x[2]) for x in dictionary]))
print(*dictionary[1:], sep="\n\n")

• Developed by: [Sergio Torres Aguilar]
• Model type: [TrOCR]
• Language(s) (NLP): [Medieval Latin, Spanish, French, Middle German]
• Finetuned from model [optional]: [Handwritten Text Recognition]