rajesh-1902/greek-htr-crnn-ctc-spacefixweighted

Greek Sentence-Level HTR — CRNN + CTC

A sentence-level Handwritten Text Recognition model for Greek polytonic script, trained on rithwikn/greek_combined_dataset.

Architecture

Component	Details
CNN	5-block (64→128→256→256→512 channels, BatchNorm + ReLU + MaxPool)
RNN	2× Bidirectional LSTM (512 units each, 30% dropout between layers)
Decoder	CTC (BestPath greedy / BeamSearch width-100)
Input size	1024 × 64 px (grayscale)
Max text len	128 characters
Batch size	8

Space-Awareness Improvements

• Space-weighted CTC loss — samples with more words get higher gradient weight (SPACE_WEIGHT=3.0)
• Trainable space logit bias — a learned additive offset on the space class logit (warm-started at +0.5)
• Space-dense oversampling — training samples with ≥6 words are duplicated to increase exposure to word boundaries

Usage — Restore from Checkpoint

from huggingface_hub import hf_hub_download, list_repo_files

REPO_ID   = "rajesh-1902/greek-htr-crnn-ctc-spacefixweighted"
LOCAL_DIR = "./model_sentence"

# Download all files from latest/
for f in list_repo_files(REPO_ID, repo_type="model"):
    if f.startswith("latest/"):
        hf_hub_download(REPO_ID, filename=f, repo_type="model", local_dir=LOCAL_DIR)

Then run the Colab notebook — it will auto-restore from the checkpoint.

Training Details

• Optimiser: Adam, lr=1e-4
• Early stopping: 10 epochs no CER improvement
• Data augmentation: rotation ±2°, brightness jitter, Gaussian blur, horizontal stretch
• Framework: TensorFlow 2.13 (TF1 graph mode via tf.compat.v1)