magles/parseq-malayalam-v3-all-checkpoints

PARSeq Malayalam OCR checkpoints

This repository contains multiple PARSeq checkpoint versions for Malayalam OCR experiments.

Files

• parseq_finetuned_v3_rl_kth_v2.pth — best RL+kth v2 fine-tuned checkpoint
• parseq_finetuned_v3_rl_kth_v2_errordriven.pth — error-driven polished checkpoint
• parseq_v3_rl_kth_v2_latest.pth — latest RL+kth v2 checkpoint
• parseq_finetuned_v3_rl_kth.pth — earlier fine-tuned RL+kth checkpoint
• parseq_v3_rl_kth_latest.pth — earlier RL+kth latest checkpoint
• parseq_v3_rl_latest.pth — RL latest checkpoint

Notes

• Vocabulary size: 99
• Framework: PyTorch
• Task: Malayalam scene text recognition
• Architecture: PARSeq-based OCR model

Metrics

• RL-kth-v2 full test: 76.58 word acc / 85.98 char acc
• RL-kth-v2 chillu-only: 70.00 word acc / 83.45 char acc
• Error-driven full test: 76.90 word acc / 86.16 char acc

Usage

import torch

ckpt = torch.load("parseq_finetuned_v3_rl_kth_v2.pth", map_location="cpu")
print(ckpt.keys())

Available checkpoint metadata

Each checkpoint may include:

• model
• epoch
• word_acc
• char_acc
• char2idx
• idx2char