SVTR Text Recognition: Zero Accuracy Issue & Need Guidance for Custom Pipeline

[btwitsayush]

Description

I'm new to PaddleOCR and currently building a custom text recognition pipeline for address labels and tracking numbers. While implementing this, I encountered multiple issues—most critically, the SVTR model always trains to 0% accuracy. I'm seeking help with:

• Resolving these training/export errors
• Getting guidance to correctly set up a custom text recognition pipeline

Current Issues

1. SVTR model trains but stays at 0% accuracy
2. CRNN model attempts lead to segmentation faults
3. Exporting fails with SVTRRecResizeImg error

Environment

• PaddlePaddle version: 2.6.2
• Running on: Google Colab (Linux)
• GPU: Available, CUDA enabled

Dataset Details

• Total images: 50
• Training: 45
• Testing: 5
• Image format: BGR, 3x32x320
• Unique characters: 63
• Content: Address labels and tracking numbers (data is encrypted/masked for privacy)

Data Formats:

- Address Labels:

Name XXXXX Address XXXXX Ltd.
Floor XX, Building XXX
City - XXXXXX
Phone XXXXXXXXXX Email Pincode XXXXXX

Tracking Numbers:

XXXXXXXXXXXXX (12–15 digit numbers)

Model Configuration (YAML)

Architecture:
model_type: rec
algorithm: SVTR
Backbone:
name: MobileNetV1Enhance
scale: 0.5
Neck:
name: SequenceEncoder
encoder_type: svtr
dims: 64
depth: 2
hidden_dims: 120
use_guide: True
Head:
name: CTCHead
fc_decay: 0.00001

Training Results

• Epochs: 20
• Best Epoch: 19
• Accuracy: 0.0
• Normalized Edit Distance: 0.06458
• Loss: From 32.9 to 31.0
• Learning Rate: 5e-6
• Pretrained Model: ch_PP-OCRv3_rec_train
• Batch Size: 4
• Max Text Length: 500

Questions

1. Why is accuracy stuck at 0?
2. Is learning rate 5e-6 too low for fine-tuning?
3. How to fix the SVTRRecResizeImg export error?
4. What is the correct pipeline for training with a small, custom dataset?

Pipeline Help Needed

Please guide me on:

• Preparing a custom dataset properly (format, label structure, etc.)
• Generating and validating the dict.txt (character dictionary)
• Adjusting training config for small datasets
• Full steps to set up a working custom text recognition pipeline
• Choosing the right architecture (SVTR, CRNN, etc.) for small & large datasets with limited labeled images

Additional Notes

• Dataset contains sensitive personal information, hence encrypted.
• Custom pipeline is for logistics use cases (address & tracking number recognition).
• Willing to contribute findings back to the community.

Request for Assistance

1. Help me fix the SVTR training and export issues
2. Share a working config or example for small-scale text recognition training
3. Recommend best practices for datasets with private content
4. Guide me in building a robust custom recognition pipeline

/label ~bug ~text-recognition ~training ~help-wanted ~documentation
/cc @PaddleOCR-maintainers