Skip to content

Latest commit

 

History

History
92 lines (63 loc) · 5.25 KB

README.md

File metadata and controls

92 lines (63 loc) · 5.25 KB
Raw

SoftCTC – Semi-Supervised Learning for Text Recognition using Soft Pseudo-Labels

This repository contains source codes for SoftCTC paper. In the soft_ctc subdirectory you can find:

  • models/
    • connections.py - Model for storing transcription variants of single text line
    • batch_connections.py - Model for batched transcription variants
  • libs/ - Compiled libraries for CUDA and OpenCL
  • equations.py - Implementation of SoftCTC loss equations as proposed in the paper
  • multi_ctc_loss.py - Parallel MultiCTC loss implementation based on PyTorch CTC
  • soft_ctc_loss.py - SoftCTC loss implementation in PyTorch
  • soft_ctc_loss_cuda.py - SoftCTC loss adapter for CUDA implementation
  • soft_ctc_loss_opencl.py - SoftCTC loss adapter for OpenCL implementation

CUDA and OpenCL

Current CUDA and OpenCL implementations are compiled for Python 3.7-11 and PyTorch 1.13.1. In case you need it for a different configuration, you can compile it on your own from the source codes in soft_ctc_gpu_lib or SoftCTC-GPU using the included instructions.

Paper abstract

This paper explores semi-supervised training for sequence tasks, such as Optical Character Recognition or Automatic Speech Recognition. We propose a novel loss function – SoftCTC – which is an extension of CTC allowing to consider multiple transcription variants at the same time. This allows to omit the confidence based filtering step which is otherwise a crucial component of pseudo-labeling approaches to semi-supervised learning. We demonstrate the effectiveness of our method on a challenging handwriting recognition task and conclude that SoftCTC matches the performance of a finely-tuned filtering based pipeline. We also evaluated SoftCTC in terms of computational efficiency, concluding that it is significantly more efficient than a naïve CTC-based approach for training on multiple transcription variants, and we make our GPU implementation public.

Please cite

If you use SoftCTC, please cite:

Kišš, M., Hradiš, M., Beneš, K., Buchal, P., & Kula, M. (2022). SoftCTC – Semi-Supervised Learning for Text Recognition using Soft Pseudo-Labels. arXiv preprint arXiv:2212.02135.

Usage

Import

from soft_ctc.soft_ctc_loss import SoftCTCLoss

Convert confusion network to Connections model

You can use convert_confusion_network_to_connections from soft_ctc.models.connections to convert confusion network (a list of dictionaries with characters as keys and probabilities as values) into a transition matrix (Connections model) and labels (ordered sequence of keys from the confusion network).

# confusion_network = [{'C': 0.5, None: 0.5}, {'A': 0.7, 'U': 0.2, None: 0.1}, {'T': 1.0}, {'E': 0.6, 'S': 0.4}]
# character_set = ["<BLANK>"] + list("ACESTU")

from soft_ctc.models.connections import convert_confusion_network_to_connections, convert_characters_to_labels

blank = character_set.index("<BLANK>")
confusion_network = convert_characters_to_labels(confusion_network, character_set)
labels, connections = convert_confusion_network_to_connections(confusion_network, blank)

Stack Connections and labels into BatchConnections and batched labels

To stack multiple transition matrices and labels into a batch structures, you can use BatchConnections and stack_labels from soft_ctc.models.batch_connections

# all_connections: list of Connections in batch
# all_labels: list of labels in batch

from soft_ctc.models.batch_connections import BatchConnections, stack_labels, calculate_target_size

target_size = calculate_target_size([connections.size() for connections in all_connections])
batch_connections = BatchConnections.stack_connections(all_connections, target_size)
batch_labels = stack_labels(all_labels, blank, target_size)

Calculate the SoftCTC loss

The initialization os the SoftCTCLoss has two optional parameters: norm_step and zero_infinity. The norm_step specifies how often the normalization is performed to prevent underflow (i.e. norm_step=10 means that the normalization is done every 10-th frame in logits). If the zero_infinity flag is set, the individual losses which are equal to infinity are zeroed to prevent the training process from collapsing.

# logits: raw network output of shape (N,C,T) where N is batch size, C is number of characters including blank, and T is 
# the number of frames (width) in logits

softctc = SoftCTCLoss(norm_step=10, zero_infinity=True)
loss = softctc(logits, batch_connections, batch_labels)

loss.mean().backward()

Numerical stability

Currently, the SoftCTC loss is calculated in probabilities and NOT log-probabilities. Due to this fact, the calculation may be mathematically unstable (underflow may occur), especially when the logits contain a very flat probability distribution at the beginning of the training process. Options to eliminate this are to use a smaller normalization step (norm_step), or to calculate in double precision.