Models

Models

Available Variants

Model	Params	FP16 Size	Accuracy (IIIT5k)	GPU Latency
FlashOCR-m-0.5x	0.6M	~1.5 MB	—	1.2 ms
FlashOCR-m	1.5M	~3.0 MB	85.2%	2.1 ms
FlashOCR-m-1.5x	3.2M	~6.5 MB	88.7%	3.4 ms

Architecture

• Backbone: ShuffleNetV2 (width multiplier configurable: 0.5x, 1.0x, 1.5x)
• Encoder: CNN with adaptive pooling (height → 1)
• Decoder: BiLSTM + CTC (default) or GRU-based Attention
• Loss: CTC Loss (connectionist temporal classification)

Decoder Types

Decoder	Description	Best For
CTC	BiLSTM + CTC loss	Fast inference, fixed-length outputs
Attention	GRU + Bahdanau attention	Variable-length, complex scripts

Pretrained Weights

Pretrained weights are auto-downloaded on first use:

trainer = Trainer(model_size="m", pretrained=True)