CRNN

This repository researched the influence of cnn and rnn on crnn.

data

The training and validation dateset is 364w Synthetic Chinese String Dataset and the dict contains 5990 chinese characters.

hardware

GPU: GTX 2080, 11G CPU: AMD Ryzen 7 3700X 8-Core Processor, 16G

cnn

cnn_name	epoch	accuracy	model_size
DefaultCNN_512	10	99.4%	45.5M
ResNet18_512	10	99.3%	68.0M
DenseNet_128	24	99.01%	21.0M
DenseNet18_256	10	99.11%	27.0M

rnn

rnn_name	epoch	accuracy	model_size
DenseNet18_256_doubleLSTM, layer=1, dropout= 0	10	99.11%	27.0M
DenseNet18_256_singleLSTM, layer=2, dropout= 0.8	10	96.17%	29.0M

models

the models trained in baidudisk passwd: ub4j