JaidedAI
Recognizing custom model result is bad compared to text predicted during training for Greek Language
Hello All
I am trying to train custom model training for Greek Language
please find my config.yaml file
number: '0123456789' symbol: "!"#$%&'()*+,-./:;<=>?@[\]^_`{|}~ €" lang_char: 'ΨτκφΦηΔΟΖΚθΘβδΧαωΞουλΠψςνΡγρΥΜεμΤΛΙσξΗΩΒΓΣΑιχΕπζΝ' experiment_name: 'greek' train_data: './all_data' valid_data: './all_data/Greek_val' manualSeed: 1111 workers: 6 batch_size: 32 #32 num_iter: 300000 valInterval: 100 saved_model: '' #'saved_models/en_filtered/iter_300000.pth' FT: False optim: False # default is Adadelta lr: 1.0 beta1: 0.9 rho: 0.95 eps: 0.00000001 grad_clip: 5 #Data processing select_data: 'Greek_train' # this is dataset folder in train_data batch_ratio: '1' total_data_usage_ratio: 1.0 batch_max_length: 128 imgH: 64 imgW: 600 rgb: False contrast_adjust: False sensitive: True PAD: True contrast_adjust: 0.0 data_filtering_off: False
Model Architecture
Transformation: 'None' FeatureExtraction: 'VGG' SequenceModeling: 'BiLSTM' Prediction: 'CTC' num_fiducial: 20 input_channel: 1 output_channel: 256 hidden_size: 256 decode: 'greedy' new_prediction: True freeze_FeatureFxtraction: False freeze_SequenceModeling: False
Generated opt.txt file is below
Model Architecture is below
Filtering the images containing characters which are not in opt.character Filtering the images whose label is longer than opt.batch_max_length
dataset_root: ./all_data opt.select_data: ['Greek_train'] opt.batch_ratio: ['1']
dataset_root: ./all_data dataset: Greek_train ./all_data/Greek_train sub-directory: /Greek_train num samples: 295 num total samples of Greek_train: 295 x 1.0 (total_data_usage_ratio) = 295 num samples of Greek_train per batch: 32 x 1.0 (batch_ratio) = 32
Total_batch_size: 32 = 32
dataset_root: ./all_data/Greek_val dataset: / ./all_data/Greek_val/ sub-directory: /. num samples: 264
No Transformation module specified model input parameters 64 600 20 1 256 256 94 128 None VGG BiLSTM CTC Model: DataParallel( (module): Model( (FeatureExtraction): VGG_FeatureExtractor( (ConvNet): Sequential( (0): Conv2d(1, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): ReLU(inplace=True) (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (3): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (4): ReLU(inplace=True) (5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (6): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (7): ReLU(inplace=True) (8): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (9): ReLU(inplace=True) (10): MaxPool2d(kernel_size=(2, 1), stride=(2, 1), padding=0, dilation=1, ceil_mode=False) (11): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (12): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (13): ReLU(inplace=True) (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (15): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (16): ReLU(inplace=True) (17): MaxPool2d(kernel_size=(2, 1), stride=(2, 1), padding=0, dilation=1, ceil_mode=False) (18): Conv2d(256, 256, kernel_size=(2, 2), stride=(1, 1)) (19): ReLU(inplace=True) ) ) (AdaptiveAvgPool): AdaptiveAvgPool2d(output_size=(None, 1)) (SequenceModeling): Sequential( (0): BidirectionalLSTM( (rnn): LSTM(256, 256, batch_first=True, bidirectional=True) (linear): Linear(in_features=512, out_features=256, bias=True) ) (1): BidirectionalLSTM( (rnn): LSTM(256, 256, batch_first=True, bidirectional=True) (linear): Linear(in_features=512, out_features=256, bias=True) ) ) (Prediction): Linear(in_features=256, out_features=94, bias=True) ) ) Modules, Parameters module.FeatureExtraction.ConvNet.0.weight 288 module.FeatureExtraction.ConvNet.0.bias 32 module.FeatureExtraction.ConvNet.3.weight 18432 module.FeatureExtraction.ConvNet.3.bias 64 module.FeatureExtraction.ConvNet.6.weight 73728 module.FeatureExtraction.ConvNet.6.bias 128 module.FeatureExtraction.ConvNet.8.weight 147456 module.FeatureExtraction.ConvNet.8.bias 128 module.FeatureExtraction.ConvNet.11.weight 294912 module.FeatureExtraction.ConvNet.12.weight 256 module.FeatureExtraction.ConvNet.12.bias 256 module.FeatureExtraction.ConvNet.14.weight 589824 module.FeatureExtraction.ConvNet.15.weight 256 module.FeatureExtraction.ConvNet.15.bias 256 module.FeatureExtraction.ConvNet.18.weight 262144 module.FeatureExtraction.ConvNet.18.bias 256 module.SequenceModeling.0.rnn.weight_ih_l0 262144 module.SequenceModeling.0.rnn.weight_hh_l0 262144 module.SequenceModeling.0.rnn.bias_ih_l0 1024 module.SequenceModeling.0.rnn.bias_hh_l0 1024 module.SequenceModeling.0.rnn.weight_ih_l0_reverse 262144 module.SequenceModeling.0.rnn.weight_hh_l0_reverse 262144 module.SequenceModeling.0.rnn.bias_ih_l0_reverse 1024 module.SequenceModeling.0.rnn.bias_hh_l0_reverse 1024 module.SequenceModeling.0.linear.weight 131072 module.SequenceModeling.0.linear.bias 256 module.SequenceModeling.1.rnn.weight_ih_l0 262144 module.SequenceModeling.1.rnn.weight_hh_l0 262144 module.SequenceModeling.1.rnn.bias_ih_l0 1024 module.SequenceModeling.1.rnn.bias_hh_l0 1024 module.SequenceModeling.1.rnn.weight_ih_l0_reverse 262144 module.SequenceModeling.1.rnn.weight_hh_l0_reverse 262144 module.SequenceModeling.1.rnn.bias_ih_l0_reverse 1024 module.SequenceModeling.1.rnn.bias_hh_l0_reverse 1024 module.SequenceModeling.1.linear.weight 131072 module.SequenceModeling.1.linear.bias 256 module.Prediction.weight 24064 module.Prediction.bias 94 Total Trainable Params: 3780574 Trainable params num : 3780574 Optimizer: Adadelta ( Parameter Group 0 eps: 1e-08 foreach: None lr: 1.0 maximize: False rho: 0.95 weight_decay: 0 )
As per https://github.com/JaidedAI/EasyOCR/blob/master/custom_model.md copied best _accuracy.pth file and renamed as greek.pth as per custom_example.zip renamed custom_example.py to greek.py and modified greek.yaml as below
moved all greek.py and greek.yaml file to /root/.EasyOCR/user_network/ and greek.pth file to /root/.EasyOCR/model this character list i have added from opt.txt file character parameter
created greek_char.txt file in easyocr/character (python dist-packages) as below
Ψ τ κ φ Φ η Δ Ο Ζ Κ θ Θ β δ Χ α ω Ξ ο υ λ Π ψ ς ν Ρ γ ρ Υ Μ ε μ Τ Λ Ι σ ξ Η Ω Β Γ Σ Α ι χ Ε π ζ Ν
Please find my code for using custom model
import easyocr reader=easyocr.Reader(lang_list=['greek'],recog_network='greek') result=reader.readtext('0.png',detail=0) for x in result : print(x)
0.png is Εμφ and my output is NH
Please let me know how to fix this issue
I am facing the same kind of issue. I am also training the custom model. I kept the settings pretty much the same. Only Added the dict file and characters. But I am facing the same issue. During training accuracy is 85+ but as I used that custom model it is giving very bad results.
@vamshi-1611 @bh1nv have you found the cause or solution? I also have the same problem.
