Utils(2).py

import torch
from torch.utils.data import Dataset,DataLoader
from PIL import Image

import os 
import cv2

import tensorflow as tf
import numpy as np
from tensorflow import keras


############ extractdetection  ################


#alpha = "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
alpha =['A', 'A1', 'A2' ,'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L','M', 
      'N','O', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z',
      'Z1', 'Z2','Z3','Z4','Z5']


############    decode_detection     ################

# labels from KHATT dataset
"""characters = [' ', '!', '"', '#', '%', '(', ')', '*', '+', ',', '-', '.', '/', '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', ':', ';', '=', '>', '?', '[', '\\', ']', 'x', '\xa0', '×', '،', '؛', '؟', 'ء', 'آ', 'أ', 'ؤ', 'إ', 'ئ', 'ا', 'ب', 'ة', 'ت', 'ث', 'ج', 'ح', 'خ', 'د', 'ذ', 'ر', 'ز', 'س', 'ش', 'ص', 'ض', 'ط', 'ظ', 'ع', 'غ', 'ـ', 'ف', 'ق', 'ك', 'ل', 'م', 'ن', 'ه', 'و', 'ى', 'ي', 'ً', 'ٌ', 'ٍ', 'َ', 'ُ', 'ِ', 'ّ', 'ْ', '–', '‘']
characters.sort()"""


# labels from old arabic
characters = [' ', '.', '[', ']', '؟', 'ء', 'آ', 'أ', 'ؤ', 'إ', 'ئ', 'ا', 'ب', 'ة', 'ت', 'ث', 'ج', 'ح', 'خ', 'د', 'ذ', 'ر', 'ز', 'س', 'ش', 'ص', 'ض', 'ط', 'ظ', 'ع', 'غ', 'ـ', 'ف', 'ق', 'ك', 'ل', 'م', 'ن', 'ه', 'و', 'ى', 'ي', 'ً', 'ٌ', 'ٍ', 'َ', 'ُ', 'ِ', 'ّ', 'ْ', 'ٔ', 'ٕ', '١', '٢', '٣', '٤', '٥', '٧', '٨', 'ٮ', 'ٯ', 'ٰ', 'ڡ', 'ک', 'ں', 'ی', '۴', '\u202c', 'ﭐ', 'ﺟ', 'ﺣ', 'ﻛ', '�']
characters.sort()

characters.sort()
max_length = 132
img_height, img_width = 1056,64

def label_to_num(label,max_length=max_length):
    label_num = []
    for ch in label:        
        try:
            label_num.append(characters.index(ch))
        except:
            pass

    return   keras.utils.pad_sequences( [label_num], maxlen=max_length, dtype='int32',
                                       padding='post',truncating='pre',value=len(characters)+2)[0]


def num_to_label(num):
    #if  isinstance(num,torch.Tensor):
       #num=num.to(torch.int8)
    ret = ""
    for ch in num:
        if int(ch)==-1 or ch ==len(characters)+2 :# or ch==-1 : # pad symbole
            break
        try:
            ret+=characters[int(ch)]
        except:
            pass
    return ret



def decode_predictions(pred,greedy = True):
    input_len = np.ones(pred.shape[0]) * pred.shape[1]
    results = keras.backend.ctc_decode(pred, input_length=input_len, greedy=greedy)[0][0][
        :, :max_length ]
    output_text = []
    for res in results:
        #decoded = tokenizer.sequences_to_texts([res.numpy()])
        #decoded = list(map(num_to_label,results))
        decoded = num_to_label( res)
        output_text.append(decoded)
    return output_text


############ dataloader ################


 
# def encode_single_sample(path_dir,label = None):
#     img = tf.io.read_file(path_dir)
#     img = tf.io.decode_jpeg( img,  name=None)  
#     img.set_shape([img.shape[0], img.shape[1],img.shape[-1]])
#     img = tf.image.rot90(img, k=1, name=None)
#     img = tf.image.resize(img, [img_height, img_width])
#     img=img/255.0
#     #img=tf.concat([img,img,img],axis=-1)

#     return img 

def encode_single_sample(path_dir, label=None):
    img = tf.io.read_file(path_dir)
    img = tf.io.decode_jpeg(img, name=None)
    img.set_shape([img.shape[0], img.shape[1], img.shape[-1]])
    img = tf.image.rot90(img, k=1, name=None)
    img = tf.image.resize(img, [img_height, img_width])
    img = tf.image.rgb_to_grayscale(img)  # Convert image to grayscale
    img = img/255.0
    return img
    

batch_size = 16 

def Loadlines(path_lines):
    path_lines.sort()
    test_dataset = tf.data.Dataset.from_tensor_slices(path_lines )
    test_dataset = (
        test_dataset.map(
            encode_single_sample, num_parallel_calls=tf.data.experimental.AUTOTUNE
        )
        .batch(batch_size,drop_remainder=False )
        .prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
    )     
    return test_dataset


############    load model     ################

""" load model_finetuned """
def load_model():
    return keras.models.load_model('/home/user/app/text_line_model.h5')