generate handwriting process is successful but no image appears

[dhea1323]

Hello @Belval and community,

I run this code for generate handwritten images in colab. Process is successfull but no image appears in the output directory. When I generate an image with some kind of font without -hw, the image appears in the output directory

Anyone can help?

[Belval]

Handwritten text generation is not working well right now, it's very possible that something broke in the latest TensorFlow version. I'll check if I can repro.

[hoangtv2000]

In my case, I downgrade the Python to 3.7 and Install Tensorflow==1.13.1 and the model works fine.

[RickyGunawan09]

In my case, I downgrade the Python to 3.7 and Install Tensorflow==1.13.1. I still don't get the image, so I try to see the code and I found this error in section # Comparing average pixel value of text and background image # when you want to generate HW (try and except temporary I comment). I'm not using CLI but using python program using dict for language indonesia ("id")

(generate) D:\ocr_handwriting\generate_data\TextRecognitionDataGenerator>python test_hw.py 2022-08-16 10:20:19.738744: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 WARNING:tensorflow:From C:\Users\1000.conda\envs\generate\lib\site-packages\tensorflow\python\training\saver.py:1266: checkpoint_exists (from tensorflow.python.training.checkpoint_management) is deprecated and will be removed in a future version. Instructions for updating: Use standard file APIs to check for files with this prefix. image <PIL.Image.Image image mode=RGBA size=419x336 at 0x2470D1A3348> mask <PIL.Image.Image image mode=RGB size=640x480 at 0x2470D171308> background_img <PIL.Image.Image image mode=RGBA size=37x32 at 0x2470704B3C8> background_mask <PIL.Image.Image image mode=RGB size=37x32 at 0x24707051C48> Traceback (most recent call last): File "test_hw.py", line 22, in for img, lbl in generator: File "D:\ocr_handwriting\generate_data\TextRecognitionDataGenerator\trdg\generators\from_strings.py", line 97, in next return self.next() File "D:\ocr_handwriting\generate_data\TextRecognitionDataGenerator\trdg\generators\from_strings.py", line 134, in next self.output_bboxes, File "D:\ocr_handwriting\generate_data\TextRecognitionDataGenerator\trdg\data_generator.py", line 191, in generate resized_img_px_mean = sum(resized_img_st.mean[:2]) / 3 File "C:\Users\1000.conda\envs\generate\lib\site-packages\PIL\ImageStat.py", line 47, in getattr v = getattr(self, "_get" + id)() File "C:\Users\1000.conda\envs\generate\lib\site-packages\PIL\ImageStat.py", line 103, in _getmean v.append(self.sum[i] / self.count[i]) ZeroDivisionError: float division by zero`

I don't know if it's really help or not but i don't know how to fix this. my code:

generator = GeneratorFromDict( count=1, fonts=["./fonts/arial.ttf"], language="id", background_type=1, is_handwritten=True ) for img, lbl in generator: # Do something with the pillow images here. img.save(f"./z_out/{lbl}.jpg")`

[simmerken]

I make the following modifications, it works fine~

185     ¦   ##############################################################
186     ¦   # Comparing average pixel value of text and background image #
187     ¦   ##############################################################
188     ¦   try:
189     ¦   ¦   resized_img_st = ImageStat.Stat(resized_img, resized_mask.split()[2])
190     ¦   ¦   background_img_st = ImageStat.Stat(background_img)
191 
192     ¦   ¦   if is_handwritten:
193     ¦   ¦   ¦   resized_img_px_mean = 25
194     ¦   ¦   else:
195     ¦   ¦   ¦   resized_img_px_mean = sum(resized_img_st.mean[:2]) / 3
196 #            print(resized_img_px_mean)
197     ¦   ¦   background_img_px_mean = sum(background_img_st.mean) / 3                                                                                                                                                                                                   
198 #            print(background_img_px_mean)
199 

[RickyGunawan09]

I make the following modifications, it works fine~

185     ¦   ##############################################################
186     ¦   # Comparing average pixel value of text and background image #
187     ¦   ##############################################################
188     ¦   try:
189     ¦   ¦   resized_img_st = ImageStat.Stat(resized_img, resized_mask.split()[2])
190     ¦   ¦   background_img_st = ImageStat.Stat(background_img)
191 
192     ¦   ¦   if is_handwritten:
193     ¦   ¦   ¦   resized_img_px_mean = 25
194     ¦   ¦   else:
195     ¦   ¦   ¦   resized_img_px_mean = sum(resized_img_st.mean[:2]) / 3
196 #            print(resized_img_px_mean)
197     ¦   ¦   background_img_px_mean = sum(background_img_st.mean) / 3                                                                                                                                                                                                   
198 #            print(background_img_px_mean)
199 

I tried your fix and I can get the image, but I find another issue with the width for handwriting. it's always generated an image with a resolution of 37x32 for every word. did you edit another line of code?

edit: I can resize it using size parameter. so my question is what did # Comparing average pixel value of text and background image # actually do? it's okay if we use hardcode like you did @simmerken ?

[meminbakir]

Actually, there happens a ZeroDivision error but since it is caught, it is not visible. So far, the solution suggested at this issue helped me to solve the problem. I hope the author applies the suggested solution and saves many from headache...

[stweil]

Related issue: #251. So this issue might be closed as a duplicate.