Unable to process data from flickr_logos_27_dataset\flickr_logos_27_dataset_cropped_augmented_images\Adidas\train.pickle : could not broadcast input array from shape (4412,32,64,3) into shape (35000,32,64,3)

[rmirzojonov]

Hello, I am new to tensorflow so I tried to reduce the number of logos to only one (Adidas) to make it run faster.
Removed everything except Adidas data from flickr_logos_27_dataset_query_set_annotation.txt and flickr_logos_27_dataset_training_set_annotation.txt.
Edited common.py:

CLASS_NAME = [
    'Adidas', 'Background'
]

train and test folders are created for both Adidas and Background.
So when I try to run gen_train_valid_test.py, it shows and error:

Unable to process data from flickr_logos_27_dataset\flickr_logos_27_dataset_cropped_augmented_images\Adidas\train.pickle : could not broadcast input array from shape (4412,32,64,3) into shape (35000,32,64,3)
Traceback (most recent call last):
File "c:\Users\username\Documents\Workspace\DeepLogo\gen_train_valid_test.py", line 196, in
main()
File "c:\Users\username\Documents\Workspace\DeepLogo\gen_train_valid_test.py", line 182, in main
train_datasets, TRAIN_SIZE, VALID_SIZE)
File "c:\Users\username\Documents\Workspace\DeepLogo\gen_train_valid_test.py", line 132, in merge_datasets
train_dataset[start_t:end_t, :, :, :] = train_logo
ValueError: could not broadcast input array from shape (4412,32,64,3) into shape (35000,32,64,3)

Can you help me solving this issue. Thanks!

[bruce-willis]

Hello @rmirzojonov!
Unfortunately, nowadays you can not

Removed everything except Adidas data from flickr_logos_27_dataset_query_set_annotation.txt and flickr_logos_27_dataset_training_set_annotation.txt.
Edited common.py:

So, to fix that you have to generate background, crop and aug for all classes :(

[bruce-willis]

@rmirzojonov
Size of train and valid dataset are calculated at line 114 gen_train_valid_test.py

vsize_per_class = valid_size // num_classes
tsize_per_class = train_size // num_classes

The size depends on number of classes: less classes you have - bigger vsize per class you get. I don't think it is a good idea to have such dependency. And if you have only 1 class — vsize will be bigger than size of your class

[juules87]

Could you please explain how to adjust the model when estimating the model only for one logo without including all logo's?

Thanks in advance

[fdornelasx]

I am actually having the same problem. I have 5 classes and getting the same problem. What could I do to solve this?

Here my command line error:
Unable to process data from tv_logos_dataset/tv_logos_dataset_cropped_augmented_images/Globo/train.pickle : could not broadcast input array from shape (27,32,64,3) into shape (1250,32,64,3) Traceback (most recent call last): File "gen_train_valid_test.py", line 196, in <module> main() File "gen_train_valid_test.py", line 182, in main train_datasets, TRAIN_SIZE, VALID_SIZE) File "gen_train_valid_test.py", line 127, in merge_datasets valid_dataset[start_v:end_v, :, :, :] = valid_logo ValueError: could not broadcast input array from shape (27,32,64,3) into shape (1250,32,64,3)

I'm trying just for test with 1 image for each class.