Datasets: Wrong labels in coil100 dataset, has 72 instead of 100

[duran67]

Hi, I believe the labels are wrong the dataset.

If you examine the "label" feature it says number of classes is 72, it should be 100.
info.features["label"].num_classes
72
72 is how many images there are per class, 100 is the number of objects/classes.

If you list the labels it shows the 5-degree steps.
info.features["label"].names
'0' '5' '10' '15' '20' '25' '30' ...

Should be:
'obj1', 'obj2', 'obj3', 'obj4', 'obj5' ...

Maybe solution in coil100.py

_LABELS = ['obj'+str(x) for x in range(1, 101)]

And in _generate_examples I think label should be:
label = int(file_name.split("__")[0][3:])-1
instead of:
label = file_name.split("_")[2].split(".")[0]

(and not sure if object_id is needed)

[Conchylicultor]

As far as I understand the implementation, it seems label correspond to the angle, while object_id correspond to the object id.
I agree this is confusing.

• object_id should be changed to a ClassLabel
• label should be renamed angle_label
• There likelly should be an additional 'angle': tf.int64 feature to contain the raw angle.

Don't hesitate to send a PR to update the dataset

[harsh020]

@Conchylicultor Hi, I was working on this issue, I seem to have run into an issue. When I renamed the key label to angle_label, it seemed that I was able to load the dataset successfully using load but I was unable to iterate the dataset using take.

Code: (Note: label is changed to angle_label)

def peek(ds):
  print(ds)
  return ds

ds_train,info= tfds.load("coil100", split='train', with_info=True)
ds_ = ds_train.map(peek, num_parallel_calls=tf.data.experimental.AUTOTUNE)   #works fine

ds = ds_train.shuffle(1024).batch(32).prefetch(tf.data.experimental.AUTOTUNE) 
for d in ds.take(1):   #raises error
    print(d)

• The peek function prints

{'angle_label': <tf.Tensor 'args_0:0' shape=() dtype=int64>, 'image': <tf.Tensor 'args_1:0' shape=(128, 128, 3) dtype=uint8>, 'object_id': <tf.Tensor 'args_2:0' shape=() dtype=string>}

• The error take raises is

tensorflow.python.framework.errors_impl.InvalidArgumentError: Feature: angle_label (data type: int64) is required but could not be found.

(Note: Same error is raised even when loaded as supervised)

But angle_lable seems to be of dtype: int64.

Same goes to object_id, when being converted to ClassLabel.

Thank You.

[Conchylicultor]

Thanks for looking into this. Could you open a draft PR so we can take a look ?

[Conchylicultor]

Could you make sure the data is fully regenerated by deleting ~/tensorflow_datasets/coil100/

[harsh020]

Regenerating the dataset seems to have resolved the issue. Thank you very much. I am committing rest of the changes and opening the PR for review.

Thank You.

[harsh020]

@duran67 Hi, I hope PR #2769 resolved your issue.
Now, label has become angle_label with 72 classes (0, 5, 10, 15,.... labelled as 0, 1, 2, 3, ....)
We have converted object_id to a ClassLabel with 100 classes (obj1, obj2, obj3,.... labelled as 0, 1, 2, 3....)
We have also added angle feature which stores raw angles (0, 5, 10, 15.....)

Thank You.