prabhakar-sivanesan / custom-keypoint-detection Goto Github PK

Hey, thanks so much for this tutorial. This is much less of an issue with your tutorial and more an issue with the tf model library I expect!

Basically, when I follow all your steps, during training I get

keypoint/offset': 0.0,
regression': 0.0,

and a total loss rate that very quickly converges to 0. Needless to say, the resulting model is useless.

Any ideas about what might be happening? Any ideas would be much appreciated!

Cheers

I am attaching the colab file which was used for training my document edge detection model. The colab file was cloned from this github repo. Some steps like tfrecord generation and tensorflow github clone is not shown in the colab because it was done already and I had it commented out when I reran the notebook. Please remember the total loss value was 0.08 which should be good for making predictions.

Main Concern:
Why the predictions are to far off when the loss was low? I have followed every step and can't pinpoint to the reason.

Colab file:
my_model.zip

Video for the same:

Unable to add keypoints to category

I ran this command

%cd "/content/TensorFlow/models/research/object_detection" !python dataset_tools/create_coco_tf_record.py --train_image_dir "/content/Custom-keypoint-detection/dataset/images" \ --test_image_dir "/content/Custom-keypoint-detection/dataset/images" \ --val_image_dir "/content/Custom-keypoint-detection/dataset/images" \ --train_annotations_file "/content/Custom-keypoint-detection/dataset/annotations/train.json" \ --testdev_annotations_file "/content/Custom-keypoint-detection/dataset/annotations/val.json" \ --val_annotations_file "/content/Custom-keypoint-detection/dataset/annotations/val.json" \ --train_keypoint_annotations_file "/content/Custom-keypoint-detection/dataset/annotations/train.json" \ --val_keypoint_annotations_file "/content/Custom-keypoint-detection/dataset/annotations/val.json" \ --output_dir "/content/custom_keypoint_detection/dataset/tfrecord/"

It is not generating .record files but giving this output

/content/TensorFlow/models/research/object_detection I0708 07:55:13.186966 139704579422080 create_coco_tf_record.py:401] Found groundtruth annotations. Building annotations index. I0708 07:55:13.187198 139704579422080 create_coco_tf_record.py:414] 0 images are missing annotations. I0708 07:55:13.189434 139704579422080 create_coco_tf_record.py:443] On image 0 of 44 I0708 07:55:13.236015 139704579422080 create_coco_tf_record.py:470] Finished writing, skipped 0 annotations. I0708 07:55:13.236133 139704579422080 create_coco_tf_record.py:473] Finished writing, skipped 0 keypoint annotations. I0708 07:55:13.241920 139704579422080 create_coco_tf_record.py:401] Found groundtruth annotations. Building annotations index. I0708 07:55:13.242027 139704579422080 create_coco_tf_record.py:414] 0 images are missing annotations. I0708 07:55:13.242739 139704579422080 create_coco_tf_record.py:443] On image 0 of 11 I0708 07:55:13.252087 139704579422080 create_coco_tf_record.py:470] Finished writing, skipped 0 annotations. I0708 07:55:13.252181 139704579422080 create_coco_tf_record.py:473] Finished writing, skipped 0 keypoint annotations. I0708 07:55:13.255093 139704579422080 create_coco_tf_record.py:401] Found groundtruth annotations. Building annotations index. I0708 07:55:13.255200 139704579422080 create_coco_tf_record.py:414] 0 images are missing annotations. I0708 07:55:13.255263 139704579422080 create_coco_tf_record.py:443] On image 0 of 11 I0708 07:55:13.263692 139704579422080 create_coco_tf_record.py:470] Finished writing, skipped 0 annotations.

can you please tell what wrong is there in code?

Dear Prabhakar,

Thank you for your "Custom Keypoint Detection" tutorial, it's very helpful!

I could follow most of the steps from the annotation process until the generation of TFrecord, but when I ran the training step I couldn't go further.
The script presented the following error:
tensorflow.python.framework.errors_impl.InvalidArgumentError: {{function_node _wrapped__IteratorGetNext_output_types_19_device/job:localhost/replica:0/task:0/device:GPU:0}} Input to reshape is a tensor with 76 values, but the requested shape requires a multiple of 6
[[{{node Reshape_8}}]]
[[MultiDeviceIteratorGetNextFromShard]]
[[RemoteCall]] [Op:IteratorGetNext]

I annotated 60 512x512 floor plan images with 3 categories, wall corner, door and window, each category with only one keypoint. There's an average of 50 annotations per image.
I'm using Colab with standard GPU for CenterNet HourGlass 104 512x512. I couldn't work with the splitting script for the JSON annotations, so I splitted them manually for training and validation.
When I generated the TFrecord, the generate_tfrecord_from_coco.py script created the train.record and val.record files, but it presented the following error:
File "/content/gdrive/MyDrive/projects/custom_keypoint_detection/dataset/generate_tfrecord_from_coco.py", line 242, in create_tf_example
keypoints = keypoint_annotations['keypoints']
KeyError: 'keypoints'

I've checked many Stack Overflow and Github questions about the issue from the subject, but none of them present specific solutions.

I keep myself available.

Best regards,
Igor

I am getting the following error whenever I am going to train the model:

TypeError: in user code:

File "C:\Users\Abhipray\AppData\Local\Programs\Python\Python37\lib\site-packages\object_detection\inputs.py", line 886, in transform_and_pad_input_data_fn  *
    tensor_dict = pad_input_data_to_static_shapes(
File "C:\Users\Abhipray\AppData\Local\Programs\Python\Python37\lib\site-packages\object_detection\inputs.py", line 319, in transform_input_data  *
    out_tensor_dict[flds_gt_kpt_weights] = (
File "C:\Users\Abhipray\AppData\Local\Programs\Python\Python37\lib\site-packages\object_detection\core\keypoint_ops.py", line 349, in keypoint_weights_from_visibilities  *
    per_keypoint_weight_mult = tf.ones((1, num_keypoints,), dtype=tf.float32)

TypeError: Expected int32, but got None of type 'NoneType'.

@prabhakar-sivanesan I noticed that you have encountered with similar error; how did you address this issue?