0
$\begingroup$

I am working on a Segmentation task, where I planned to use U-Net

for the input_image of shape (224,224,3), the output should be the mask image of shape (224,224,1)

The mask image contains two unique values - black[0] and white[1]

the output layer from the UNet is having the tensor of shape (None, 224, 224, 1)

I used the softmax activation function for the output layer

The shapes and type for training data

print(trainX.shape) # (200, 224, 224, 3)
print(testX.shape) # (50, 224, 224, 3)
print(trainY.shape) # (200, 224, 224, 1)
print(testY.shape) # (50, 224, 224, 1)

print(trainX.dtype) # float64
print(testX.dtype) # float64
print(trainY.dtype) # int16
print(testY.dtype) # int16

In the mask image, the black pixel are very much more than white pixel. For balancing the black and white pixels, I planned to use class weight [for black - 0.53083749, for white - 8.60701406] in the trianing. So, I write this function

def lossFunc(true, pred):
  weightsList = K.constant([0.53083749, 8.60701406])
  true = K.reshape(true, [-1])
  pred = K.squeeze(pred, axis=3)
  sample_weightsList = K.gather(weightsList, true)
  loss = keras.losses.sparse_categorical_crossentropy(true,pred)
  loss*sample_weightsList

  return loss

But when I started training, I got this error

InvalidArgumentError                      Traceback (most recent call last)
<timed exec> in <module>

/usr/local/lib/python3.10/dist-packages/keras/utils/traceback_utils.py in error_handler(*args, **kwargs)
     68             # To get the full stack trace, call:
     69             # `tf.debugging.disable_traceback_filtering()`
---> 70             raise e.with_traceback(filtered_tb) from None
     71         finally:
     72             del filtered_tb

/usr/local/lib/python3.10/dist-packages/tensorflow/python/eager/execute.py in quick_execute(op_name, num_outputs, inputs, attrs, ctx, name)
     50   try:
     51     ctx.ensure_initialized()
---> 52     tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,
     53                                         inputs, attrs, num_outputs)
     54   except core._NotOkStatusException as e:

InvalidArgumentError: No OpKernel was registered to support Op 'GatherV2' used by {{node lossFunc/GatherV2}} with these attrs: [Tparams=DT_FLOAT, Tindices=DT_INT16, batch_dims=0, Taxis=DT_INT32]
Registered devices: [CPU, GPU]
Registered kernels:
  device='XLA_CPU_JIT'; Taxis in [DT_INT32, DT_INT64]; Tindices in [DT_INT32, DT_INT16, DT_INT64]; Tparams in [DT_FLOAT, DT_DOUBLE, DT_INT32, DT_UINT8, DT_INT16, 930109355527764061, DT_HALF, DT_UINT32, DT_UINT64, DT_FLOAT8_E5M2, DT_FLOAT8_E4M3FN]
  device='XLA_GPU_JIT'; Taxis in [DT_INT32, DT_INT64]; Tindices in [DT_INT32, DT_INT16, DT_INT64]; Tparams in [DT_FLOAT, DT_DOUBLE, DT_INT32, DT_UINT8, DT_INT16, 930109355527764061, DT_HALF, DT_UINT32, DT_UINT64, DT_FLOAT8_E5M2, DT_FLOAT8_E4M3FN]
  device='CPU'; Tparams in [DT_QINT16]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_QINT16]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_QUINT16]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_QUINT16]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_QINT32]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_QINT32]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_QUINT8]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_QUINT8]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_QINT8]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_QINT8]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_VARIANT]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_VARIANT]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_RESOURCE]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_RESOURCE]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_STRING]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_STRING]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_BOOL]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_BOOL]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_COMPLEX128]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_COMPLEX128]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_COMPLEX64]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_COMPLEX64]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_DOUBLE]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_DOUBLE]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_FLOAT]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_FLOAT]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_BFLOAT16]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_BFLOAT16]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_HALF]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_HALF]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_INT32]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_INT32]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_INT8]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_INT8]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_UINT8]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_UINT8]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_INT16]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_INT16]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_UINT16]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_UINT16]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_UINT32]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_UINT32]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_INT64]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_INT64]; Tindices in [DT_INT32]
  device='CPU'; Tparams in [DT_UINT64]; Tindices in [DT_INT64]
  device='CPU'; Tparams in [DT_UINT64]; Tindices in [DT_INT32]
  device='GPU'; Tparams in [DT_BOOL]; Tindices in [DT_INT64]
  device='GPU'; Tparams in [DT_BOOL]; Tindices in [DT_INT32]
  device='GPU'; Tparams in [DT_COMPLEX128]; Tindices in [DT_INT64]
  device='GPU'; Tparams in [DT_COMPLEX128]; Tindices in [DT_INT32]
  device='GPU'; Tparams in [DT_COMPLEX64]; Tindices in [DT_INT64]
  device='GPU'; Tparams in [DT_COMPLEX64]; Tindices in [DT_INT32]
  device='GPU'; Tparams in [DT_DOUBLE]; Tindices in [DT_INT64]
  device='GPU'; Tparams in [DT_DOUBLE]; Tindices in [DT_INT32]
  device='GPU'; Tparams in [DT_FLOAT]; Tindices in [DT_INT64]
  device='GPU'; Tparams in [DT_FLOAT]; Tindices in [DT_INT32]
  device='GPU'; Tparams in [DT_BFLOAT16]; Tindices in [DT_INT64]
  device='GPU'; Tparams in [DT_BFLOAT16]; Tindices in [DT_INT32]
  device='GPU'; Tparams in [DT_HALF]; Tindices in [DT_INT64]
  device='GPU'; Tparams in [DT_HALF]; Tindices in [DT_INT32]
  device='GPU'; Tparams in [DT_INT64]; Tindices in [DT_INT64]
  device='GPU'; Tparams in [DT_INT64]; Tindices in [DT_INT32]
  device='GPU'; Tparams in [DT_INT32]; Tindices in [DT_INT64]
  device='GPU'; Tparams in [DT_INT32]; Tindices in [DT_INT32]

     [[lossFunc/GatherV2]] [Op:__inference_train_function_6870]

What is this issue and How to solve this

Thanks

$\endgroup$

1 Answer 1

0
$\begingroup$

From the docs the gather function requires int32 or int64 indices, while you seem to provide int16. Basically try to cast them to int32:

def lossFunc(true, pred):
  weightsList = K.constant([0.53083749, 8.60701406])
  true = K.cast(K.reshape(true, [-1]), "int32")  # or tf.cast(..., tf.int32)
  pred = K.squeeze(pred, axis=3)
  sample_weightsList = K.gather(weightsList, true)
  loss = keras.losses.sparse_categorical_crossentropy(true,pred)
  loss*sample_weightsList

  return loss
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.