assertion 'dim[0] > 0 && dim[1] == 1' failed

[alex-bronia]

Hello! I have a model that outputs a tensor with shape (2,) (Softmax probabilities of two classes). Its output is fed to a tensor_decoder mode=image_labeling option1={LABEL_TEXT_FILE}. I get those errors:

** (python3:31280): CRITICAL **: 15:46:43.233: il_getOutCaps: assertion 'dim[0] > 0 && dim[1] == 1' failed

** (python3:31280): CRITICAL **: 15:46:43.233: il_getOutCaps: assertion 'dim[0] > 0 && dim[1] == 1' failed

** (python3:31280): CRITICAL **: 15:46:43.233: il_getOutCaps: assertion 'dim[0] > 0 && dim[1] == 1' failed

** (python3:31280): CRITICAL **: 15:46:43.233: il_getOutCaps: assertion 'dim[0] > 0 && dim[1] == 1' failed

** (python3:31280): WARNING **: 15:46:43.252: Invalid param, dimensions (0) tensors (1)

What does tensor_decode expect exactly? Or is the error somewhere else, for example in tensor_filter framework=tensorflow2-lite model={MODEL_PATH}? I can't seem to understand solely by log messages + documentation.

If tensor_decode expects a shape of (batch, 2) instead, how can i reshape the tensor, adding a dummy first dimension?

[alex-bronia]

Okay update: I figured out it my last idea was correct and tensor_decode indeed expects a shape of e.g. (1, 2) instead of just (2,). Is there a way to reshape the model output tensor to [1, -1] via some pipeline element? As far as I can see tensor_transform only allows swapping dimensions around but not actual reshaping.

[myungjoo]

@niley7464 It is probably related with the rank-limit updates, which allowed (N, 1) == (N, 0). Please take a look.