You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
According to https://coral.ai/docs/edgetpu/models-intro/#supported-operations coral does not seem to support Transpose op. I would presume that under some limiting circumstances (e.g., not transposing batch dim, etc.) the device should be able to support this. And the operation would be very helpful -- given size limitations, using grouped convolutions might be an easy way to get better accuracy. However, such architectures usually involve a channel "shuffle" so that channels from different groups can get mixed and such shuffle is usually implemented by reshape + transpose + reshape.
The text was updated successfully, but these errors were encountered:
For any one face this issue
by converting pytorch to keras (using this rep https://github.com/nerox8664/pytorch2keras ) i was able to fully convert my model to NHWC and by that avoiding the TRANSPOE nodes, edge tpu comile was able to map all operations to TPU (:
According to https://coral.ai/docs/edgetpu/models-intro/#supported-operations coral does not seem to support Transpose op. I would presume that under some limiting circumstances (e.g., not transposing batch dim, etc.) the device should be able to support this. And the operation would be very helpful -- given size limitations, using grouped convolutions might be an easy way to get better accuracy. However, such architectures usually involve a channel "shuffle" so that channels from different groups can get mixed and such shuffle is usually implemented by reshape + transpose + reshape.
The text was updated successfully, but these errors were encountered: