I have been trying to learn TensorFlow. But just trying to get the tutorial examples to work is killing me. Most of the examples on TensorFlow's GitHub page just can't work out of the box. It seems that TensorFlow's API is changing so fast, with methods being moved from module to module, that slightly older examples just can't work with the newer version of the API.
It is frustrating, trying to figure out what has been moved to contrib, what has been moved to legacy, what has been moved to deprecated, etc. Then having to change the source code, try running, get errors, change source code again, and run again, and the cycle continues. Until it finally runs for a while, you spend an hour reading and preparing the large data file, and then it gives another error.
Especially since the data for deep learning is usually very big. Huge.
It is a terrifying experience.
But I think it is also can't be helped. TensorFlow is the bleeding edge, and changes daily. Just that there should be a bit more organization, a bit more planning out of the API, so that we can avoid a bit more heart pain.
It is frustrating, trying to figure out what has been moved to contrib, what has been moved to legacy, what has been moved to deprecated, etc. Then having to change the source code, try running, get errors, change source code again, and run again, and the cycle continues. Until it finally runs for a while, you spend an hour reading and preparing the large data file, and then it gives another error.
Especially since the data for deep learning is usually very big. Huge.
It is a terrifying experience.
But I think it is also can't be helped. TensorFlow is the bleeding edge, and changes daily. Just that there should be a bit more organization, a bit more planning out of the API, so that we can avoid a bit more heart pain.
No comments:
Post a Comment