Release 5.0 and installation with pip #1153
Conversation
…d rename old setup.py to not infer with new install. This file will be delete ultimately.
…exes which was missing its implementation in the datastore. Added tf to pi install.
… works from withing the donkeycar project directory.
This reverts commit 568c0e6.
…pandas / matplotlib don't work. Updated version and pushed that to pypi test.
|
I have reviewed the changes included for Release 5.0 and would like to encourage you to include changes to the MacOS setup to leverage the Mac GPU. METAL is the GPU technology that powers hardware-accelerated graphics and AI on Apple platforms. METAL is supported by Mac computers from the year 2012. See METAL on Mac Computers. Apple has released tensorflow-macos and tensorflow-metal to speedup tensorflow on the Mac. I have used this installation on both an INTEL and ARM mac for the last couple of years. Here is a set of instructions I have used for installation on the Macbook Air M1. Install Tensorflow-macos. There is one difference for the ARM Mac. For mac computers prior to 2012, the setup would use standard tensorflow. Once you merge the files into the MAIN branch I will work through suggested changes to the code and instructions for each type of Mac. Maybe you could solicit others with Macs to test. |
@cfox570 - this is a very good comment. Having a quick look at your gist above, it looks like we should simply replace |
For Conda, the ARM Mac (M1/M2/M3) needs tensorflow-deps installed via conda not pip. Not sure how you change your setup to accomplish. |
This is already part of this PR. Everything now gets installed via pip. Going from |
|
To clarify, both tensorflow-macos and tensorflow-metal are required.
|
|
I tried out the installation of [torch]. I found two issues:
Also I suggest you update the mac instructions to install [macos] instead of [pc] |
Thanks @cfox570 - we have tested on Mac Intel with the standard TF install that is in the |
| @@ -78,8 +78,7 @@ def write_record(self, record=None): | |||
| # Private properties | |||
| contents['_timestamp_ms'] = int(round(time.time() * 1000)) | |||
| contents['_index'] = self.manifest.current_index | |||
| contents['_session_id'] = self.manifest.session_id | |||
|
|
|||
| contents['_session_id'] = self.manifest.session_id[1] | |||
There was a problem hiding this comment.
why the index of 1 here? It's magic-number-ish. Maybe worth a comment.
There was a problem hiding this comment.
Sorry, saw that comment too late... session_id now returns a tuple of the (number, session_id_string) which allows to simplify the code a lot, so we are picking the string here.
Release 5.0 and installation with pip
This PR is in preparation for the Donkey Car 5.0 release. Compared to current HEAD, this release has the following major changes:
setup.cfgandpyproject.tomlinstallationpip install donkeycar[XX], where XX contains the platform.keras-vizdependency from which we only required two utility functions which have now been copied into the code directlyAs part of the 5.X release we will also be updating the documentation accordingly. The changes for the documentation are in https://github.com/autorope/donkeydocs/tree/install_with_pip.