tensorflow/models
I installed tensorflow through anaconda, and here I want to include the tensorflow/model.
Can i install tensorflow/models package with pip install?
I want to import the code as shown below.
import nets.inception_resnet_v2
We do not yet, nor do we currently have any plans to, package tensorflow/models in a pip package. You'll have to clone the repository and/or copy-paste the code.
I understand this issue is closed, but wanted to put this back on the radar. I believe that installing pre-trained models is a fair use case, and having the ability to run pip install tensorflow-models
would be fantastic. Could you elaborate a bit more on the reasons for not considering a pip package for tensorflow/models?
21 votes not enough to reopen this issue, or at least provide a close justification?
Apologies, this was closed more than a year ago (before TensorFlow 1.0, i.e., before API stability guarantees from main TensorFlow).
This may be worth revisiting now.
@karmel - any thoughts?
We have discussed this, and everyone agrees it is a good idea. Let's leave this open to track for now.
@asimshankar @karmel it would also be useful to have official.utils.logs import hooks_helper
and official.utils.misc import model_helpers
as an installable package as well. There are very few examples out there of using them.
The amount of IT work involved in just getting official/wide_deep to run due to the use of official.utils is pretty ridiculous. Please either provide a write-up on how to get working easier or update the packages used within the py files.
@bstreit -- please see the Official Models README for instructions in setting up the Python path. If you have further issues with that, please open up a new issue, and we can help resolve.
Could you also create a Pip installable package for the research folder?
That's an interesting suggestion, but pip installation for /research is not planned. Each model is owned and maintained by the researchers in question, so providing tooling on top of all of those would be exceedingly hard to maintain.
Actually, I was thinking about the object_detection folder. It would make it more easy and reproducible its usage. We should be able to easily pip install it instead of cloning the whole repository. What do you think?
That would be under the purview of the researchers-- can you open a separate issue for that?
Any progress on the pip installable update to object detection or (even better) all of the models? Eagerly awaiting this.
For the official models, we have not made any progress-- though if someone wants to pitch in, I'm marking this as Contributions Welcome. For object detection, please see #4887
Hello, is there an update on this ?
There are no current contributors who have picked this up, AFAIK. @lintian06 -- you may also be interested, as pip installation is an oft requested feature.
Thank you @karmel! I'll take a look at this.
A general thought is that we need some sorts of design for different subfolders inside the project, since they may have different configurations and prerequisites.
Would be interested in contributing to making the package pip installable.
What a pity for two years waiting...
Working in progress. We have an initial package out for official models.
We plan to release an light pip package along with TF 2.1 release.
Yes, You can install it
pip install tensorflow (for python2)
pip3 install tensorflow (for python3)
Where is the module actually being installed?
import tensorflow.models
raise an error.
Where is the module actually being installed?
import tensorflow.models
raise an error.
Please install the nightly pip first:
pip install tf-models-nightly
And then import in the following way for example:
from official.nlp.bert import ...
We may make the import easier in the future.
Thanks, Chen!
@jaeyounkim Shall we close this issue?
Yes, we can close this issue.
Most helpful comment
I understand this issue is closed, but wanted to put this back on the radar. I believe that installing pre-trained models is a fair use case, and having the ability to run
pip install tensorflow-models
would be fantastic. Could you elaborate a bit more on the reasons for not considering a pip package for tensorflow/models?