3 ms·

A perfect example is the TensorFlow project's support policy which the article references (i.e. they don't do support.) As long as a project is clear about wha

by blihp 8y ago

A perfect example is the TensorFlow project's support policy which the article references (i.e. they don't do support.) As long as a project is clear about what they will and won't do/provide, just about anything is workable. Far worse to claim to provide support and then be inconsistent about it... better to either say 'we don't provide support' or 'contact us for support options and pricing'.

mlthoughts2018 8y ago
TensorFlow is a very bad example of exactly the failure mode I’m talking about. In professional deep learning, TensorFlow is very widely regarded as a mess, with hostile developers who make breaking changes without warning and refuse to fix huge usability problems, let alone add important enhancements or support-specific bug fixes. Nobody seriously uses TensorFlow without Keras, except in situations when they have to make custom modifications precisely to support broken things that the core devs won’t fix. Contrast this with Keras and it’s night and day. Keras actually has support guidelines and a GitHub project and, even if lacking capacity, actually cares about its users instead of being willfully hostile to users like TensorFlow is. The original creator of Keras, Francois Chollet, can even be found very often in GitHub issue threads assisting totally novice newcomers to Keras and explaining tips and tricks with custom layers, etc.
- blihp 8y ago
  I don't disagree, I find support across the board for Google products dismal. But isn't it better that they say 'we don't do support' than hold out hope that they might or provide spotty support? That way developers and users of the products, including TensorFlow, know what they're getting into. For the free projects that actually do provide good support, that's excellent and kudos to the developer/community that do so... they seem to be the exception these days. I compare it to the 'old days' of software: it used to be common to have a phone number to call. Then it became email only and people complained there should be a support phone number. Then it was $0.99 apps and maybe you got a support forum. Now in the era of freemium, you're lucky if you get an online community support forum... and people complain when nobody monitors it. When you move to free, the support budget goes to $0 which leaves developers to shoulder support in addition to everything else. Unsurprisingly, many don't want to do this. If you're willing to pay, sure you can still get anything you want (up to and including on-site support for enterprise stuff if you have deep enough pockets) but as long as the default support model is clear at least one can make an informed decision. Support is one of those things that the $300+ (phone support), $30+ (email support), $10- (forum?), $0 (good luck) pays/paid for after all...
  - mlthoughts2018 8y ago
    I guess I do agree with you that it’s better for a project to brazenly denounce support than to feign an effort. In the case of TensorFlow in particular, it allows forward thinking people to realize it’s not a safe project to bet the farm on at all, not even anywhere close, even without factoring in Google’s tendency to sunset projects regardless of their relative success or community needs. This is why nobody’s using it. Keras gives you the option of other backends, and there’s a large shift to PyTorch as well. Even being new and not as featureful, you simply just cannot stake the long term health of a project on the constant rug pull that is TensorFlow.

by blihp 8y ago

mlthoughts2018 8y ago

TensorFlow is a very bad example of exactly the failure mode I’m talking about. In professional deep learning, TensorFlow is very widely regarded as a mess, with hostile developers who make breaking changes without warning and refuse to fix huge usability problems, let alone add important enhancements or support-specific bug fixes. Nobody seriously uses TensorFlow without Keras, except in situations when they have to make custom modifications precisely to support broken things that the core devs won’t fix. Contrast this with Keras and it’s night and day. Keras actually has support guidelines and a GitHub project and, even if lacking capacity, actually cares about its users instead of being willfully hostile to users like TensorFlow is. The original creator of Keras, Francois Chollet, can even be found very often in GitHub issue threads assisting totally novice newcomers to Keras and explaining tips and tricks with custom layers, etc.

blihp 8y ago
I don't disagree, I find support across the board for Google products dismal. But isn't it better that they say 'we don't do support' than hold out hope that they might or provide spotty support? That way developers and users of the products, including TensorFlow, know what they're getting into. For the free projects that actually do provide good support, that's excellent and kudos to the developer/community that do so... they seem to be the exception these days. I compare it to the 'old days' of software: it used to be common to have a phone number to call. Then it became email only and people complained there should be a support phone number. Then it was $0.99 apps and maybe you got a support forum. Now in the era of freemium, you're lucky if you get an online community support forum... and people complain when nobody monitors it. When you move to free, the support budget goes to $0 which leaves developers to shoulder support in addition to everything else. Unsurprisingly, many don't want to do this. If you're willing to pay, sure you can still get anything you want (up to and including on-site support for enterprise stuff if you have deep enough pockets) but as long as the default support model is clear at least one can make an informed decision. Support is one of those things that the $300+ (phone support), $30+ (email support), $10- (forum?), $0 (good luck) pays/paid for after all...
- mlthoughts2018 8y ago
  I guess I do agree with you that it’s better for a project to brazenly denounce support than to feign an effort. In the case of TensorFlow in particular, it allows forward thinking people to realize it’s not a safe project to bet the farm on at all, not even anywhere close, even without factoring in Google’s tendency to sunset projects regardless of their relative success or community needs. This is why nobody’s using it. Keras gives you the option of other backends, and there’s a large shift to PyTorch as well. Even being new and not as featureful, you simply just cannot stake the long term health of a project on the constant rug pull that is TensorFlow.