I am not impressed, here is why.

[Der Zurechtweiser]

I have been observing the development of tensorflow for a while and I have to say that I am not impressed by the results.

Everything tensorflow has dealt with so far was done better with other means before.

Let's take speech synthesis for instance, wavenet. When downloading the github-sourcecode and trying it with your own files or the files given, it just is not performing as well as old-school tts. It sounds unnatural and I don't think they created their samples with the given source code.

Or let's take deep dream. What do you really use this for, it just looks bad.

Or let's take neural style. There is never a satisfying outcome that you would prefer over the input image. The result just looks artificial and cold and not in the slightest similar to the original creator.

It is always the same, it is for those who enjoy playing with cold data and derive some enjoyment from it, but real machine "learning", realizing the emotions behind the images in the case of neural style or understanding what sounds good in the case of wavenet is just not happening with the code that is out there for tensorflow.

It might be solvabe by tinkering with the parameters, but that seems endless and you never know if it's even possible to improve the result.

[Der Zurechtweiser]

Example output of wavenet: https://soundcloud.com/user-731806733/tensorflow-wavenet-500-msec-88k-train-steps

Taken from https://github.com/ibab/tensorflow-wavenet

[Benjamin Ellenberger]

The problem you are observing is not related to the development of tensorflow or the code you were mentioning, but is more related to the data efficiency of deep learning in general. Most of the code you mention just does not get into the range of quality you would expect if you do not feed it with massive amounts of data. There have been several blog posts on this topic which explain why large companies give you their toolset. It boils down to "data is the new oil". Algorithms might improve in the future, so that data can be used more efficiently, but for now it is just about who has the highest amount of data. So without enough data, tts performs better than wavenet because the wavenet barely captured the modes of your distribution, let alone any of the details (that is why there is so much blabbering in the wavenet outputs you added, it just was trained to the point where it understood that it requires blabbering in order to be similar.)

[Der Zurechtweiser]

How would you know that more data generated a better result? You can always say that and repeat the sentence; do you just want to be "right" or can you clarify how the data with for instance wavenet has to look like for it to become the way it was presented in the papers of wavenet?

[Edvard Fagerholm]

I think you're complaining to the wrong people. TensorFlow is a framework for implementing neural networks (as well as some other things), while WaveNet is a particular NN for which there's a TensorFlow implementation. You don't complain to the Clang project that their C++ compiler sucks, because you cloned someone's code from Github, which was in C++, and it didn't work. None of those networks you mention would work any better had you ported the same code to e.g. PyTorch.

BTW, speech synthesis and speech-to-text are quite well known to be very data hungry problems for which it's also expensive to produce good training data. Making these work well for some smaller local language is a pretty big undertaking due to the required data collection and annotation.

Best,

Edvard

On Fri, Mar 2, 2018 at 1:10 PM, 'Der Zurechtweiser' via Discuss <dis...@tensorflow.org> wrote:

How would you know that more data generated a better result? You can always say that and repeat the sentence; do you just want to be "right" or can you clarify how the data with for instance wavenet has to look like for it to become the way it was presented in the papers of wavenet?

--
You received this message because you are subscribed to the Google Groups "Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to discuss+unsubscribe@tensorflow.org.
To post to this group, send email to dis...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/988602c5-5226-4377-a6f5-a9af835a0b9c%40tensorflow.org.

[Der Zurechtweiser]

What does "expensive" mean in " it's also expensive to produce good training data"?

Your argument is like saying that Silicium was perfect, the problems came with people building turing machines with it. But isn't it surprising that none of the projects being built upon tensorflow are convincing? Contrary to that, a lot of projects built on c++ are convincing.

[Edvard Fagerholm]

Most companies can't open source ML stuff, because the training data can't be published and the code itself would be completely useless. Hence, you see very few production quality open source ML projects.

Expensive means recording, say, a hundred thousand audio clips, annotating the precise words that were spoken and aligning the text to the audio clips. This will require time per clip. Now multiply by the number of clips and make up a number based on the estimated hourly pay.

--

You received this message because you are subscribed to the Google Groups "Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to discuss+unsubscribe@tensorflow.org.
To post to this group, send email to dis...@tensorflow.org.

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/40e1de45-b0bc-4c41-ac27-cd1e6a19c228%40tensorflow.org.

[Der Zurechtweiser]

Which production quality open source wavenet implementation do you know?

Am Freitag, 2. März 2018 13:16:46 UTC+1 schrieb Edvard Fagerholm:

Most companies can't open source ML stuff, because the training data can't be published and the code itself would be completely useless. Hence, you see very few production quality open source ML projects.

Expensive means recording, say, a hundred thousand audio clips, annotating the precise words that were spoken and aligning the text to the audio clips. This will require time per clip. Now multiply by the number of clips and make up a number based on the estimated hourly pay.

On Fri, Mar 2, 2018 at 2:01 PM, 'Der Zurechtweiser' via Discuss <dis...@tensorflow.org> wrote:

What does "expensive" mean in " it's also expensive to produce good training data"?

Your argument is like saying that Silicium was perfect, the problems came with people building turing machines with it. But isn't it surprising that none of the projects being built upon tensorflow are convincing? Contrary to that, a lot of projects built on c++ are convincing.

Am Freitag, 2. März 2018 12:28:20 UTC+1 schrieb Edvard Fagerholm:

I think you're complaining to the wrong people. TensorFlow is a framework for implementing neural networks (as well as some other things), while WaveNet is a particular NN for which there's a TensorFlow implementation. You don't complain to the Clang project that their C++ compiler sucks, because you cloned someone's code from Github, which was in C++, and it didn't work. None of those networks you mention would work any better had you ported the same code to e.g. PyTorch.
BTW, speech synthesis and speech-to-text are quite well known to be very data hungry problems for which it's also expensive to produce good training data. Making these work well for some smaller local language is a pretty big undertaking due to the required data collection and annotation.
Best,
Edvard

--

You received this message because you are subscribed to the Google Groups "Discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to discuss+u...@tensorflow.org.

To post to this group, send email to dis...@tensorflow.org.

[Edvard Fagerholm]

None.

To unsubscribe from this group and stop receiving emails from it, send an email to discuss+unsubscribe@tensorflow.org.

To post to this group, send email to dis...@tensorflow.org.

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/3a9bf8cb-77bd-4b41-a748-dd4234a384da%40tensorflow.org.

[Der Zurechtweiser]

In the case of wavenet, as a first step tensorflows hyperparameters have to adjust themselves to the current data. If I supply an audio file containing a speaker speaking a sentence and a text file with it's content that is all I should need. The number of epochs and other hyperparameters have to be predicted and automatically set, so I get something like "Training this set of data will take X hours to achieve an accuracy of Y per cent similarity". Tensorflow/Google itself should offer this implementation tailored to it's framework, so we can finally attribute success or failure to the underlying framework. Otherwise it will always be blame pushing.

[Der Zurechtweiser]

TensorFlow 1.7.0-rc1 has been released,

• Eager execution is moving out of contrib, try tf.enable_eager_execution().
• Graph rewrites emulating fixed-point quantization compatible with TensorFlow Lite, supported by new tf.contrib.quantize package.
• Easily customize gradient computation with tf.custom_gradient.
• TensorBoard Debugger Plugin, the graphical user interface (GUI) of TensorFlow Debugger (tfdbg), is now in alpha.
• Experimental support for reading a sqlite database as a Dataset with new tf.contrib.data.SqlDataset.
• Distributed Mutex / CriticalSection added to tf.contrib.framework.CriticalSection.
• Better text processing with tf.regex_replace.
• Easy, efficient sequence input with tf.contrib.data.bucket_by_sequence_length

Do you guys even have usefulness of applications set as your goal?

For instance in android, you tried to at least give guidelines for design with material design. There went some thought into it. But what you are adding is working on another level. You are solving problems nobody has.

I presume, sooner or later, people will put deep learning aside again and focus on reliable, old-fashioned tinkering again.

There is way too much hype and way too little reproducable results.

[Edd Wilder-James]

Please keep the discussion in this forum helpful: its purpose is for discussion and support around TensorFlow.  This criticism isn't really constructive feedback, or useful in keeping this forum welcoming and engaged. I respectfully request you find an alternate forum for these criticisms of deep learning at large.

[mingya...@gmail.com]

I know a new product, it may answer your doubts.          https://play.google.com/store/apps/details?id=com.my.texttomp3

在 2018年3月2日星期五 UTC+8上午6:44:35，Der Zurechtweiser写道：