I'm trying to implement the code described in this blogpost with deeplearning4j.
http://www.wildml.com/2015/12/implementing-a-cnn-for-text-classification-in-tensorflow/
I'm having problems implementing multiple kernels on the same channel and concatenating the resulting tensors. Is there even a concat function in deeplearning4j? And is it even possible to have different kernels on the same channel?
Tank you in advance.
You will have to use our computation graph api:
http://deeplearning4j.org/compgraph
Related
I'm planning a TFF scheme in which the clients send to the sever data besides the weights, like their hardware information (e.g CPU frequency). To achieve that, I need to call functions of third-party python libraries, like psutils. Is it possible to serialize (using tff.tf_computation) such kind of functions?
If not, what could be a solution to achieve this objective in a scenario where I'm using a remote executor setting through gRPC?
Unfortunately no, this does not work without modification. TFF uses TensorFlow graphs to serialize the computation logic to run on remote machines. TFF does not interpret Python code on the remote machines.
There maybe a solution by creating a TensorFlow custom op. This would mean writing C++ code to retrieve CPU frequency, and then a Python API to add the operation to the TensorFlow graph during computation construction. TensorFlow's guide for Create an op can provide detailed instructions.
I am trying to use a Pytorch 3D UNet for inference (from here: https://github.com/wolny/pytorch-3dunet) which receives images of size (96, 96, 96). I would like to use it on CPU instances, but I am getting very high memory usages (~18 GB). After researching on the subject I found out that this was due to the way convolutions are implemented on CPU (see https://discuss.pytorch.org/t/pytorch-high-memory-demand/2798/5). I thus have the following questions:
Is there a way to use a more memory-efficient implementation of the convolution in Pytorch?
How can I optimize my model for CPU inference? I saw that some tools like AWS Neo, Intel OpenVINO, etc. exist; could they solve my problem?
Does Tensorflow have a similar problem for using convolutions on CPU?
Any other tip, link on how to deploy such models in an efficient way is welcome!
Thanks!
You could benchmark your model's performance with DNN-Bench and choose the best inference engine for your application and your hardware. You might need to convert your model to ONNX first.
So as a fun project, I've been messing with the TensorFlow API (in Java unfortunately.. but I should be able to get some results out anyways). My first goal is to develop a model for 2D point cloud filtering. So I have written code that generates random clouds in 224x172 resolution, computes the result of a neighbor density filter, and stores both (see images below).
So basically I have generated data for both an input and expected output, which can be done as much as needed for a massive dataset.
I have both the input and output arrays stored as 224x172 binary arrays (0 for no point at index, 1 for a point at that index). So my input and output are both 224x172. At this point, I'm not sure how to translate my input to my expected result. I'm not sure how to weight each "pixel" of my cloud, or how to "teach" the program the expected result. Any suggestions/guidance on whether this is even possible for my given scenario would be appreciated!
Please don't be too hard on me... I'm a complete noob when it comes to machine learning.
Imagine, that Tensorflow is a set of building blocks (like LEGO) that allows constructing machine learning models. After the model is constructed, it could be trained and evaluated.
So basically your question could be divided into three steps:
1. I'm new to machine learning. Please guide me how to choose the model that fits the task.
2. I'm new to tensorflow. I have the idea of model (see 1) and I want to construct it via Tensorflow.
3. I'm new to Java's API of tensorflow. I know how to build model using tensorflow (2), but I'm stuck in Java.
This sounds scaring, but that's not too bad really. And I'd suggest you the following plan to do:
1. You need to look through the machine learning models to find the model that suits your case. So you need to ask yourself: what are the models that could be used for cloud filtering? And basically, do you really need some machine learning model? Why don't you use simple math formulas?
That are the questions you may ask to yourself.
Ok, assume you've found some model. For example, you've found a paper describing neural network able to solve your tasks. Then go to the next step
2. Tensorflow has a lot of examples and code snippets. So you could even
find the code with your model implemented yet.
The bad thing, that most code examples and API are Python based. But as you want to go into machine learning, I'd suggest you studying Python. It's easy to enter. It's very common to use python in the science world as it allows not to waste time on wrappers, configuration, etc. (as Java needs). We just start solving our task from the first line of the script.
As I've told initially about tensorflow and LEGO similarity, I'd like to add that there are more high-level additions to the tensorflow. So you work with the not building blocks but some kind of layers of blocks.
Something like tflearn. It's very good especially if you don't have deep math or machine learning background. It allows building machine learning models in a very simple and understandable way. So do you need to add some neural network layer? Here you are. And that's all without complex low-level tensor operations.
The disadvantage that you won't be able to load tflearn model from Java.
Anyway, we assume, at the end of this step you are able to build your model, to train it and to evaluate the model and prediction quality.
3. So you have your machine learning model, you understand Tensorflow mechanics, and if you still need to work with Java that should be much easier yet.
I note, that you won't be able to load tflearn model from Java. You can try to use jython to call python's functions directly from Java, though I haven't tried it.
And on this way (1-3) you will definitely have some more questions. So welcome to SO.
I've got a bunch of images (~3000) which have been manually classified (approved/rejected) based on some business criteria. I've processed these images with Google Cloud Platform obtaining annotations and SafeSearch results, for example (csv format):
file name; approved/rejected; adult; spoof; medical; violence; annotations
A.jpg;approved;VERY_UNLIKELY;VERY_UNLIKELY;VERY_UNLIKELY;UNLIKELY;boat|0.9,vehicle|0.8
B.jpg;rejected;VERY_UNLIKELY;VERY_UNLIKELY;VERY_UNLIKELY;UNLIKELY;text|0.9,font|0.8
I want to use machine learning to be able to predict if a new image should be approved or rejected (second column in the csv file).
Which algorithm should I use?
How should I format the data, especially the annotations column? Should I obtain first all the available annotation types and use them as a feature with the numerical value (0 if it doesn't apply)? Or would it be better to just process the annotation column as text?
I would suggest you try convolutional neural networks.
Maybe the fastest way to test your idea if it will work or not (problem could be the number of images you have, which is quite low), is to use transfer learning with Tensorflow. There are great tutorials made by Magnus Erik Hvass Pedersen, who published them on youtube.
I suggest you go through all the videos, but the important ones are #7 and #8.
Using transfer learning allows you to use the models they build at google to classify images. But with transfer learning, you are able to use your own data with your own labels.
Using this approach you will be able to see if this is suitable for your problem. Then you can dive into convolutional neural networks and create the pipeline that will work the best for your problem.
Im working on an automatic image annotation problem in which im trying to associate tags with images. For that im trying for SIFT features for learning. But the problem is all the SIFT features are a set of keypoints, each of which have a 2-D array, and the number of keypoints are also huge.How many and how do I give them for my learning algorithm which typically accepts only one-d features?
You can represent single SIFT as "visual word" which is one number and use it as SVM input, I think it is what you need. It is usually done by k-means clustering.
This method is called "bag-of-words" and described in this paper.
Short presentation review of method.
You should read the original paper about SIFT, it tells you what is SIFT and how to use it, you should carefully read the chapter 7 and rest for understanding how to use it practically.
Here is the link for original paper.
You can use the Bag of Words approach, of which you can read about in the following post:
http://gilscvblog.wordpress.com/2013/08/23/bag-of-words-models-for-visual-categorization/
Sift and Surf are invariant feature extractors. There for matching features will help solving lots of problems.
But there is matching problem since all points may not be same in two different image. (and in the case of similarity problem). Therefore you should use the features which is matched the others may.
Another problem is this algorithms extract lots of features which is not possible to match in large datasets.
There is a good solution to those problems which is called "Bag of Visual Word"
https://github.com/dermotte/LIRE complete bag of visual word is fully implemented. Here is the lire Demo site.
Code is very simple if you know the bag of visual word you can modify also.
After getting visual word you should use information retrieval approaches used in search engines. By the way Lire also include an information retrieval library called lucene. You should fallow the lire way until you get the complete idea and implement your own.