Hello, we need someone to create a MATLAB program for us with the following requirements:
1) We have several images for several locations and a layer with classifications for each pixel (N samples, each M x N image, where M and N may be different for each sample). Each location/sample has several images stacked together which we want to use to predict the classification of each pixel. The data for each location/sample is flattened into N matrices (the matrices have row, col of the image and columns for all inputs and targets). That is, each of the N locations/sample consists of a matrix that has been prepared which is of size N*M x K +1 where K is the number of predictors. Each column of the matrix represents an MxN image/raster that has been flattened into a vector. If you require a different form of the data format for input, then you would need to create the code to convert from the stacked images matrix representation to your preferred representation. Contractor should create a CNN model that can be used to predict value of each pixel in an image based on the inputs images. We can provide a few sample files.
2) May only use the following licensed libraries:
Deep Learning Toolbox
Global Optimization Toolbox
Parallel Computing Toolbox
Statistics and Machine Learning Toolbox
It may not use any other matlab licensed libraries.
3) The final model should be compile-able using the matlab C coder to create a DLL. We have done this before with regular patternnet NN in matlab so as long as the model interface is similar it should be fine. We can verify this based on your approach in milestone 1.
4) It should allow for the model to be trained in chunks. All samples would be in total too large to store in memory at once, so the model training framework needs to be set up to successively read in files, train, purge data from memory (but save network), read in more files, further train network, etc.
5) The machine we will run the model on has many GPU cards on a single Windows VM. It should be set up to leverage multiple GPUs to accelerate training.
6) Should be able to deal with missing data (either missing pixels for predictor variable in a sample, a missing predictor variable in entirety in a sample both in training and prediction; or a missing pixel in the target variable in training).
Note: depending on complexity to deploy, an Octave model may also be proposed but must be able to be compiled into a dll either by Matlab C Coder or by you, and meet all requirements above. We would also consider a Python based model.
8 freelanceria on tarjonnut keskimäärin $763 tähän työhön
I have extensive experience with ML, Python and Matlab, in special with computer-vision application and remote-sensing applications, including hyperspectral satelite image classification.