# Structure of machine Learning algorithm

I am experimenting on automation on video editing. The idea is obtaining "strange results" and experiments, not looking for any precise result.

So, I want to design a simple neural network that takes as a dataset some .txt files, that contain strings that encode steps from a video editing software. This is an example of strings contained in this file:

ID";"Track";"StartTime";"Length";"PlayRate";"Locked";"Normalized";"StretchMethod";"Looped";"OnRuler";"MediaType";"FileName";"Stream";"StreamStart";"StreamLength";"FadeTimeIn";"FadeTimeOut";"SustainGain";"CurveIn";"GainIn";"CurveOut";"GainOut";"Layer";"Color";"CurveInR";"CurveOutR";"PlayPitch";"LockPitch";"FirstChannel";"Channels"
1; 1; 3482.8288; 5035.0080; 1.000000; FALSE; FALSE; 0; TRUE; FALSE; VIDEO; "path/to/video"; 0; 1987135.9834; 5045.0080; 0.0000; 0.0000; 1.000000; 4; 0.000000; 4; 0.000000; 0; -1; 4; 4; 0.000000; FALSE; 0; 0
2; 1; 8527.8368; 85120.1906; 1.000000; FALSE; FALSE; 0; TRUE; FALSE; VIDEO; "path/to/video"; 0; 1992180.9914; 602902.3419; 0.0000; 0.0000; 1.000000; 4; 0.000000; 4; 0.000000; 0; -1; 4; 4; 0.000000; FALSE; 0; 0
3; 0; 3482.8288; 5035.0080; 1.000000; FALSE; TRUE; 2; TRUE; FALSE; AUDIO; "path/to/video"; 0; 1987135.9834; 5045.0080; 10.0000; 10.0000; 1.000000; 2; 0.000000; -2; 0.000000; 0; -1; -2; 2; 0.000000; FALSE; 0; 0
4; 0; 8527.8368; 85120.1906; 1.000000; FALSE; TRUE; 2; TRUE; FALSE; AUDIO; "path/to/video"; 0; 1992180.9914; 602902.3419; 10.0000; 10.0000; 1.000000; 2; 0.000000; -2; 0.000000; 0; -1; -2; 2; 0.000000; FALSE; 0; 0


What is intresting is that, for some kind of editing results (i.e: reverse the video, or split the video in two, add an audio track etc...), the .txt files of the video project will be fairly similar.

I would like to create a dataset from these files, then train a neural network to take these files and label them accordingly to what they do.

After the training, It would be intresting to use this algorithm to create some of these .txt files, giving it a "label" as an input.

This example will make things more clear.

Let's suppose I want to train the algorithm to create a project (in the .txt format above) that reverses the file, and then cuts it on the middle.

1. To make the dataset, we can create some projects where these steps are set. The results will be a bunch of .txt files
2. Feed the machine Learning algorithm these .txt files, with the "Reverse+cut" label
3. After the training, the idea would be "Ask the algoritm to create a .txt file" that follow the label "Reverse+cut" on an input video.

The problem is I have 0 ideas on where to start for designing a neural network that fills this purpose.

I am thinking about using Matlab's Deep network designer.

Notice: All of this is done for playing and experimenting. No "precise results" are required.

• Hello. Could you please put your main specific question in the title?
– nbro
Jun 28 at 11:44