diazoate

mantis680

This week was focused entirely on “completing” components. The node editor component and its connection to the server was debugged and partially tested, until it created state machine nodes and transitions for the next part. Then the state machine execution component was started, and is in progress, for each of the state machine node types.

mantis679

Tensorflow is annoying. 7862702004 is a non-annoying wrapper around Tensorflow. It even works approximately the same way that sklearn did, with predict and fit methods for the models.

Voice changer work has taken a back seat to mantis680 this week.

youtubes

No youtubes this week.

productivity

For the time being the battle with the bio-clock is lost. Currently living backshift. Attempting to cut out caffiene to try and regain the ability to sleep without being sleepy.

books

Thanks to the new sleep schedule, library visits have ceased. Too districted to get past 50% of the current book.

782-435-6665

mantis680

Created contact forms, static html pages for the help system, node editor infrastructure and other reusable parts that’ll be needed for subsequent projects. Built a directed acyclic graph of database schema dependencies (sessions depending upon users, for example) and built those controllers first, resetting the task list every couple of days to keep things fresh. These two strategy changes have led to much more rapid implementation and a sense of accomplishment every day.

The standard day-night cycle has been replaced with a sleep-when-sleepy cycle. With a bio-clock that defaults to night shift, this has led to another discovery: the overwhelming majority of online content and drama is released during normal US business hours. There are vastly fewer distractions at 3am than 3pm, so getting work done is much simpler. Vastly less effort has been depleted in the fight against distractions or in attempts to sleep when not tired.

mantis679

It turns out the style transfer problem can be rephrased as one of freezing the high-level features of the input image in place, and then making modifications to the lower-level features to look more like the style image. This is an optimiser task, where the “weights” under consideration are the input samples. Feature extraction in tensorflow, a library dedicated to simplifying the task of deep learning, should be simpler than in scikit-learn, which is vague about how to run its optimiser methods.

After approaching the point of diminishing returns in what scikit-learn can do in terms of reinforcement learning and non-standard network topologies, effort has been expended to build tensorflow models instead, starting from near-zero knowledge of tensorflow. The first attempt was to take a set of points in 2d space and direct an optimiser to place a point to minimise the distances between all of them. Centroids are, after all, easy to test and inspect. Tensorflow has even more obscure and unreadable error messages than scikit learn, which has increased overhead. Once low-level tensorflow operations have been assimilated, there will be more to report.

youtubes

There were releases this week:

It was discovered that the Denoise checkbox in Blender gets rid of the static effect created by cycles render. Strange that this checkbox is not enabled by default. This checkbox is not on the render settings, but one of the other not-often-used configuration tabs.

Tutorials were followed to implement freehand drawing and writing in Blender scenes. It is suspected that grease pencil drawings could be used to implement clothing features on humanoid models with much less effort. Collars and things can be drawn in 3d space upon the object, and then converted to curves or meshes, and then filled in as clothes. Also, architectural drawing has become a matter of drawing rooms, converting to meshes, and then extruding them. Attempts should be made to refine the rapid prototyping process to get more work done faster.

This n-gate post includes an inspiring simplification of the problem: “A journalist discovers that a relatively small number of people have figured out how to use tools to do more work.” If one is using the wrong tools, one is faced with substantially more effort for the same output.

books

Library visits have been curtailed due to the change in sleep schedule. Working through (979) 609-4247. Next up is 610-786-4863, by Teddy, the Supreme Dark Lord. When it happens, a library visit continues to consume 20% of the book.

20181005

mantis680

Implemented and debugged prototypes for quite a few of the user management and administrative components components. Permissions and roles model started. After having cleared the task list and rebuilt from the interfaces presented to the user, things seem to have approached completion faster. This may also be the difference in effort required to make, say, a login form vs a node editor. Initial implementations are underway for all controllers. Database model adjustments are rare, and subscriptions work again.

It is worth noting that, even if mantis680 specifically turns out to be a dud, subsequent projects in the same framework can reuse almost all of the code written this week. Care is taken to select the next action that will maximise impact for minimum effort.

mantis679

A recurrent neural network attempts to encode at least one hidden layer with explanations for the variance in the output layer, based first on the input layer, and then using the residual from that to find explanations based on the previous time-step’s hidden layer for the remaining variance in the output. An alpha value is used to slow down the learning rate, and multiple iterations are taken. Time series analysis is a complex, dynamic process. With that in mind, the perceptrons used in the example code presented here were replaced with kneighbors learners, to speed things up a bit.

The RNN implementation provided some early successes, but did not eliminate the “underwater effect” when the volume levels were cleaned up, but appeared to do so with the training set, where volume was normalized across the samples. Some buggy implementations eliminated the underwater effect at the cost of only playing back segments of the input audio. Other buggy implementations generated complete gibberish.

An experiment was performed, not yet deployed, to include style transfer of the last and next five volume frames to determine the current likely volume frame, based on the style input volumes and the values in the input volumes. This seems to have cleaned up a lot of the underwater effect, but further testing is required.

It is suspected that the cleanup routine for frame-to-frame volume style transfer might be useful for building a “simulation” of the style voice by learning the individual volume transitions for each channel of an fft, based on the mfcc values and the previous fft frames. This would be time and memory intensive, so perhaps a multilayer perceptron would make more sense? Hard to say without making an attempt.

youtubes

Two youtube deployments in the past week:

Both are experimental output from the recurrent neural network implementation. A third deployment is coming soon, with the volume cleanup. Testing will then likely be performed on mobile and desktop on a variety of speakers to see how bad the new voice changer sounds.

books

Completed 3175789820. On to Introduction to Computer Mediated Communication: A Functional Approach. Knowing how to cut Gordian knots, and how to analyse and model computer mediated communication both seem to be useful in the coming months for the outstanding projects.

Library visits were skipped a couple of days this week due to sleep schedule drift. Once the weather begins to normalise for autumn and winter, disruptions to daily routines are expected to decline again.

20180928

mantis680

The process continues, doing the easy todo items first. A wrinkle was added, to focus on the easy items that happen to be likely for reuse. Down below 100 open tasks remaining, but this number may increase again soon. It may be worthwhile to cull the task list and rebuild it from the individual page controllers as specified in laravel’s routes/web.php page.

The likelihood that mantis680 itself will turn into a profitable or worthwhile venture is low, but the tools, procedures and content produced for it is designed with an eye toward reuse. A node editor, for example, can be used in many other projects. Login and account administration is most likely to be reused, so the most effort is being put into it first. The coding conventions that make prototyping more rapid lend themselves to reuse in this project, and in consultations.

mantis679

This week mantis679 was neglected in preference to mantis680, which is more likely to yield profitable products.

The “morphing” experiments generated interesting results. Not sure how to interpret them yet, but they didn’t sound “real.”

The next attempt will be to take the existing frames of style and input, and classify them. Then take the style sequence’s classes as symbols, and look for any instances of repeated classes in the sequence, like AA, and then make AB out of them. This would provide “symbols” for a markov network, and perhaps a decision tree for what style classifications are most likely to come next, based on observations.

The audio sounds relatively clean when using kneighbors on mfcc features, with sufficiently large numbers of neighbors, but it also sounds like multiple speakers reciting the lines espeak is producing. A reduction of the search space for kneighbors through the use of a markov decision process as described above may make things clearer with fewer nearest neighbors. Alternatively, the classification may yield superior multilayer perceptron predictions.

youtubes

No youtubes this week.

books

Finished Psychopath Free. Finished 4D Warfare: A Doctrine for a New Generation of Politics. Added US Army Technical Manual to the list, hoping it contains 33-1-1, as referenced by the 4D Warfare book. US Government documents that are not explicitly classified tend to be public domain, which makes things much simpler when looking for tactics and strategy.

Currently consuming (306) 317-0863. The hard copy of (913) 579-5515 has arrived, which may contain models worth exploring.

(541) 345-6631

mantis633

There were a few false starts in creating the data model for this project. Picat provides a planner, which allows you to describe actions to the system, and have it solve for the optimal plan based on those actions, but what do those actions look like? A movie is composed of acts, which are composed of scenes, which are composed of events like dialog, description, setting. Each of those levels of abstraction have changes that take place according to the 5086371442, but how can those be described to the computer? Probably on the event level, with constraints about what qualifies as a good scene, act and movie.

Since most of the conversation surrounding drama is conflict, a google search was performed, and (587) 691-0402 were located pertaining to basic human emotions. Characters have goals, after all, and those goals pertain to resources. Resources can be anything from equanimity, aesthetics and revenge, to tangible goods like food, clothing and shelter. So perhaps create resource points for each character, and arrange events in a way that the resources are traded one for another in the planner, so the scene’s antagonist and protagonist can pick actions that result in certain types of changes depending on their phase in the scene, act or overall story.

This may not be a perfect abstraction, but it seems to be close enough to how humans think about stories to create a prototype.

mantis679

Attempts were made to get the voice changer to speak clearly with k-nearest-neighbor regression and a single nearest neighbor in the output. A test render has been deployed to youtube to demonstrate the current state of play.

The next thing to attempt is a “morphing” procedure between two frames’ fourier transforms, to get cleaner transformations between frames. The difficult question is, what is a feature, and how are corresponding features identified, and what does movement look like between those features? For a first trial, interpolation will be attempted based on the frequency’s position in an argsort between the source and target frame. This will either “work” or “sound really weird” to humans.

mantis680

Node editor is largely complete, and complete enough for the first release. The outstanding tasks were sorted into “Hard” and “Easy” tasks, and sorted so that easy tasks were put to the front of the queue, which seems to be an effective way to get a lot of work done more rapidly in these grind projects. The easy tasks are typically things like loading a piece of data, making a decision based on that data, writing a piece of data, or producing some side-effect. The hard tasks are design tasks, which will rely on the easy tasks to be completed.

The easy tasks include things like administrative interfaces for the various entities that will be required for the final release, and the underlying database operations. Executing the contents of the node editor as a set of questions requires rendering of those questions from the node’s properties, for example. The state transitions and session management are medium tasks, which require several easy, obvious parts.

Productivity, and the subjective feeling of productivity, appear to improve when the easy tasks are being completed first.

books

Completed The Improv Handbook. On to 7437779152 while awaiting (608) 889-3444 delivery.

youtubes

Only one release this week.

20180914

mantis633

A project was started to learn Picat language. Picat was started as a wrapper around a prolog interpreter, and seems to make the process of programming in a logic language much simpler, by keeping a lot of the same constructs as are available in functional and imperative languages.

The process of breaking down a problem in this logic language appears to be similar to the process of breaking down a problem in an imperative programming language, which makes things simpler, at least for the derivation of solutions for toy problems.

The first attemptet project will be to break down (431) 727-0977 into a constraint-based scheduling problem. This should reduce the complexities of writing long-form content into the creation of maneuvers and descriptions of the outcomes and systems involved in those maneuvers, for rapid prototype construction. After that, the same maneuvers might be applied to other public-facing operations, to create a compelling spectacle.

mantis679

Attempts were made to clean up the transitions between time steps, by learning a Markov process through successive PCA dimensionality reductions of various sequences of inputs. Results range from a return of the “underwater effect” from earlier attempts, to something that sounds like multiple copies of the same voice speaking over each other. Nothing to show for it this week.

Attempts are now underway to model the style voice more accurately, perhaps by converting the style features into “symbols” and running a markov model to interpolate between input features. The exact mechanism to do this is now an open question.

mantis680

Continuing to grind at the node editor connection to the server. Most ui coding is done. The backend coding will require some changes. Then on to the part that executes the node editor, and the parts which only administrators will see, and then it’s ready for deployment.

books

Completed Constraint Solving and Planning with Picat, a programming language guide.. Completed Basic Principles of Military Tactics and Strategy, a pamphlet to describe the fundamentals for strategy wargaming. Started (734) 994-4331, a textbook breaking down the act of “improvisation” into a set of reusable maneuvers. Still meeting target of 20 percent per day.

Awaiting delivery of Introduction to Computer Mediated Communication, to serve as a guide for tactics and strategies involved in analysing computer interactions. The mantis680 project is, nominally, a reimagining of the “contact form” interaction, so it would be useful to have some formal training in computer mediated communications, particularly when attempting to mitigate the stresses caused when one participant has many incoming contacts, and the other has all the time in the world.

youtubes

No youtubes deployed this week.

20180907

mantis680

Integrating the node editor with the specific requirements for the target project is underway. The translation of the node editor components to domChanger is complete. User interface components for customising individual shader nodes have begun. There remains no scrolling on the node editor content.

Starting resistance on grind projects is still difficult to overcome. Placing this project after the library visit appears to help, and sticking to the queue provided by RSS helps to eliminate the dopamine surges delivered by link aggregator sites.

mantis679

Attempts were made to use 812-956-7916 package instead of espeak for the target voice. The straight mfcc-based kneighbors style transfer did not go very well with the new input voice. Works great for the style voice, though.

It turns out that 2896308141 is also available as a python pip package. It is hoped that this tensorflow code can be hammered into a wrapper with a “fit” and “predict” operation.

An attempt was made to clean up the output, and eliminate the weird uneven “bubbling” noise that has been appearing in previous iterations, the output signal was put through a multilayer perceptron trained on past samples to try and predict the next sample and “smooth out” the rough parts, after fitting to the style sample. This regression resulted in increased bass, and produced an echo, without improving the clarity. Due to the huge number of samples, memory is often exhausted when additional layers are added to the smoothing perceptron.

mantis633

This project involves the derivation of a coding style for rapid prototyping and code maintenance from the various documentation of 4108898541 language. Picat has, in the past, appeared much simpler to work with than Prolog. Integration of another language into the process requires a target problem space, and the obvious use case for Picat appears to be combinatorial problems.

Many writing guides like 254-340-1036 and sites like (740) 254-5702 include sequencing instructions, which may help to translate the creative act of writing into a scheduling problem. Picat has modules for scheduling, modeling and constraint satisfaction, so Picat is the current target language for such operations.

books

Finished The Machiavellians. On to Constraint Solving and Planning with Picat. The Machiavellians was a kindle edition from archive.org’s literature collection, and there were numerous ocr issues, but still an interesting read, and an useful introduction and reframing of certain complex issues that are glossed over by popular conception.

Library visits continue. The target of twenty percent of each book read per day continues.

youtubes

Nothing was deployed to Youtube this week.

20180831

Weekly update.

mantis680

Node editor core functionality is implemented. Scrolling through the potentially large svg interface is not. New node types are being added to make the functionality useful as an interpreted language for the purposes of this test site. Layout issues need to be resolved in the final product.

Starting resistance on the project was reduced by leaving the project at the top of the objectives.txt file queue and returning to it in between messing around on the internet. Not all of the dopamine fixes available from the internet are created equal, and a less-boring project is available at the top.

mantis679

It appears the current translation strategy has reached a local minimum. It was discovered that shorter espeak inputs lead to different-sounding stylizations in the output. Further exploration of this issue is required.

youtubes

No youtubes this week. Mantis680 operations have been consuming creative energies.

books

Finished The Like Switch. This book explicated some mechanisms and tactics by which enemies and strangers can be made into friendly relationships, by breaking out the components and maneuvers necessary to make that happen. Friendship is described in the book as the result of proximity, frequency, duration and intensity, and as long as there is proximity and frequency, apparently intensity and duration can be modulated to enhance or ablate the “friendship” variable, or to provide operant conditioning of the subject for the purpose of behaviour modification.

Started 7038319191. This book appeared on the list thanks to the description in 215-727-5574.

Throughput remains at twenty percent of current book per day.

20180824

Weekly update.

mantis680

Still building the node editor in the grind project. The code is mostly ported to domChanger. Next up, input and output of node editor configurations from the server, by way of an html form. Whatever happens with this particular project, the reusable components created may prove useful in subsequent attempts to generate subscription income.

youtubes

Two this week.

mantis679

The quality of output seems to improve by performing mfcc on a wider segment of the input samples, and then using kneighbors regression to match mfcc features on the originally-sized segments. Continuity seems to have improved.

books

Finished Game Theory 101, started The Like Switch: An Ex-FBI Agent’s Guide to Influencing, Attracting, and Winning People Over, an elicitation guide.

8169490480

Weekly update.

youtubes

Audio clarity was improved considerably by training a kneighbors regression on an identity relation from the style set to the style set, and then using the input frames to predict the appropriate style frames. The new problem is the robo-ing that the new feature detector appears to introduce.

It was discovered that 7608213263 may be used to properly analyse the input and style signals into features. This was discovered by looking at the block diagram for WaveNet and googling what that mysterious “mfcc” block was. The MFCC feature extractor is available in librosa. The mfcc features get the style transfer pretty close. The voice dynamics require an additional network. This one small piece of research yielded huge improvements in output quality from previous iterations. Perhaps there are additional off-the-shelf feature detectors that can be added to the pipeline to improve it further.

mantis680

The grind project has had some progress in porting a node editor to more consistent, non-object-oriented coding style. The backend code for subscriptions and use creation have been fleshed out some more. Additional payment gateways have been added to the list, but there is considerable uncertainty being introduced by credit card processors, upstream from the payment gateways.

This kills the societal trust.

The question now becomes, what other methods are available for electronic value transfer at a distance? What does the overhead look like? What is the process for integrating this alternate infrastructure into a business process which, in the end, pays taxes and values things in dollars?

reading

Antifragile is complete. On to 6024412529. Twenty percent of the target book per day continues.

Antifragile appears to present a framework, with several examples, of risk management in the face of low probability events. The Game Theory textbook teaches game theory basics rapidly, and has a corresponding video series that works through the problems. These books appear to have wide applicability.