Rebecca Fiebrink Perry Cook, Advisor FPO, 12/13/2010

Real-time Human Interaction with Supervised Learning Algorithms for Music Composition and PerformanceRebecca FiebrinkPerry Cook, AdvisorFPO, 12/13/2010

function [x flag hist dt] = pagerank(A,optionsu)[m n] = size(A);if (m ~= n) error('pagerank:invalidParameter', 'the matrix A must be square');end; options = struct('tol', 1e-7, 'maxiter', 500, 'v', ones(n,1)./n, … 'c', 0.85, 'verbose', 0, 'alg', 'arnoldi', … 'linsys_solver', @(f,v,tol,its) bicgstab(f,v,tol,its), … 'arnoldi_k', 8, 'approx_bp', 1e-3, 'approx_boundary', inf,… 'approx_subiter', 5);if (nargin > 1) options = merge_structs(optionsu, options);end;if (size(options.v) ~= size(A,1)) error('pagerank:invalidParameter', … 'the vector v must have the same size as A');end;if (~issparse(A)) A = sparse(A);end;% normalize the matrixP = normout(A);switch (options.alg) case 'dense’ [x flag hist dt] = pagerank_dense(P, options); case 'linsys’ [x flag hist dt] = pagerank_linsys(P, options) case 'gs’ [x flag hist dt] = pagerank_gs(P, options); case 'power’ [x flag hist dt] = pagerank_power(P, options); case 'arnoldi’ [x flag hist dt] = pagerank_arnoldi(P, options); case 'approx’ [x flag hist dt] = pagerank_approx(P, options); case 'eval’ [x flag hist dt] = pagerank_eval(P, options); otherwise

error('pagerank:invalidParameter', ...

'invalid computation mode specified.');

function [x flag hist dt] = pagerank(A,optionsu)

useful

usableeffective, efficient, satisfying

Machine learning

algorithms?

Outline

• Overviews of interactive computer music and machine learning

• The Wekinator software• Live demo• User studies• Findings and Discussion• Contributions, Future Work, and Conclusions

interactive computer music

Interactive computer music

sensed action

interpretation

response (music, visuals, etc.)

computer

human with microphone, sensors, control interface, etc.

audio synthesis or processing,

visuals, etc.

Example 1: Gesture recognition

sensed action

identification

response

computer

Bass drum:

“Gesture 1”

Example 1: Gesture recognition

sensed action

response

computer

Bass drum:

Hi-hat“Gesture 2”

identification

Model of sensed action to meaning

sensed action

response

computer

meaning

16 computer

Example 2: Continuous gesture-to-sound mappings

sensed action

interpretation

sound generation

computer

mapping

human + control interface

Example 2: Continuous gesture-to-sound mappings

A composed system

sensed action

mapping/model/

interpretation

response

mapping/model/

interpretation

supervised learning

algorithm

trainingdata

Training

Supervised learning

inputs

outputs

algorithm

trainingdata

Training

Supervised learning

inputs

outputsRunning

“Gesture 1” “Gesture 2” “Gesture 3”

“Gesture 1”

Supervised learning is useful

• Models capture complex relationships from the data and generalize to new inputs. (accurate)

• Supervised learning circumvents the need to explicitly define mapping functions or models. (efficient)

Has been demonstrated to be useful in musical applications, but no usable, general-purpose tools exist for composers to apply algorithms in their work.

Criteria for a usable supervised learning tool for composers

1. General-purpose: many algorithms & applications

2. Runs on real-time signals

3. Compatible with other compositional tools

4. Supports interaction via a GUI

Weka and similar

general-purpose tools

✓✓

✓(some)

Existing computer

music tools

✗✓

✓(some)

✗ ✗ ✓5. Supports appropriate end-user interactions with the supervised learning process

✗✗

Criteria for a usable supervised learning tool for composers

Weka and similar

general-purpose tools

✓✓

✓(some)

Existing computer

music tools

✗✓

✓(some)

✗ ✗ ✓5. Supports appropriate end-user interactions with the supervised learning process

✗✗

Interactively…• create training data• evaluate models• modify training data

Interactive machine learning (IML)

• Fails and Olsen (2003): training set editing for computer vision systems

• More recent work:– application to other domains

(Shilman et al. 2006; Fogarty et al. 2008; Amershi et al. 2009; Baker et al. 2009)

– other types of interactions, e.g. manipulating confusion matrices(e.g., Talbot et al. 2009, Kapoor et al. 2010)

Research questions

•Which interactions are possible and useful?•What are the essential differences between IML and conventional ML?

•How can IML be used in real-time and creative contexts?

Outline

The Wekinator

• Built on Weka API• Downloadable at http://code.google.com/p/wekinator/

(Fiebrink, Cook, and Trueman 2009; Fiebrink, Trueman, and Cook 2009; Fiebrink et al. 2010)

✓✓✓✓

5. Supports appropriate end-user interactions with the supervised learning process

Running models in real-time

model(s)

.01, .59, .03, ....01, .59, .03, ....01, .59, .03, ....01, .59, .03, ...

5, .01, 22.7, …5, .01, 22.7, …5, .01, 22.7, …5, .01, 22.7, …

Feature extractor(s)

Parameterizable process

Real-time, interactive design

The Wekinator supports user interaction with all stages of the model creation process.

3.3098 Class24

Under the hood

Model1 Model2 ModelM

Feature1 Feature2 Feature3 FeatureN…

Parameter1 Parameter2 ParameterM

joystick_x joystick_y

pitchvolume

webcam_1

3.3098 Class24

Under the hood

Model1 Model2 ModelM

Feature1 Feature2 Feature3 FeatureN…

Parameter1 Parameter2 ParameterM

Learning algorithms:Classification:

AdaBoost.M1J48 Decision TreeSupport vector machineK-nearest neighbor

Regression:MultilayerPerceptron

Tailored but not limited to music

The Wekinator• Built-in feature extractors for music & gesture• ChucK API for feature extractors and synthesis

classes

Other modules for sound synthesis,

animation, …?

Other feature extraction modules

Open Sound Control (UDP)

Outline

Study 1: Participatory design process with composers

• Method:– 7 composers– 10 weeks, 3 hours / week– Group discussion, experimentation,

and evaluation– Iterative design– Final questionnaire

• Outcomes:– Composers focused on instrument-building, 2 ultimately used

in performance– Much-improved software and lots of feedback

(Fiebrink, Trueman, et al., 2010)

Study 2: Teaching interactive systems building in an undergraduate course

• COS/MUS 314 Spring 2010 (PLOrk)• Focus on building interactive music

systems• Method:– One component of the midterm assignment:

students building interactive systems using supervised learning

– 1 continuous + 1 discrete system– Logging + short answer questions

• Outcomes:– Successful project completion, use in midterm and final

performances– Logs from 21 students

Study 3: Bow gesture recognition

• Work with a composer/cellist to build gesture classifiers for a commercial sensor bow, the “K-Bow”

Study 3: Bow gesture recognition

• Work with a composer/cellist to build gesture classifiers for a commercial sensor bow

• 8 classifiers for standard bow gestures:e.g., up/down bow (2 classes), articulation (7 classes)

• Method:– All classification tasks defined by cellist and addressed in order of her

choice• All 8 tasks in 1st session, 5 tasks in 2nd session

– Logging, observations, final questionnaire– Cellist assigned each iteration’s classifier a quality rating (1 to 10)

• Outcome:– Successful classifiers created for all 8 tasks (rated “9” or “10”)

Study 4: Composer case studies

• Clapping Music Machine Variations (CMMV) by Dan Trueman, faculty

• CMMV by Dan Trueman, faculty• The Gentle Senses / MARtLET by Michelle Nagai, graduate student

• CMMV by Dan Trueman, faculty• The Gentle Senses / MARtLET by Michelle Nagai, graduate student• G by Raymond Weitekamp, undergraduate

Outline

Discussion of Findings

1. Users took advantage of interaction in their work with the Wekinator.

2. Users employed a variety of model evaluation criteria, and subjective evaluation did not always correlate with cross-validation accuracy.

3. Feedback from the Wekinator influenced users’ actions and goals.

4. The Wekinator was a useful and usable tool.5. Interactive supervised learning can be a tool for

supporting creativity and embodiment.

An iterative approach to model-building

(mean per student for each task)

(mean per classification task)

PLOrk con-tinuous

PLOrk discrete

KBow 1st session

KBow 2nd session

Mean # Trainings

Frequent modifications to training data in-between re-trainings

Add Data Edit Data Delete Data Clear All Data

Change Learner

Change Learn

Params

Change Features

Add Data

Edit Data

Delete

Clear A

ll Data

e Learn

Change

Cont.Disc.P

Interaction and the training dataset

• Training data is an interface for key tasks:– defining the learning problem– clarifying the learning problem to fix errors– communicating changes in problem over time

• User incrementally makes model more complex• User changes his mind about how a model should work

– providing a “sketch” that the computer fills in

Thoughts: Data editing makes sense

• The training data is the most effective and appropriate interface for accomplishing these tasks.

• Potentially appropriate for many problem domains where:– User has knowledge of learning problem– User is capable of creating examples

• The interfaces for creating and editing the training data are crucial to the user’s ability to effectively and efficiently accomplish these tasks.

Playalong data recording

•Employed by– 3 of 8 composers in Studies 1 and 4– 8 of 21 PLOrk students

•These users found playalong highly useful– Allowed training data to represent more fine-grained

information– Enabled some composers to engage their musical and

physical expertise• Allowed practice and attention to “feel”

“Conventional” model evaluation

Available data

Training set

Evaluation set

Evaluate

Cross-validation repeats this process with different data partitions.

“Direct” evaluation in Wekinator

Training set Train

Evaluate

Direct evaluation used most frequently

• Composers in participatory design and case studies: only direct evaluation

• KBow and PLOrk:

Cross-val. Acc.

Train. Acc. Direct Eval.0

PLOrk cont.PLOrk disc.KBow 1KBow 2

Roles of cross-validation and training accuracy

• K-bow: Cross-validation used to quickly and objectively compare different feature selections and learning algorithms

• PLOrk:– Treated as reliable evidence a model was performing

well– Used to validate the user’s own ability

Roles of direct evaluation

• Used to assess behavior of the model against subjective criteria

• Used to obtain feedback that shapes the users’ future interactions with the system

Accuracy measured by direct evaluation

• Important in building discrete gesture classifiers– Accuracy = models outputs are correct according to

learning concept definition

• Accuracy still important for open-ended instrument-building tasks– Accuracy = matching a user’s expectations, especially on

inputs like the training examples

Other criteria

•For discrete classifiers:– Cost: consequences and locations of model errors– Decision boundary smoothness

•For continous mappings:– Complexity, difficulty – Unexpectedness and surprise– “Feel”

Subjective evaluation criteria & CV

• K-bow:– Cross-validation sometimes correlates with subjective

quality, but sometimes it doesn’t!

Task: Horizontal Position

Vertical Position

Bow Direction

On/Off String

Speed Articulation

R: -0.59 -0.44 -0.74 -0.50 0.65 0.93

Pearson’s correlation for tasks with > 3 iterations:

Vertical Position

Bow Direction

On/Off String

Speed Articulation

R: -0.59 -0.44 -0.74 -0.50 0.65 0.93

Vertical Position

Bow Direction

On/Off String

Speed Articulation

R: -0.59 -0.44 -0.74 -0.50 0.65 0.93

Thoughts: Is generalization accuracy important?

• Yes!– Human and environmental variations are inevitable

• BUT it may not be the only or most important factor• Generalization estimated from the training set (e.g.,

using cross-validation) is not always informative• Implies that models designed for human use should

be evaluated by human use.

Thoughts: Choosing or designing better algorithms• Most algorithms’ training process aims for a model with

good generalization• BUT the user is employing the training data as an

interface– User relies on good training accuracy to influence the model’s

behavior

• Better algorithms might– Privilege training accuracy (e.g., k-nearest neighbor)– Provide parameters for interactive improvement against other

subjective criteria (e.g., using regularization parameter for boundary smoothness)

Interaction involves control and feedback

control

feedback**Running the models (“direct evaluation”)Cross-validation and training accuracy

Machine learning

algorithms

Running models informs future actions

• For example:– locate errors add correctly-labeled examples– detect total failure delete all the data

Running models trains users to be more effective supervised learning practitioners

•Users especially learned to create better training datasets– Minimize noise– Balance the number of examples in each class– Vary examples along all the dimensions that might vary

in performance

• Important for novice users

Running models informs users’ goals for machine learning• Some PLOrkers and composers relied on the Wekinator to

help formulate the problem definition– composers liked being inspired by mapping behaviors

• Users learned what was most easily accomplished

… and exploited flexibilities in the learning concept definition to create a model that most easily met their most important goals• PLOrk: Define classes based on what’s easy to classify• K-Bow: Add more classes when a model performed well (e.g.,

Bow Speed)• A key benefit of IML!

Running models teaches users about themselves and their work

K-Bow cellist:Model’s confusion of spiccato and riccocet

realization that her spiccato was too much like riccocet

improved technique

improved classifiers

Thoughts: Impact in other domains

• Direct evaluation on user-generated inputs could offer similar advantages in other applied machine learning domains– Appropriate when user is qualified to create inputs and

evaluate results

Barriers to usability

• Long training times (a few users)• Algorithms’ inability to model the desired concept

[easily]– e.g., producing a linear function with a neural network

• Difficulty in debugging – No guidance on choosing a better algorithm or

algorithm parameters– Especially due to novice users

Usability and usefulness: Study 1 composers

Statement 5-point Likert mean (std. dev.)

“The Wekinator allows me to create more expressive mappings than other techniques.”

4.5 (.8)

“The Wekinator allows me to create mappings more easily than other techniques.”

4.7 (.5)

Usability and usefulness: PLOrk students

Statement 5-point Likert mean (std. dev.)

“I can reliably predict what sound my model will make for a given inputgesture.”

4.5 (.7)

“Wekinator eventually learned what I wanted it to.”

4.3 (.9)

“My model provides reliable gesture classifications” (discrete task)

4.9 (.2)

“My model is musically expressive” (continuous task)

4.1 (.7)

Usability and usefulness: PLOrk students

• Model-building was fast– 27.1 minutes for continuous mapping– 16.1 minutes for discrete classifier

• Students enjoyed the Wekinator– “Learning by experimentation was a lot of fun!”– “It’s so cool, the Wekinator rocks.”

Usability and usefulness: K-Bow

Task Rating (1 to 10) CV Accuracy (%)Direction 10 87.3On/Off String 10 83.5Grip 10 100.0Roll 10 98.2Horizontal Position 10 89.3Vertical Position 10 90.0Speed 9 87.5Articulation 9 98.8

• Models successfully created for all 8 tasks:

Usability and usefulness: K-Bow

Statement 5-point Likert response

“The Wekinator was able to create accurate bow stroke classiers in our work so far”

“The Wekinator was able tocreate bow stroke classiers more easily than other approaches”

“10 (so 5)"

Usability and usefulness: Case studies

create mappings more easily

Series1

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Trueman Nagai Weitekamp

Agreement (1-5)

create mappings that were more expressive

create a kind of music thatisn't possible or that is hard to

create using other techniques

approach the process of composition in a new way

The Wekinator allowed me to:

“There is simply no way I would be able to manually create the mappings that the Wekinator comes up with; being able to playfully explore a space that I've roughly mapped out, but that the Wekinator has provided the detail for, is inspiring.”

“The ability to map sound and gesture, in a very immediate and intuitive (yet unpredictable) way is really the most inspiring and useful aspect of the wekinator for me right now. I can see the possibility of building interfaces or instruments as needed, flexibly, on the fly, for different kinds of projects, and being able to quickly map them out to existing sound sets with only minor programming changes.”

Qualities important to composers

• Speed and ease of creating and exploring mappings (especially complex mappings)– IML is useful: Demonstration can be faster and more

efficient than coding.• Access to surprise and discovery– IML is useful: Neural networks fill in the details of the

training data sketch.• Balancing surprise and complexity with predictability and

control– IML is useful: Users can reliably steer model behavior

using the training data.

Creativity support in HCI

• Qualities of the Wekinator align with “creativity support tool” guidelines proposed by Shneiderman (2000, 2007) and Resnick et al. (2005)– Support exploration, discovery, and sketching– Support diverse users (e.g., novices and experts) and applications– Operate seamlessly with other [composition] tools

• IML integral to Wekinator’s realization of guidelines• Interactive supervised learning may be an effective tool

for supporting creativity in many other domains.

Embodiment is also important• Composers valued the ability to take an embodied approach

to design, which engaged their physical expertise as musicians• Designing this way also produced instruments that “felt right”

“I have never before been able to work with a musical interface … that allowed me to really ‘feel’ the music as I was playing it and developing it. The Wekinator allowed me to approach composing with electronics and the computer more in the way I might if I was writing a piece for cello, where I would actually sit down with a cello and try things out.”

Embodiment is also important• Wekinator engaged users’ physical expertise as musicians• And produced instruments that “felt right”– Users were physically engaged in the creation of the data and the

evaluation of the models (especially with playalong)• Embodied cognition: body plays an important role in learning,

understanding, and thinking• Physical knowledge, kinesthetic memory, and bodily skill are

important considerations in the design of interactive systems (Klemmer et al. 2006)

• Interactive machine learning may be useful for supporting embodied design in other real-time, interactive domains

Outline

Contributions

1. The Wekinator software and “playalong” interaction for training data creation

“Well, I had basically lost interest in the whole process of digital controller-based instrument building, so the Wekinator's very existence has enabled and inspired me to get back into the game... The Wekinator enables you to focus on what your primary sonic and physical concerns are, and takes away the need to address so many details, and it does so in such a way that even if you DID spend all the time on building the mappings manually, you would *never* come up with what the Wekinator comes up with. So, the process becomes more focused, more musical, more creative, more playful. I actually *want* to do it.”

Contributions

2. A demonstration of the important roles that interaction can play in the development of supervised learning systems…

…and a greater understanding of the differences between interactive and conventional machine learning contexts.

communicating evolving problem definitions, evaluating models against subjective criteria, training users to become better at machine learning, embodied design, …

emphasis on generalization accuracy and role of the training set, training dataset size, time constraints

Contributions

3. A better understanding of the requirements and challenges in the analysis and design of algorithms and interfaces for interactive supervised learning in real-time and creative problem domains.

enabling inspiration, supporting novice users, providing access to complexity, enabling evaluation against subjective criteria, supporting embodiment, rapid prototyping, …

useful

usable

Contributions

4. A clearer characterization of composers’ goals and priorities for interacting with computers in music composition and instrument design and a demonstration that interactive supervised learning is useful in supporting composers in their work.

Contributions

5. A demonstration of the usefulness of interactive supervised learning as a creativity support tool.

Supporting exploration, discovery, and sketchingSupporting diverse users (e.g., novices and experts) and applications

Operating seamlessly with other tools

Future work

• Improve the Wekinator as a compositional tool and a research platform– Collaborate with musicians, researchers, and other

users in participatory processes

• Investigate new algorithms, interactions, and interfaces for music performance and beyond– Further explore how interactive machine learning can

be made more useful and usable by more people applying it to more problems

Final Conclusions

• It is important to consider the human-computer interfaces and context of algorithms in practice

• Interactive supervised learning has the potential to significantly improve the usability and usefulnesss of conventional learning algorithms, and to enable application to new problems by new users

• Computer music is a fantastic area to be doing research– Musicians make great collaborators

Thanks!• Perry Cook• Dan Trueman• Dan Morris• Ken Steiglitz• Adam Finkelstein• Szymon Rusinkiewicz• Michelle Nagai• Cameron Britt• Konrad Kaczmarek• Michael Early• MR Daniel• Anne Hege• Raymond Weitekamp• All the PLOrk students

• Meg Schedel• Andrew McPherson• Barry Threw• Keith McMillen Instruments• Ge Wang• Jeff Snyder• Xiaojuan Ma• Sonya Nikolova• Matt Hoffmann• Merrie Morris• Sumit Basu• Ichiro Fujinaga

• National Science Foundation GRFP

• Francis Lathrop Upton Fellowship

• National Science Foundation grants 0101247 and 0509447

• The Kimberly and Frank H. Moss '71 Research Innovation Fund

• The David A. Gardner '69 Magic Project

• The John D. and Catherine T. MacArthur Foundation

• Everyone else I’m forgetting

Related publications• Fiebrink, R. 2006. An exploration of feature selection as an optimization tool for musical genre

classification. Master’s thesis, McGill University. • Fiebrink, R., P. R. Cook, and D. Trueman. 2009. “Play-along mapping of musical controllers.” Proc.

International Computer Music Conference.• Fiebrink, R., M. Schedel, and B. Threw. 2010. “Constructing a personalizable gesture-recognizer

infrastructure for the K-Bow.” International Conference on Music and Gesture (MG3).• Fiebrink, R., D. Trueman, C. Britt, M. Nagai, K. Kaczmarek, M. Early, M.R. Daniel, A. Hege, and P. R.

Cook. 2010. “Toward understanding human-computer interactions in composing the instrument.” Proc. International Computer Music Conference.

• Fiebrink, R., D. Trueman, and P. R. Cook. 2009. “A meta-instrument for interactive, on-the-fly learning.” Proc. New Interfaces for Musical Expression.

• Fiebrink, R., G. Wang, and P. R. Cook. 2007. “Don't forget the laptop: Using native input capabilities for expressive musical control.” Proc. International Conference on New Interfaces for Musical Expression.

• Fiebrink, R., G. Wang, and P. R. Cook. 2008. “Support for MIR prototyping and real-time applications in the ChucK programming language.” Proc. International Conference on Music Information Retrieval.

• Wang, G., R. Fiebrink, and P. R. Cook. 2007. “Combining analysis and synthesis in the ChucK programming language.” Proc. International Computer Music Conference.

References

• Amershi, S., Fogarty, J., Kapoor, A., and Tan, D. 2010. “Examining Multiple Potential Models in End-User Interactive Concept Learning.” Proc CHI 2010.

• Baker, K., A. Bhandari, and R. Thotakura. 2009. “Designing an Interactive Automatic Document Classification System.” Proc. HCIR 2009, pp. 30–33.

• Fails, Jerry, and Dan Olsen. 2003. “Interactive machine learning.” Proc. IUI, pp. 39–45.• Fels, S. S. and G. E. Hinton. 1993. “Glove-Talk: A neural network interface between a data-glove and a

speech synthesizer.” IEEE Trans. on Neural Networks, vol. 4.• M. Lee, A. Freed, and D. Wessel. 1992. “Neural networks for simultaneous classification and parameter

estimation in musical instrument control.” Adaptive and Learning Systems, vol. 1706, pp. 244-55.• Raphael, Chris. 2001. “A probabilistic expert system for automatic musical accompaniment.” Journal of

Computational and Graphical Statistics, vol. 10, no. 3, pp. 487-512.• Shneiderman, B. 2000. “Creating Creativity: User interfaces for supporting innovation.” ACM Trans. CHI, vol.

7, no. 1, pp. 114–138.• Shneiderman, B. 2007. “Creativity support tools: Accelerating discovery and innovation.” Comm. ACM vol.

50, no. 12, Dec. 2007, pp. 20–32.• Witten, I., and E. Frank. 2005. Data Mining: Practical Machine Learning Tools and Techniques, 2nd ed. San

Francisco: Morgan Kaufmann.

Training set size – why so small?

• Learning concepts were “easier” ? (i.e., lower sample complexity)

• Users learned to provide the most useful training examples for representing the problem?– like active learning, but the user is in charge

• Users defined the learning concept definition in order to negotiate the tradeoffs between what they wanted and what was possible in a given amount of time to create training data and train the algorithms?

Running models enables users to practice employing them more effectively

• Through practice, they learn to use models more effectively• Users accepted or expected the need to adapt their behaviors

An HCI view on algorithms

• Algorithms afford certain possible interactions, control, and feedback – i.e., they have an innate potential to be useful

• User interfaces can hide or expose these affordances– And can expose them in more or less usable ways

• The Wekinator exploits the fact that supervised learning models can be manipulated through the training dataset

• Algorithms can be made more useful and usable– through more appropriate interfaces– through affording more appropriate interactions, control, and

feedback

Rebecca Fiebrink Perry Cook, Advisor FPO, 12/13/2010

options case gs x flag

endfunction x flag hist

case dense x flag hist

search algorithms

machine learning algorithms

supervised learning

computer human

squareend options

Documents

FPO - Vizio

FPO: Maria Mendoza

Rebecca Fiebrink Perry Cook, Advisor FPO, 12/13/2010

Lalit FPO Processes 04.10

Tema 3 FPO

Amortized Analysis Explained (by Rebecca...

Introduction about FPO

FPO - archive.org

Cursos FPO Málaga

Presentacion power point fpo

FPO - ViewSonic

Bios Curso Fpo

Indiana republican.pdf · Perry B, Perry D, Perry H, Perry....

TAMIL NADE FPO

Basic Guia Fpo

Manual da FPO Magnética