path: root/helps/help-ann_td.pd

#N canvas 117 9 858 468 12;
#N canvas 445 134 647 348 creation 0;
#X obj 52 235 outlet;
#X text 246 38 specifying all;
#X text 159 216 TIP:don't set the layers param too high;
#X msg 49 10 create 2 1 5;
#X text 175 6 create with 2 inputs \, 1 output and 5 frames;
#X msg 59 38 create 2 1 5 3 3 1 0.7;
#X text 159 179 params: num_input \, num_output \, frames \, num_layers
\, num_neurons_hidden \, connection_rate \, learning_rate;
#N canvas 218 152 650 413 what 0;
#X text 37 134 you pass [0 0.1] to ann_tdnn;
#X text 34 152 internally now there is this array: [0 0.1 0 0 0 0]
;
#X text 38 196 next input is [0.2 1];
#X text 36 211 internally now there is this array: [0.2 1 0 0.1 0 0]
;
#X text 37 255 next input is [0.3 0.4];
#X text 35 270 internally now there is this array: [0.3 0.4 0.2 1 0
0.1];
#X text 36 317 next input is [0.7 0];
#X text 34 332 internally now there is this array: [0.7 0 0.3 0.4 0.2
1];
#X text 35 168 a normal ann_mlp is run with this inputs;
#X text 38 225 a normal ann_mlp is run with this inputs;
#X text 33 284 a normal ann_mlp is run with this inputs;
#X text 33 347 a normal ann_mlp is run with this inputs;
#X text 12 139 1);
#X text 14 197 2);
#X text 15 258 3);
#X text 13 319 4);
#X text 33 4 this implementation od tdnn is simply a normal ann_mlp
with num_input*frame inputs and num_output outputs. ann_tdnn simply
helps managing the delay \, frames and buffers.;
#X text 65 385 ...and so on...;
#X text 34 64 frames can be seen as the delay feedback: how many times
an input is internally held in the input array;
#X text 35 104 eg: 2 inputs 3 frames = internally 6 inputs;
#X restore 155 109 pd what frames are?;
#X connect 3 0 0 0;
#X connect 5 0 0 0;
#X restore 93 68 pd creation examples;
#N canvas 136 60 728 356 run 0;
#X obj 90 219 outlet;
#X msg 123 69 0 1;
#X msg 124 92 1 0;
#X msg 125 115 1 1;
#X msg 126 140 0 0;
#X text 40 17 now you can run your nn passing it a list with inputs
;
#X text 169 70 send a list of data and watch the console for output
;
#X text 39 35 the output is sent as a list of float;
#X text 184 134 these inputs are good for a nn like the one in example1
directory;
#X connect 1 0 0 0;
#X connect 2 0 0 0;
#X connect 3 0 0 0;
#X connect 4 0 0 0;
#X restore 107 180 pd run the net;
#N canvas 0 0 619 610 other 0;
#X obj 43 401 outlet;
#X msg 102 37 train;
#X msg 103 63 run;
#X msg 152 37 setmode 0;
#X msg 153 63 setmode 1;
#X text 249 40 set training/running mode;
#X text 247 63 training mode currently not implemented;
#N canvas 265 255 690 335 training 0;
#X obj 71 288 outlet;
#X msg 82 195 FANN_TRAIN_INCREMENTAL;
#X msg 82 216 FANN_TRAIN_BATCH;
#X msg 81 238 FANN_TRAIN_RPROP;
#X msg 81 258 FANN_TRAIN_QUICKPROP;
#X text 40 28 you can set the training algorithm simply sending a message
with the name of the algorithm chosen. possible values are: FANN_TRAIN_INCREMENTAL
FANN_TRAIN_BATCH FANN_TRAIN_RPROP FANN_TRAIN_QUICKPROP the default
is: FANN_TRAIN_RPROP see the FANN manual for details on each algorithm:
http://fann.sourceforge.net/html/r1996.html;
#X connect 1 0 0 0;
#X connect 2 0 0 0;
#X connect 3 0 0 0;
#X connect 4 0 0 0;
#X restore 150 153 pd training algorithm;
#X text 360 175 some advanced param;
#N canvas 371 92 698 395 training 0;
#X obj 52 230 outlet;
#X msg 69 118 desired_error 0.01;
#X msg 79 146 max_iterations 500000;
#X msg 90 178 iterations_between_reports 1000;
#X text 58 28 you can change training parameters. see FANN manual for
details (http://fann.sourceforge.net);
#X connect 1 0 0 0;
#X connect 2 0 0 0;
#X connect 3 0 0 0;
#X restore 151 179 pd training params;
#N canvas 371 92 694 391 activation 0;
#X obj 49 335 outlet;
#X text 40 28 you can set ti output activation algorithm passing a
message to nn. see the FANN manual for description of the algorithms
;
#X msg 69 118 set_activation_function_output FANN_THRESHOLD;
#X msg 83 139 set_activation_function_output FANN_THRESHOLD_SYMMETRIC
;
#X msg 95 163 set_activation_function_output FANN_LINEAR;
#X msg 98 184 set_activation_function_output FANN_SIGMOID;
#X msg 106 206 set_activation_function_output FANN_SIGMOID_STEPWISE
;
#X msg 108 233 set_activation_function_output FANN_SIGMOID_SYMMETRIC
;
#X msg 115 256 set_activation_function_output FANN_SIGMOID_SYMMETRIC_STEPWISE
;
#X connect 2 0 0 0;
#X connect 3 0 0 0;
#X connect 4 0 0 0;
#X connect 5 0 0 0;
#X connect 6 0 0 0;
#X connect 7 0 0 0;
#X connect 8 0 0 0;
#X restore 150 203 pd activation algorithm;
#X msg 151 287 details;
#X text 229 285 details on the current nn;
#X msg 145 333 help;
#X connect 1 0 0 0;
#X connect 2 0 0 0;
#X connect 3 0 0 0;
#X connect 4 0 0 0;
#X connect 7 0 0 0;
#X connect 9 0 0 0;
#X connect 10 0 0 0;
#X connect 11 0 0 0;
#X connect 13 0 0 0;
#X restore 128 258 pd other commands;
#N canvas 0 0 665 525 save 0;
#X obj 39 264 outlet;
#X msg 64 20 filename test.net;
#X msg 66 46 save;
#X msg 82 103 load;
#X text 221 19 set the filename;
#X text 214 42 save the net to the file;
#X text 138 104 you can reload it too;
#X text 144 182 nn can be loaded from a file at creation time simply
passing the filename as argument;
#X msg 68 71 save test.net;
#X msg 93 130 load test.net;
#X text 144 217 like [ann_td num_inputs frames filename];
#X connect 1 0 0 0;
#X connect 2 0 0 0;
#X connect 3 0 0 0;
#X connect 8 0 0 0;
#X connect 9 0 0 0;
#X restore 118 218 pd save the net;
#X text 270 66 create a nn;
#X text 244 179 run your net;
#X text 258 215 save your net;
#N canvas 0 0 712 542 tips 0;
#X text 51 84 for better performances inputs value should be normalized
\, all input should have the same range (if one input has a larger
range it will be more "important"). the range of each input should
be 0 centered. so [-1 \, 1] is good [-2 \, 2] is good \, [0 \, 1] not
so good [1 \, 2] is bad. the range sould not be too small ([-0.1 \,
0.1] is bad).;
#X text 41 19 TIPS;
#X text 41 56 inputs;
#X text 39 211 outputs;
#X text 50 235 each class of outputs should have its own output value:
don't use the same output for 2 meanings \, use 2 outputs intead \,
1 for each.;
#X restore 167 285 pd tips;
#X text 272 371 an interface to fann classes (http://fann.sourceforge.net)
;
#X text 274 389 by Davide Morelli - info@davidemorelli.it;
#N canvas 227 183 580 411 train 0;
#X obj 32 241 outlet;
#N canvas 100 44 892 558 train 0;
#X obj 57 397 outlet;
#X msg 60 31 train;
#X text 126 33 1- set the train mode;
#X text 116 81 2- build a list with inputs and desired output;
#X text 139 101 be shure you provide the correct numbers of inputs
and outputs;
#X obj 168 202 pack s f f f;
#X obj 197 248 pack f f f;
#X obj 168 225 unpack s f f f;
#X msg 192 374 run;
#X obj 198 170 tgl 15 0 empty empty in1 0 -6 0 8 -262144 -1 -1 0 1
;
#X obj 228 170 tgl 15 0 empty empty in2 0 -6 0 8 -262144 -1 -1 0 1
;
#X obj 259 170 tgl 15 0 empty empty output 0 -6 0 8 -262144 -1 -1 0
1;
#X obj 148 169 bng 15 250 50 0 empty empty train! 0 -6 0 8 -262144
-1 -1;
#X text 299 183 set inputs and output value \, then send the list clicking
on the "train!" bang;
#X text 229 374 3- when you are ready switch again to run mode before
exiting;
#X text 311 308 NOTE2: while training the right outlet gives you the
mean square error after each training pattern.;
#X msg 316 278 create 2 1 5;
#X text 315 226 NOTE1: before training with this example you should
have created a nn with 2 ins and 1 out and 5 frames with a command
like:;
#X connect 1 0 0 0;
#X connect 5 0 7 0;
#X connect 6 0 0 0;
#X connect 7 1 6 0;
#X connect 7 2 6 1;
#X connect 7 3 6 2;
#X connect 8 0 0 0;
#X connect 9 0 5 1;
#X connect 10 0 5 2;
#X connect 11 0 5 3;
#X connect 12 0 5 0;
#X connect 16 0 0 0;
#X restore 68 50 pd train it on the fly;
#X text 62 5 there are 2 ways to train your net;
#X text 253 47 on the fly is simpler;
#X text 86 128 with a trainfile the net could be more accurate;
#X msg 89 149 train-on-file test.txt;
#X connect 1 0 0 0;
#X connect 5 0 0 0;
#X restore 115 119 pd train;
#X text 190 118 train a nn;
#X obj 113 360 print mse;
#X obj 54 391 print out;
#X text 150 315 2 args needed: num_inputs and frames;
#X text 148 331 see [pd creation examples] for details;
#X obj 33 319 ann_td 2 5;
#X text 9 2 ann_td: time delay neural networks in pd;
#X connect 0 0 16 0;
#X connect 1 0 16 0;
#X connect 2 0 16 0;
#X connect 3 0 16 0;
#X connect 10 0 16 0;
#X connect 16 0 13 0;
#X connect 16 1 12 0;