Caffe - How to make Python API generate runtime input prototxt - machine-learning

I need the python API to generate such a prototxt :
input: "data"
input_dim: 7
input_dim: 3
input_dim: 368
input_dim: 368
input: "heatmap"
input_dim: 7
input_dim: 15
input_dim: 46
input_dim: 46
input: "center_map"
input_dim: 7
input_dim: 1
input_dim: 368
input_dim: 368
input: "h_t_1"
input_dim: 7
input_dim: 48
input_dim: 46
input_dim: 46
input: "cell_t_1"
input_dim: 7
input_dim: 48
input_dim: 46
input_dim: 46
How does one go about doing this? I tried doing this:
caffeNet.tops['x'] = L.Data(name="x")
caffeNet.tops['y'] = L.Data(name="y")
caffeNet.tops['z'] = L.Data(name="z")
But this only generates in this format which is invalid syntax:
layer {
name: "h_t"
type: "Data"
top: "h_t"
}
layer {
name: "cell_t"
type: "Data"
top: "cell_t"
}
layer {
name: "last_paf"
type: "Data"
top: "last_paf"
}
If I try to load my net with this format, it crashes, it only seems to accept the other format. I could write a hacky script to convert it but surely there is a better way

You need to use "Input" layers:
caffeNet.tops['data'] = L.Input(name="data",
input_param={"shape":{"dim":[7, 3, 368, 368]}})
caffeNet.tops['heatmap'] = L.Input(name="heatmap",
input_param={"shape":{"dim":[7, 15, 46, 46]}})
caffeNet.tops['center_map'] = L.Input(name="center_map",
input_param={"shape":{"dim":[7, 1, 368, 368]}})
and so on.

Related

ActiveRecord::Fixture::FixtureError: table has no columns named "false"

I am getting the error: ActiveRecord::Fixture::FixtureError: table "creatures" has no columns named "false". I have no column named false in this model.
What is going on?
Here is my fixture:
3 one:
4 name: MyString
5 no: 1
6 type1: 1
7 type2: 1
8 total: 1
9 hp: 1
10 attack: 1
11 defense: 1
12 special_attack: 1
13 special_defense: 1
14 speed: 1
15 generation: 1
16 legendary: false
17
Putting the no in single quotes solved the problem.
If I put a debugger call just before the error is raised:
[475, 484] in /usr/local/bundle/gems/activerecord-7.0.4.2/lib/active_record/connection_adapters/abstract/database_statements.rb
475: fixture = fixture.stringify_keys
476:
477: unknown_columns = fixture.keys - columns.keys
478: if unknown_columns.any?
479: debugger
=> 480: raise Fixture::FixtureError, %(table "#{table_name}" has no columns named #{unknown_columns.map(&:inspect).join(', ')}.)
481: end
482:
483: columns.map do |name, column|
484: if fixture.key?(name)
(byebug) fixtures
It looks like the no got interpreted as a false:
[{"name"=>"MyString", false=>1, "type1"=>1, "type2"=>1, "total"=>1, "hp"=>1, "attack"=>1, "defense"=>1, "special_attack"=>1, "special_defense"=>1, "speed"=>1, "genneration"=>1, "legendary"=>false, "created_at"=>2023-02-05 18:53:31.63881045 UTC, "updated_at"=>2023-02-05 18:53:31.63881045 UTC, "id"=>980190962}, {"name"=>"MyString", false=>2, "type1"=>1, "type2"=>1, "total"=>1, "hp"=>1, "attack"=>1, "defense"=>1, "special_attack"=>1, "special_defense"=>1, "speed"=>1, "genneration"=>1, "legendary"=>false, "created_at"=>2023-02-05 18:53:31.63881045 UTC, "updated_at"=>2023-02-05 18:53:31.63881045 UTC, "id"=>298486374}]
Putting the no in single quotes solved the problem:
3 one:
4 name: MyString
5 'no': 1
6 type1: 1
7 type2: 1
8 total: 1
9 hp: 1
10 attack: 1
11 defense: 1
12 special_attack: 1
13 special_defense: 1
14 speed: 1
15 generation: 1
16 legendary: false
17
18 two:
19 name: MyString
20 'no': 2
21 type1: 1
22 type2: 1
23 total: 1
24 hp: 1
25 attack: 1
26 defense: 1
27 special_attack: 1
28 special_defense: 1
29 speed: 1
30 generation: 1
31 legendary: false

FitFailedWarning: Estimator fit failed. The score on this train-test partition for these parameters will be set to nan when using greater int values

I've recently watched a YouTube (DataSchool) video where the guy used only 3 columns from the Titanic dataset and made a pipeline. I wanted to add more columns to get better accuracy so I added Age and Fare.
I think it's probably because of the values of Age and Fare that I'm getting this error when I perform cross_val_score
columns_trans = make_column_transformer(
(OneHotEncoder(), ['Sex', 'Embarked']),
remainder='passthrough')
logreg = LogisticRegression(solver='lbfgs')
pipe = make_pipeline(columns_trans, logreg)
cross_val_score(pipe, X, y, cv=5, scoring='accuracy').mean()
/opt/conda/lib/python3.7/site-packages/sklearn/model_selection/_validation.py:552: FitFailedWarning: Estimator fit failed. The score on this train-test partition for these parameters will be set to nan.
If I remove Age and Fare, everything works fine. I was wondering if the Column Transformer or the make_pipeline had a problem with values like that.
I also tried scaling the values of Fare and Age, then it gave a cross_val_score but failed in pipe.predict() giving an error:
ValueError: Input contains NaN, infinity or a value too large for dtype('float64')
Traceback:
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
/tmp/ipykernel_119/4279568460.py in <module>
----> 1 cross_val_score(pipe, X, y, cv=5, scoring='accuracy', error_score="raise").mean()
/opt/conda/lib/python3.7/site-packages/sklearn/utils/validation.py in inner_f(*args, **kwargs)
70 FutureWarning)
71 kwargs.update({k: arg for k, arg in zip(sig.parameters, args)})
---> 72 return f(**kwargs)
73 return inner_f
74
/opt/conda/lib/python3.7/site-packages/sklearn/model_selection/_validation.py in cross_val_score(estimator, X, y, groups, scoring, cv, n_jobs, verbose, fit_params, pre_dispatch, error_score)
404 fit_params=fit_params,
405 pre_dispatch=pre_dispatch,
--> 406 error_score=error_score)
407 return cv_results['test_score']
408
/opt/conda/lib/python3.7/site-packages/sklearn/utils/validation.py in inner_f(*args, **kwargs)
70 FutureWarning)
71 kwargs.update({k: arg for k, arg in zip(sig.parameters, args)})
---> 72 return f(**kwargs)
73 return inner_f
74
/opt/conda/lib/python3.7/site-packages/sklearn/model_selection/_validation.py in cross_validate(estimator, X, y, groups, scoring, cv, n_jobs, verbose, fit_params, pre_dispatch, return_train_score, return_estimator, error_score)
246 return_times=True, return_estimator=return_estimator,
247 error_score=error_score)
--> 248 for train, test in cv.split(X, y, groups))
249
250 zipped_scores = list(zip(*scores))
/opt/conda/lib/python3.7/site-packages/joblib/parallel.py in __call__(self, iterable)
1039 # remaining jobs.
1040 self._iterating = False
-> 1041 if self.dispatch_one_batch(iterator):
1042 self._iterating = self._original_iterator is not None
1043
/opt/conda/lib/python3.7/site-packages/joblib/parallel.py in dispatch_one_batch(self, iterator)
857 return False
858 else:
--> 859 self._dispatch(tasks)
860 return True
861
/opt/conda/lib/python3.7/site-packages/joblib/parallel.py in _dispatch(self, batch)
775 with self._lock:
776 job_idx = len(self._jobs)
--> 777 job = self._backend.apply_async(batch, callback=cb)
778 # A job can complete so quickly than its callback is
779 # called before we get here, causing self._jobs to
/opt/conda/lib/python3.7/site-packages/joblib/_parallel_backends.py in apply_async(self, func, callback)
206 def apply_async(self, func, callback=None):
207 """Schedule a func to be run"""
--> 208 result = ImmediateResult(func)
209 if callback:
210 callback(result)
/opt/conda/lib/python3.7/site-packages/joblib/_parallel_backends.py in __init__(self, batch)
570 # Don't delay the application, to avoid keeping the input
571 # arguments in memory
--> 572 self.results = batch()
573
574 def get(self):
/opt/conda/lib/python3.7/site-packages/joblib/parallel.py in __call__(self)
261 with parallel_backend(self._backend, n_jobs=self._n_jobs):
262 return [func(*args, **kwargs)
--> 263 for func, args, kwargs in self.items]
264
265 def __reduce__(self):
/opt/conda/lib/python3.7/site-packages/joblib/parallel.py in <listcomp>(.0)
261 with parallel_backend(self._backend, n_jobs=self._n_jobs):
262 return [func(*args, **kwargs)
--> 263 for func, args, kwargs in self.items]
264
265 def __reduce__(self):
/opt/conda/lib/python3.7/site-packages/sklearn/model_selection/_validation.py in _fit_and_score(estimator, X, y, scorer, train, test, verbose, parameters, fit_params, return_train_score, return_parameters, return_n_test_samples, return_times, return_estimator, error_score)
529 estimator.fit(X_train, **fit_params)
530 else:
--> 531 estimator.fit(X_train, y_train, **fit_params)
532
533 except Exception as e:
/opt/conda/lib/python3.7/site-packages/sklearn/pipeline.py in fit(self, X, y, **fit_params)
333 if self._final_estimator != 'passthrough':
334 fit_params_last_step = fit_params_steps[self.steps[-1][0]]
--> 335 self._final_estimator.fit(Xt, y, **fit_params_last_step)
336
337 return self
/opt/conda/lib/python3.7/site-packages/sklearn/linear_model/_logistic.py in fit(self, X, y, sample_weight)
1415 penalty=penalty, max_squared_sum=max_squared_sum,
1416 sample_weight=sample_weight)
-> 1417 for class_, warm_start_coef_ in zip(classes_, warm_start_coef))
1418
1419 fold_coefs_, _, n_iter_ = zip(*fold_coefs_)
/opt/conda/lib/python3.7/site-packages/joblib/parallel.py in __call__(self, iterable)
1039 # remaining jobs.
1040 self._iterating = False
-> 1041 if self.dispatch_one_batch(iterator):
1042 self._iterating = self._original_iterator is not None
1043
/opt/conda/lib/python3.7/site-packages/joblib/parallel.py in dispatch_one_batch(self, iterator)
857 return False
858 else:
--> 859 self._dispatch(tasks)
860 return True
861
/opt/conda/lib/python3.7/site-packages/joblib/parallel.py in _dispatch(self, batch)
775 with self._lock:
776 job_idx = len(self._jobs)
--> 777 job = self._backend.apply_async(batch, callback=cb)
778 # A job can complete so quickly than its callback is
779 # called before we get here, causing self._jobs to
/opt/conda/lib/python3.7/site-packages/joblib/_parallel_backends.py in apply_async(self, func, callback)
206 def apply_async(self, func, callback=None):
207 """Schedule a func to be run"""
--> 208 result = ImmediateResult(func)
209 if callback:
210 callback(result)
/opt/conda/lib/python3.7/site-packages/joblib/_parallel_backends.py in __init__(self, batch)
570 # Don't delay the application, to avoid keeping the input
571 # arguments in memory
--> 572 self.results = batch()
573
574 def get(self):
/opt/conda/lib/python3.7/site-packages/joblib/parallel.py in __call__(self)
261 with parallel_backend(self._backend, n_jobs=self._n_jobs):
262 return [func(*args, **kwargs)
--> 263 for func, args, kwargs in self.items]
264
265 def __reduce__(self):
/opt/conda/lib/python3.7/site-packages/joblib/parallel.py in <listcomp>(.0)
261 with parallel_backend(self._backend, n_jobs=self._n_jobs):
262 return [func(*args, **kwargs)
--> 263 for func, args, kwargs in self.items]
264
265 def __reduce__(self):
/opt/conda/lib/python3.7/site-packages/sklearn/linear_model/_logistic.py in _logistic_regression_path(X, y, pos_class, Cs, fit_intercept, max_iter, tol, verbose, solver, coef, class_weight, dual, penalty, intercept_scaling, multi_class, random_state, check_input, max_squared_sum, sample_weight, l1_ratio)
762 n_iter_i = _check_optimize_result(
763 solver, opt_res, max_iter,
--> 764 extra_warning_msg=_LOGISTIC_SOLVER_CONVERGENCE_MSG)
765 w0, loss = opt_res.x, opt_res.fun
766 elif solver == 'newton-cg':
/opt/conda/lib/python3.7/site-packages/sklearn/utils/optimize.py in _check_optimize_result(solver, result, max_iter, extra_warning_msg)
241 " https://scikit-learn.org/stable/modules/"
242 "preprocessing.html"
--> 243 ).format(solver, result.status, result.message.decode("latin1"))
244 if extra_warning_msg is not None:
245 warning_msg += "\n" + extra_warning_msg
AttributeError: 'str' object has no attribute 'decode'
I solved this error by changing solver=lbfgs to solver=liblinear in LogisticRegression()
logreg = LogisticRegression(solver='lbfgs')
to
logreg = LogisticRegression(solver='liblinear')
And for the following error:
ValueError: Input contains NaN, infinity or a value too large for dtype('float64')
It's best to check if your test data contains any null values or strings.

LSTM predicting time series yields odd results

I'm trying to predict time series data for the next few days looking at past few days, using Keras. My label data is target values for multiple future days, regression model has multiple output neurons (the "direct approach" for time series).
Here is test data with predictions for 10 days, using 60 days history.
10 days prediction for test data
As you can see, future values for all days are about the same. I've spent quite some time on it, and must admit that I'm probably missing something with respect to LSTM...
Here is training data with prediction:
10 days prediction for training data
In order to confirm that I'm preparing data properly, I've created a "tracking data set" which I used to visualize data transformations. Here it is...
Data set:
Open,High,Low,Close,Volume,OpenInt
111,112,113,114,115,0
121,122,123,124,125,0
131,132,133,134,135,0
141,142,143,144,145,0
151,152,153,154,155,0
161,162,163,164,165,0
171,172,173,174,175,0
181,182,183,184,185,0
191,192,193,194,195,0
201,202,203,204,205,0
211,212,213,214,215,0
221,222,223,224,225,0
231,232,233,234,235,0
241,242,243,244,245,0
251,252,253,254,255,0
261,262,263,264,265,0
271,272,273,274,275,0
281,282,283,284,285,0
291,292,293,294,295,0
Training set using 2 days history, predicting 3 days future values (I used different values of history days and future days, and it all makes sense to me), without feature scaling in order to visualize data transformations:
X train (6, 2, 5)
[[[111 112 113 114 115]
[121 122 123 124 125]]
[[121 122 123 124 125]
[131 132 133 134 135]]
[[131 132 133 134 135]
[141 142 143 144 145]]
[[141 142 143 144 145]
[151 152 153 154 155]]
[[151 152 153 154 155]
[161 162 163 164 165]]
[[161 162 163 164 165]
[171 172 173 174 175]]]
Y train (6, 3)
[[131 141 151]
[141 151 161]
[151 161 171]
[161 171 181]
[171 181 191]
[181 191 201]]
Test set
X test (6, 2, 5)
[[[201 202 203 204 205]
[211 212 213 214 215]]
[[211 212 213 214 215]
[221 222 223 224 225]]
[[221 222 223 224 225]
[231 232 233 234 235]]
[[231 232 233 234 235]
[241 242 243 244 245]]
[[241 242 243 244 245]
[251 252 253 254 255]]
[[251 252 253 254 255]
[261 262 263 264 265]]]
Y test (6, 3)
[[221 231 241]
[231 241 251]
[241 251 261]
[251 261 271]
[261 271 281]
[271 281 291]]
Model:
def CreateRegressor(self,
optimizer='adam',
activation='tanh', # RNN activation
init_mode='glorot_uniform',
hidden_neurons=50,
dropout_rate=0.0,
weight_constraint=0,
stateful=False,
# SGD parameters
learn_rate=0.01,
momentum=0):
kernel_constraint = maxnorm(weight_constraint) if weight_constraint > 0 else None
model = Sequential()
model.add(LSTM(units=hidden_neurons, activation=activation, kernel_initializer=init_mode, kernel_constraint=kernel_constraint,
return_sequences=True, input_shape=(self.X_train.shape[1], self.X_train.shape[2]), stateful=stateful))
model.add(Dropout(dropout_rate))
model.add(LSTM(units=hidden_neurons, activation=activation, kernel_initializer=init_mode, kernel_constraint=kernel_constraint,
return_sequences=True, stateful=stateful))
model.add(Dropout(dropout_rate))
model.add(LSTM(units=hidden_neurons, activation=activation, kernel_initializer=init_mode, kernel_constraint=kernel_constraint,
return_sequences=True, stateful=stateful))
model.add(Dropout(dropout_rate))
model.add(LSTM(units=hidden_neurons, activation=activation, kernel_initializer=init_mode, kernel_constraint=kernel_constraint,
return_sequences=False, stateful=stateful))
model.add(Dropout(dropout_rate))
model.add(Dense(units=self.y_train.shape[1]))
if (optimizer == 'SGD'):
optimizer = SGD(lr=learn_rate, momentum=momentum)
model.compile(optimizer=optimizer, loss='mean_squared_error')
return model
...which I create with these params:
self.CreateRegressor(optimizer = 'adam', hidden_neurons = 100)
... and then fit like this:
self.regressor.fit(self.X_train, self.y_train, epochs=100, batch_size=32)
... and predict:
y_pred = self.regressor.predict(X_test)
... or
y_pred_train = self.regressor.predict(X_train)
What am I missing?

Parsing Blocks of Data in REBOL

I have (games scores) data in this format:
Hotspurs Giants 356 6 275 4 442 3
Fierce Lions Club 371 3 2520 5 0 4
Mountain Tigers 2519 2 291 6 342 1
Shooting Stars Club 2430 5 339 1 2472 2
Gun Tooters 329 4 2512 2 2470 6
Banshee Wolves 301 1 2436 3 412 5
The first two/three words represent the club's names, thereafter follows 6 blocks of data per row which represents the club's round-by-round scores and opponent index (starting from 1). In the data above 3 rounds have been played by each team. Hotspurs Giants (index 1) played Banshee Wolves (6) in the 1st round scoring 356 to Banshee's 301, in round 2 Hotspurs Giants played Shooting Stars Club (4) scoring 275 - 339 and in round 3 played Mountain Tigers (3) scoring 442 to Tiger's 342
My question is how to parse this blocks of data in the most efficient way possible such that each club's data will be in the format below considering that a club's name may comprise of two (2) or more words.
Viz
[Club Round Score Opponent Opponent-Score] for each club
Assuming data is:
data: [
Hotspurs Giants 356 6 275 4 442 3
Fierce Lions Club 371 3 2520 5 0 4
Mountain Tigers 2519 2 291 6 342 1
Shooting Stars Club 2430 5 339 1 2472 2
Gun Tooters 329 4 2512 2 2470 6
Banshee Wolves 301 1 2436 3 412 5
]
I think this solves the problem, please check the result:
clubs: copy []
parse data [
some [
copy club some word!
copy numbers some number!
(append clubs reduce [form club numbers])
|
skip
]
]
new-line/all/skip clubs yes 2
list: copy []
parse clubs [
some [
set club string! into [
copy numbers some number! (
i: 1
foreach [score index] numbers [
append list reduce [
club score
pick clubs index * 2 - 1
pick pick clubs index * 2 i
]
i: i + 2
]
)
]
| skip
]
]
new-line/all/skip list yes 4
Afterwards if you probe clubs you should get:
CLUBS is a block of value: [
"Hotspurs Giants" [356 6 275 4 442 3]
"Fierce Lions Club" [371 3 2520 5 0 4]
"Mountain Tigers" [2519 2 291 6 342 1]
"Shooting Stars Club" [2430 5 339 1 2472 2]
"Gun Tooters" [329 4 2512 2 2470 6]
"Banshee Wolves" [301 1 2436 3 412 5]
]
And if you probe list the output is:
LIST is a block of value: [
"Hotspurs Giants" 356 "Banshee Wolves" 301
"Hotspurs Giants" 275 "Shooting Stars Club" 339
"Hotspurs Giants" 442 "Mountain Tigers" 342
"Fierce Lions Club" 371 "Mountain Tigers" 2519
"Fierce Lions Club" 2520 "Gun Tooters" 2512
"Fierce Lions Club" 0 "Shooting Stars Club" 2472
"Mountain Tigers" 2519 "Fierce Lions Club" 371
"Mountain Tigers" 291 "Banshee Wolves" 2436
"Mountain Tigers" 342 "Hotspurs Giants" 442
"Shooting Stars Club" 2430 "Gun Tooters" 329
"Shooting Stars Club" 339 "Hotspurs Giants" 275
"Shooting Stars Club" 2472 "Fierce Lions Club" 0
"Gun Tooters" 329 "Shooting Stars Club" 2430
"Gun Tooters" 2512 "Fierce Lions Club" 2520
"Gun Tooters" 2470 "Banshee Wolves" 412
"Banshee Wolves" 301 "Hotspurs Giants" 356
"Banshee Wolves" 2436 "Mountain Tigers" 291
"Banshee Wolves" 412 "Gun Tooters" 2470
]
Here is one example (using Rebol 3) showing how this could be done:
club-data: map [] ; store data in hash map is one option
foreach line read/lines %games-scores.txt [
fields: split line space
; lets take last 6 cols of data
scores: reverse collect [loop 6 [keep to-integer take/last fields]]
; and whats left is the club name
club-name: form fields
; build club data
club-data/(club-name): scores
]
Above assumes data is in file games-scores.txt and returns you a MAP! (hash map) called club-data where your club data would look like this:
make map! [
"Hotspurs Giants" [356 6 275 4 442 3]
"Fierce Lions Club" [371 3 2520 5 0 4]
"Mountain Tigers" [2519 2 291 6 342 1]
"Shooting Stars Club" [2430 5 339 1 2472 2]
"Gun Tooters" [329 4 2512 2 2470 6]
"Banshee Wolves" [301 1 2436 3 412 5]
]
One caveat... READ/LINES will load whole file into memory. So if games-scores.txt is big you should look at using OPEN instead and read in one line at a time.
Update - re: your comment here is same example in Rebol 2 [tested in REBOL/Core 2.7.8.2.5 (2-Jan-2011)]:
club-data: make hash! [] ; of course doesn't have to be hash!
foreach line read/lines %games-scores.txt [
fields: parse line none
scores: reverse collect [loop 6 [keep to-integer take/last fields]]
club-name: form fields
append club-data reduce [club-name scores]
]

Try to simulate a neural network in MATLAB by myself

I tried to create a neural network to estimate y = x ^ 2. So I created a fitting neural network and gave it some samples for input and output. I tried to build this network in C++. But the result is different than I expected.
With the following inputs:
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49
50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 -1
-2 -3 -4 -5 -6 -7 -8 -9 -10 -11 -12 -13 -14 -15 -16 -17 -18 -19 -20 -21 -22 -23 -24 -25 -26 -27 -28 -29 -30 -31 -32 -33 -34 -35 -36 -37 -38 -39 -40 -41 -42 -43 -44 -45 -46 -47 -48 -49 -50 -51 -52 -53 -54 -55 -56 -57 -58 -59 -60 -61 -62 -63 -64 -65 -66 -67 -68 -69 -70 -71
and the following outputs:
0 1 4 9 16 25 36 49 64 81 100 121 144 169 196 225 256 289 324 361 400
441 484 529 576 625 676 729 784 841 900 961 1024 1089 1156 1225 1296
1369 1444 1521 1600 1681 1764 1849 1936 2025 2116 2209 2304 2401 2500
2601 2704 2809 2916 3025 3136 3249 3364 3481 3600 3721 3844 3969 4096
4225 4356 4489 4624 4761 4900 5041 1 4 9 16 25 36 49 64 81 100 121 144
169 196 225 256 289 324 361 400 441 484 529 576 625 676 729 784 841
900 961 1024 1089 1156 1225 1296 1369 1444 1521 1600 1681 1764 1849
1936 2025 2116 2209 2304 2401 2500 2601 2704 2809 2916 3025 3136 3249
3364 3481 3600 3721 3844 3969 4096 4225 4356 4489 4624 4761 4900 5041
I used fitting tool network. with matrix rows. Training is 70%, validation is 15% and testing is 15% as well. The number of hidden neurons is two. Then in command lines I wrote this:
purelin(net.LW{2}*tansig(net.IW{1}*inputTest+net.b{1})+net.b{2})
Other information :
My net.b[1] is: -1.16610230053776 1.16667147712026
My net.b[2] is: 51.3266249426358
And net.IW(1) is: 0.344272596370387 0.344111217766824
net.LW(2) is: 31.7635369693519 -31.8082184881063
When my inputTest is 3, the result of this command is 16, while it should be about 9. Have I made an error somewhere?
I found the Stack Overflow post Neural network in MATLAB that contains a problem like my problem, but there is a little difference, and the differences is in that problem the ranges of input and output are same, but in my problem is no. That solution says I need to scale out the results, but how can I scale out my result?
You are right about scaling. As was mentioned in the linked answer, the neural network by default scales the input and output to the range [-1,1]. This can be seen in the network processing functions configuration:
>> net = fitnet(2);
>> net.inputs{1}.processFcns
ans =
'removeconstantrows' 'mapminmax'
>> net.outputs{2}.processFcns
ans =
'removeconstantrows' 'mapminmax'
The second preprocessing function applied to both input/output is mapminmax with the following parameters:
>> net.inputs{1}.processParams{2}
ans =
ymin: -1
ymax: 1
>> net.outputs{2}.processParams{2}
ans =
ymin: -1
ymax: 1
to map both into the range [-1,1] (prior to training).
This means that the trained network expects input values in this range, and outputs values also in the same range. If you want to manually feed input to the network, and compute the output yourself, you have to scale the data at input, and reverse the mapping at the output.
One last thing to remember is that each time you train the ANN, you will get different weights. If you want reproducible results, you need to fix the state of the random number generator (initialize it with the same seed each time). Read the documentation on functions like rng and RandStream.
You also have to pay attention that if you are dividing the data into training/testing/validation sets, you must use the same split each time (probably also affected by the randomness aspect I mentioned).
Here is an example to illustrate the idea (adapted from another post of mine):
%%# data
x = linspace(-71,71,200); %# 1D input
y_model = x.^2; %# model
y = y_model + 10*randn(size(x)).*x; %# add some noise
%%# create ANN, train, simulate
net = fitnet(2); %# one hidden layer with 2 nodes
net.divideFcn = 'dividerand';
net.trainParam.epochs = 50;
net = train(net,x,y);
y_hat = net(x);
%%# plot
plot(x, y, 'b.'), hold on
plot(x, x.^2, 'Color','g', 'LineWidth',2)
plot(x, y_hat, 'Color','r', 'LineWidth',2)
legend({'data (noisy)','model (x^2)','fitted'})
hold off, grid on
%%# manually simulate network
%# map input to [-1,1] range
[~,inMap] = mapminmax(x, -1, 1);
in = mapminmax('apply', x, inMap);
%# propagate values to get output (scaled to [-1,1])
hid = tansig( bsxfun(#plus, net.IW{1}*in, net.b{1}) ); %# hidden layer
outLayerOut = purelin( net.LW{2}*hid + net.b{2} ); %# output layer
%# reverse mapping from [-1,1] to original data scale
[~,outMap] = mapminmax(y, -1, 1);
out = mapminmax('reverse', outLayerOut, outMap);
%# compare against MATLAB output
max( abs(out - y_hat) ) %# this should be zero (or in the order of `eps`)
I opted to use the mapminmax function, but you could have done that manually as well. The formula is a pretty simply linear mapping:
y = (ymax-ymin)*(x-xmin)/(xmax-xmin) + ymin;

Resources