def train(inputs, outputs, size, participants, victors, generations, threshold, cRate, mRate, printFile=False):
'''Create and start training the NN via evolution strategy. Selection, crossover, mutation, evaluation.'''
global hero
global OrigAnswers
OrigAnswers = copy.deepcopy(outputs)
EvaluationNN = GA.create_net(inputs, outputs)
population = init_pop(EvaluationNN, inputs, outputs, size)
# Test each citizen and determine initial fitness
GA.evaluate(EvaluationNN, population, inputs, outputs)
if printFile: f = open('ES.csv', 'w')
gen = 0
children = []
# loop until a hero is found or we've reached max generations
while gen <= generations and hero == 0:
# Select our parents using tournament selection
parents = GA.tournament(population, participants, victors)
# Have our parents mate (Crossover)
children = GA.mate(parents, cRate)
# Have the children experience the world (Mutate)
for child in children:
mutate(child, mRate)
# Test each child's fitness
GA.evaluate(EvaluationNN, children, inputs, outputs)
children = GA.tournament(children, participants, victors)
population = sorted(population + children,
key=itemgetter(-1))[:-victors]
if GA.heroFound(population, threshold):
break
else:
print("Training: {:2.2%}".format(
population[0][-1]), "{:2.2%} ".format(gen / generations), end="\r")
if printFile: f.write('%f,' % population[0][-1])
if printFile: f.write('\n')
gen += 1
if printFile: f.close()
if hero == 0:
gen -= 1
hero = sorted(population, key=itemgetter(-1))[0]
EvaluationNN.SetNNWeights(hero[:-1]) # Load hero into NN, prep for usage.
# Evaluate the hero on the inputs and outputs
print('Generations: %d' % gen, ' ' * 20)
print("Error Relative: {:2.5%}".format(NN.calcRelativeError(EvaluationNN, inputs, OrigAnswers)))
print("Least Squares: %d" % NN.calcLeastSquaresError(EvaluationNN, inputs, OrigAnswers))
print("Loss Squared: %d" % NN.calcLossSquared(EvaluationNN, inputs, OrigAnswers))
#for x in inputs:
# EvaluationNN.SetStartingNodesValues(x)
# EvaluationNN.CalculateNNOutputs()
# print(x, EvaluationNN.GetNNResults(), EvaluationNN.GetNNResultsInt(), OrigAnswers[inputs.index(x)])
print()
return EvaluationNN
def test_ModelLayer():
for typ in ModelTypes:
for i in range(4):
cl = i+1 if typ == "SL2" and i >0 else 0
NNx = NN.clsNN(1,i,1,Type=typ)
assert len(NNx.model) == 1 # Only One Model (one Output)
assert len(NNx.model[0].layers) == i + 2 + cl # i+ Input and Output Layer
def xQModelPredict(self, Xraw, rund=None):
Xx = []
Ax = []
Q = []
la = len(self.actions)
# Extend X to X+Actions
for state in Xraw:
Q.append([0] * 4)
for j in range(len(self.actions)):
if str(type(Xraw[0][-1])) == "<class 'int'>":
Xx.append(state[:-1])
else:
Xx.append(state)
Ax.append(j)
X = self._RetXActAs01(Xx, Ax)
# Predict
Q01np = NN.ModelPredict(self.QModel[0], X)
# Reduce Output to State
Q01 = [float(Q01np[i][0]) for i in range(len(Q01np))]
for i in range(len(Q01)):
if rund == None:
Q[i // la][i % la] = Q01[i]
else:
Q[i // la][i % la] = round(Q01[i], rund)
return Q
def start_thread(inp, activation, out_activ, outp, learn, thresh, mmntm,
logger):
global count
training_inputs = []
training_data = []
count += 1
print(out_activ)
testNN = NN.main(inp, activation, out_activ, outp, learn, thresh, mmntm)
print("DONE TRAINING")
for i in inp:
for j in i:
training_inputs.append(random.randint(
0, 4)) #create random inputs for testing
training_data.append(training_inputs)
training_inputs = []
logger.info("ACTIVATION SET: ")
logger.info(activation)
logger.info("OUTPUT ACTIVATION: %s" % out_activ)
logger.info("TESTING INPUT: ")
logger.info(training_data)
logger.info("OUTPUT: ")
for x in training_data:
testNN.SetStartingNodesValues(x)
testNN.CalculateNNOutputs()
logger.info(str(x))
logger.info(testNN.GetNNResults())
logger.info("RB OUTPUT: %s" % rb_test.rb_test(x))
def start_thread(inp, activation, out_activ, outp, learn, thresh, mmntm, logger):
global count
training_inputs = []
training_data = []
count += 1
print(out_activ)
testNN = NN.main(inp, activation, out_activ, outp, learn, thresh, mmntm)
print("DONE TRAINING")
for i in inp:
for j in i:
training_inputs.append(random.randint(0, 4)) # create random inputs for testing
training_data.append(training_inputs)
training_inputs = []
logger.info("ACTIVATION SET: ")
logger.info(activation)
logger.info("OUTPUT ACTIVATION: %s" % out_activ)
logger.info("TESTING INPUT: ")
logger.info(training_data)
logger.info("OUTPUT: ")
for x in training_data:
testNN.SetStartingNodesValues(x)
testNN.CalculateNNOutputs()
logger.info(str(x))
logger.info(testNN.GetNNResults())
logger.info("RB OUTPUT: %s" % rb_test.rb_test(x))
def verify_frame(pred, detc, le):
global frame, NAME, available, cache, rect, correct
while True:
if cache:
state, shape, rect = facial_landmarks.test_frame(frame, pred, detc)
if state:
#FIND THE NAME
# _, embeddings = facial_landmarks.get_ratios(shape, frame)
embeddings = face_recognition.get_embeddings(
frame, rect, model)
name, percentage = NN.predict_input_from_video(embeddings,
le,
second=False)
percentage = float("{:.2f}".format(percentage))
NAME = str(name) + " " + str(percentage) + "%"
# if name != "Bill_Clinton":
# print (name)
# correct = False
# else:
# print ("Correct")
# correct = True
available = True
else:
available = False
cache = False
def __init__(self, alpha, decay, n_panels, buffer_size, path=''):
# Initialize model
self.model = NN.NN(n_panels)
self.n_panels = n_panels
# Load model
if len(path) != 0:
self.load(path)
# Initialize loss criterion, and optimizer
self.criterion = nn.MSELoss()
self.optimizer_forward = torch.optim.Adam(self.model.parameters(),
lr=alpha)
self.optimizer_backward = torch.optim.Adam(self.model.parameters(),
lr=alpha)
self.lr_scheduler_forward = torch.optim.lr_scheduler.ExponentialLR(
optimizer=self.optimizer_forward, gamma=decay)
self.lr_scheduler_backward = torch.optim.lr_scheduler.ExponentialLR(
optimizer=self.optimizer_backward, gamma=decay)
# Initialize frame buffer
self.buffer_size = buffer_size
self.panels_forward_buffer = []
self.performance_forward_buffer = []
self.panels_backward_buffer = []
self.performance_backward_buffer = []
def Alex_features_MNIST(bulk_size):
"""Pre-processing MNIST data to make them consistent with AlexNet, and
then extract the features as the output of the 7'th layer
"""
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
imagenet_mean = np.array([104., 117., 124.], dtype=np.float32)
# we cannot give the whole training data at once because of OOM
# forming the division indices
train_size = mnist.train.images.shape[0]
divisions = np.append(np.arange(0, train_size, bulk_size), train_size)
train_features = np.zeros((4096, train_size))
for t in range(len(divisions) - 1):
inds = np.arange(divisions[t], divisions[t + 1])
Alexified_images = np.zeros((len(inds), 227, 227, 3))
for i in range(len(inds)):
img = np.reshape(mnist.train.images[inds[i], :], (28, 28))
img = cv2.resize(img.astype(np.float32), (227, 227))
img = img.reshape((227, 227, 1))
img = np.repeat(img, 3, axis=2)
img -= imagenet_mean
Alexified_images[i, :, :, :] = img.reshape((1, 227, 227, 3))
train_features[:, inds] = NN.AlexNet_features(Alexified_images).T
print("%d / %d" % (t, len(divisions) - 1))
print("Extracting test features...")
test_size = mnist.test.images.shape[0]
divisions = np.append(np.arange(0, test_size, bulk_size), test_size)
test_features = np.zeros((4096, test_size))
for t in range(len(divisions) - 1):
inds = np.arange(divisions[t], divisions[t + 1])
Alexified_images = np.zeros((len(inds), 227, 227, 3))
for i in range(len(inds)):
img = np.reshape(mnist.test.images[inds[i], :], (28, 28))
img = cv2.resize(img.astype(np.float32), (227, 227))
img = img.reshape((227, 227, 1))
img = np.repeat(img, 3, axis=2)
img -= imagenet_mean
Alexified_images[i, :, :, :] = img.reshape((1, 227, 227, 3))
test_features[:, inds] = NN.AlexNet_features(Alexified_images).T
print("%d / %d" % (t, len(divisions) - 1))
return train_features, test_features
def NeuralNetworkTest(pca_option):
import NN
NN.NeuralNetworkSimulation(
NN.nn, processing.linear_pca, processing.overall_training_data, pca_option)
processing.final_validation = np.array(processing.final_validation)
FV_features = []
FV_labels = []
FV_features, FV_labels = processing.createFeatures_Labels(
processing.final_validation)
FV_features_data = None
FV_labels_data = None
FV_features_data, FV_labels_data = processing.convertToDataFrame(
FV_features, FV_labels, processing.column_titles)
global NN_final_predictions
if(pca_option == 'yes' or pca_option == 'both'):
transformed_FV = processing.linear_pca.transform(FV_features_data)
final_predictions = NN.nn.predict(transformed_FV)
NN_final_predictions = final_predictions
accuracy = metrics.accuracy_score(final_predictions, FV_labels)
precision = metrics.precision_score(
FV_labels, final_predictions, average='micro')
recall = metrics.recall_score(
FV_labels, final_predictions, average='micro')
print('NEURAL NETWORK MODEL FINAL TEST DATA ACCURACY: ', 100 * accuracy)
print('NEURAL NETWORK MODEL FINAL TEST DATA PRECISION: ', 100 * precision)
print('NEURAL NETWORK MODEL FINAL TEST DATA RECALL: ', 100 * recall)
print()
return accuracy, precision, recall
else:
final_predictions = NN.nn.predict(FV_features_data)
NN_final_predictions = final_predictions
accuracy = metrics.accuracy_score(final_predictions, FV_labels)
precision = metrics.precision_score(
FV_labels, final_predictions, average='micro')
recall = metrics.recall_score(
FV_labels, final_predictions, average='micro')
print('NEURAL NETWORK MODEL FINAL TEST DATA ACCURACY: ', 100 * accuracy)
print('NEURAL NETWORK MODEL FINAL TEST DATA PRECISION: ', 100 * precision)
print('NEURAL NETWORK MODEL FINAL TEST DATA RECALL: ', 100 * recall)
print()
return accuracy, precision, recall
def trainAndSaveNN(train_dl, test_dl, model):
'''
Train the Neural Network and save the trained model
Parameters
----------
train_dl : Training data
test_dl : Test data
model : Neural Network model
Returns
-------
None.
'''
nn.train_model(train_dl, test_dl, model)
torch.save(model.state_dict(), 'trainedNN.pt')
def split_train(ds, lab, test_ratio=0.3, regression=False, model=None):
all = len(lab)
splitor = split.TTSplit(all, 'portion', test=test_ratio)
tr_idx, te_idx = splitor.split()
xtr, ytr, xte, yte = ds[tr_idx], lab[tr_idx], ds[te_idx], lab[te_idx]
if model: pass
else:
if regression: model = NN.FcRegRDKit()
else: model = NN.FcRDKit()
print('start training')
if regression: nn = active.TorchRegressionFold(xtr, ytr, xte, yte, model, 'active', path, ['percent_of_unlabel', 1],
measure='distance', distance='linear')
else: nn = active.TorchFold(xtr, ytr, xte, yte, model, 'active', path, ['percent_of_unlabel', 1])
nn.train()
print('finish training')
def test_ModelOut():
for typ in ModelTypes:
for i in range(1,4):
NNx = NN.clsNN(1,1,i,Type=typ)
if typ == "STD":
assert len(NNx.model) == 1
else:
assert len(NNx.model) == i
def test_forward(self):
input_ = np.random.random((5, 3))
nn = NN.NN(loss_func=loss_fs.MeanSquaredLoss())
nn.add_layer(
layers.DenseLayer(n_neurons=5, activation_func=af.Sigmoid()))
nn.add_layer(
layers.DenseLayer(n_neurons=7, activation_func=af.Sigmoid()))
rv = nn.forward(input_)
assert rv.shape == (5, 7)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--df', default=df_path, metavar='df', help='Dataframe path.')
parser.add_argument('--metrics', default=True, help='Turn on metrics.')
args = parser.parse_args()
df = pd.read_csv(args.df, header=None)
X = df.loc[:, df.columns != 1].to_numpy()
Y = df[1].to_numpy()
Y = np.where((Y == 'M'), 1, 0)
X = NN.feature_scale(X, mode='standart')
split_index = round(X.shape[0] * 0.8) + 1
X_train, X_test = X[:split_index], X[split_index:]
Y_train, Y_test = Y[:split_index], Y[split_index:]
nn = NN.NN()
nn.load_weights()
nn.predict(X_test)
if args.metrics:
nn.calc_metrics(X_test, Y_test)
def neural_network_with_batch(X_c,Y_c,hidden_layer_list,activation_list,learning_rate,l2_norm,keep_prob=1,max_iter=1000,batch_size=None,\
beta1=0.9,beta2=0.999,seed=10,iter_to_print=100,optimization='gd',epsilon=1e-8,verbose=False):
m = X_c.shape[1]
nx = X_c.shape[0]
ny = Y_c.shape[0]
if batch_size is None:
batch_size = m
epochs = list()
tb = int(math.ceil(float(m) / batch_size))
for i in xrange(tb):
epochs.append((X_c[:, i * batch_size:(i + 1) * batch_size],
Y_c[:, i * batch_size:(i + 1) * batch_size, ]))
epochs = tuple(epochs)
np.random.seed(seed)
if (type(keep_prob) == int):
keep_prob = np.repeat(float(keep_prob), len(hidden_layer_list) + 1)
keep_prob[len(keep_prob) - 1] = 1
assert (len(keep_prob) == len(hidden_layer_list) + 1)
parameter, momentum, RMS = initialize_parameter_rms(
nx, ny, activation_list, hidden_layer_list)
cost_series = list()
for i in range(max_iter):
for epoch in epochs:
(X, Y) = epoch
m = X.shape[1]
AL, linear_cache = nn.forward_propagation_(X, hidden_layer_list,
parameter, keep_prob,
activation_list)
cost = nn.compute_cost(AL, Y, parameter, l2_norm)
if (math.isnan(cost)):
print("Cost went nuts")
break
# grads=nn.backward_propagation_(Y,AL,linear_cache,keep_prob)
grads = nn.backward_propagation_both(Y, AL, linear_cache,
keep_prob)
if (optimization == 'gd'):
parameter = nn.update_parameter(parameter, grads,
learning_rate, l2_norm, m)
elif (optimization == 'gdm'):
parameter, momentum = update_parameter_momentum(
parameter, grads, momentum, learning_rate, l2_norm, m,
beta1, i + 1)
elif (optimization == 'adam'):
parameter, momentum, RMS = update_parameter_adam(
parameter, grads, momentum, RMS, learning_rate, l2_norm, m,
beta1, beta2, i + 1, epsilon)
AL_c, _ = nn.forward_propagation_(X_c, hidden_layer_list, parameter,
activation_list)
cost_t = nn.compute_cost(AL_c, Y_c, parameter, l2_norm)
cost_series.append(cost_t)
if (verbose and i % iter_to_print == 0):
print("Iteration: %s Cost=%.5f" % (i, cost_t))
cache = (parameter, hidden_layer_list, activation_list)
AL, _ = nn.forward_propagation_(X_c, hidden_layer_list, parameter,
activation_list)
return AL, cache, cost_series
def test_backward(self):
input_ = np.random.random((5, 3))
output_grad = np.random.random((5, 7))
nn = NN.NN(loss_func=loss_fs.MeanSquaredLoss())
nn.add_layer(
layers.DenseLayer(n_neurons=4, activation_func=af.Sigmoid()))
nn.add_layer(
layers.DenseLayer(n_neurons=7, activation_func=af.Sigmoid()))
nn.forward(input_)
nn.backward(output_grad)
def parse_input(game_style, view):
splitter = game_style.split(" ")
switcher = {
"solo": Solo_Ai(view),
"minmax": AI.MINMAX(),
"mcts": AI.MCTS(),
"rnd": AI.VariousRnd(),
"NN": NN.BetaOne()
}
return switcher[splitter[0]], switcher[splitter[1]]
def parse_input(game_style):
splitter = game_style.split(" ")
switcher = {
"solo": AI.SOLO(),
"minmax": AI.MINMAX(),
"mcts": AI.MCTS(),
"NN": NN.BetaOne(),
"File": Network_from_file()
}
return switcher[splitter[0]], switcher[splitter[1]]
def build_network(shape, individual):
''' Builds a neural net of the input shape using the individual's array of weights. '''
nn = NN.Network(shape)
counter = 0
for layer in nn.weights:
for node in layer:
for i in range(len(node)):
node[i] = individual[counter]
counter += 1
return nn
def main_loop(pred, detc):
global frame, NAME, available, cache, rect, correct
#video path
path_to_vid = '../video/bill_bush.mp4'
video_capture = cv2.VideoCapture(path_to_vid)
size = (int(video_capture.get(cv2.CAP_PROP_FRAME_WIDTH)),
int(video_capture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
still = True
frame_num = 0
_, _, le = NN.prepare_data()
verification_thread = threading.Thread(target=verify_frame,
args=(pred, detc, le),
daemon=True)
verification_thread.start()
out = cv2.VideoWriter('../video/bill_bush_2.avi',
cv2.VideoWriter_fourcc(*'DIVX'), 30.0, size)
while still:
# Capture frame-by-frame
still, frame_clear = video_capture.read()
#if new, update frame
if not cache:
frame = frame_clear
cache = True
#DRAW ON THE FRAME
if available:
(x, y, w, h) = face_utils.rect_to_bb(rect)
if correct:
cv2.rectangle(frame_clear, (x, y), (x + w, y + h), (0, 255, 0),
2)
cv2.putText(frame_clear, NAME, (x - 10, y - 10),
cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)
else:
cv2.rectangle(frame_clear, (x, y), (x + w, y + h), (0, 0, 255),
2)
cv2.putText(frame_clear, NAME, (x - 10, y - 10),
cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 2)
# cv2.imshow('Video', frame_clear)
out.write(frame_clear)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
print(frame_num)
frame_num += 1
# time.sleep(0.001)
# When everything is done, release the capture
video_capture.release()
out.release()
cv2.destroyAllWindows()
# pred, detc = facial_landmarks.load_pred_detec()
# main_loop(pred, detc)
def oneColumn(self, name_column):
"""
Fill the NaN for one column
:param df: dataframe
:param name_column: column name to fill up
:return: Nothing. Modify self.df_ori
"""
df = self.df_ori[self.df_ori[name_column].notnull()]
df = df.fillna(0)
inter = self.extractXY_values(df, name_column)
X = inter[0]
Y = inter[1]
columns_entry = inter[2]
## print(columns_entry)
# Extract important coefficients. Just ridge regression
columns_entry = self.importantCoeff(X, Y, 12, columns_entry, alpha=12)
x_values = df.loc[:, columns_entry].values
y_values = df.loc[:, [name_column]].values
# Normalisation. MinMaxScaler may not be appropriate
mmsx = MinMaxScaler().fit(x_values)
X = mmsx.transform(x_values)
mmsy = MinMaxScaler().fit(y_values)
Y = mmsy.transform(y_values)
# handle data missing values
df_cor = self.df_ori[self.df_ori[name_column].isnull()]
len_df_cor = len(df_cor)
len_df_ori = len(self.df_ori)
## print(len_df_cor/len_df_ori)
if len_df_cor != 0 and len_df_cor / len_df_ori < 0.2:
df_cor = df_cor.fillna(0)
x_cor = df_cor.loc[:, columns_entry]
y_cor = df_cor.loc[:, [name_column]]
x_cor_values = x_cor.values
y_cor_values = y_cor.values
X_cor = mmsx.transform(x_cor_values)
Y_cor = mmsy.transform(y_cor_values)
# give the mean value to, to avoid bias. Only for RBM or Autoencoder
Y_cor = self.meanValueToNaN(Y, Y_cor)
# Model learning
nn = NN.nn(X.shape[1], [10], learning_rate=0.01)
nn.train(X, Y, X_cor, batch_size=40, n_epoches=15)
# inverse scaler
corrected_column = mmsy.inverse_transform(nn.test)
# correct columns in original dataframe
self.fillDfOneColumn(df_cor, corrected_column, name_column)
def __init__(self, n_networks):
"""
The PSO object contains an input n_networks which is the number of neural networks
that are to be initialised.
networks: is a list to store the initialised networks
global_best_value: is initialised as infinity
global_best_position: gets its shape from the Neural Network's getParams function
global_best_yHat: is initialised at floating point 0. Useful for future plotting of graphs.
"""
self.neurons = int(
input("Inform the number of neurons in hidden layer of NN: "))
self.n_networks = n_networks
self.networks = [
NN.NeuralNetwork(NN.x, NN.y, self.neurons)
for i in range(self.n_networks)
]
self.global_best_value = float("inf")
self.global_best_position = NN.NeuralNetwork(
NN.x, NN.y, self.neurons).getParams.shape
self.global_best_yHat = 0
def file_check(self):
"""Open file explorer, get file path and load net"""
filename = tk.filedialog.askopenfilename(
title="Seleccioni l'arxiu",
filetypes=(("pickled files", "*.pkl"), ("all files", "*.*")))
try:
self.net = NN.load_net(filename)
print("Successful Load")
self.root.destroy()
except:
print("Failed")
def eval_MultimgAL(expr, method_name, img_paths, start_ind=0, save_dir=[]):
m = len(expr.train_paths[0]) - 1
patch_shape = expr.pars['patch_shape'][:2] + \
(m*expr.pars['patch_shape'][2],)
model = NN.create_model(expr.pars['model_name'], expr.pars['dropout_rate'],
expr.nclass, expr.pars['learning_rate'],
expr.pars['grad_layers'],
expr.pars['train_layers'],
expr.pars['optimizer_name'], patch_shape)
model.add_assign_ops()
method_path = os.path.join(expr.root_dir, method_name)
Qs = get_queries(expr, method_name)
qnum = len(Qs)
imgnum = len(img_paths)
save_dir = os.path.join(method_path, 'test_scores.txt')
if start_ind > 0:
scores = np.loadtxt(save_dir)
else:
scores = np.zeros((imgnum, qnum))
with tf.Session() as sess:
model.initialize_graph(sess)
sess.graph.finalize()
for i in range(start_ind, qnum):
weights_path = os.path.join(method_path,
'curr_weights_%d.h5' % (i + 1))
print('Loading weights %s' % weights_path)
model.perform_assign_ops(weights_path, sess)
for j in range(imgnum):
# grid-samples from the j-th image
expr.test_paths = img_paths[j:j + 1]
stats_arr = np.zeros((1, 2 * m))
mask, _ = nrrd.read(expr.test_paths[0][-1])
for t in range(m):
img, _ = nrrd.read(expr.test_paths[0][t])
stats_arr[0, 2 * t:2 * (t + 1)] = np.array([
np.mean(img[~np.isnan(mask)]),
np.std(img[~np.isnan(mask)])
])
expr.test_stats = stats_arr
scores[j, i], test_preds = expr.test_eval(model, sess)
np.savetxt(save_dir, scores)
print(j, end=',')
print()
def test_eval(self):
input_ = np.random.random((5, 3))
true = np.random.random((5, 4))
nn = NN.NN(loss_func=loss_fs.MeanSquaredLoss())
nn.add_layer(
layers.DenseLayer(n_neurons=8, activation_func=af.Sigmoid()))
nn.add_layer(
layers.DenseLayer(n_neurons=4, activation_func=af.Sigmoid()))
loss = nn.eval(input_, true)
assert loss > 0
assert list(nn.parameters) != []
assert list(nn.parameter_gradients) != []
def finetune_multimg(expr, model, sess, all_padded_imgs, training_inds):
s = len(training_inds)
img_ind_sizes = [len(training_inds[i]) for i in range(s)]
n = np.sum(img_ind_sizes)
m = len(all_padded_imgs[0]) - 1
b = expr.pars['b']
d3 = expr.pars['patch_shape'][2]
for t in range(expr.pars['epochs']):
batch_inds = NN.gen_batch_inds(n, b)
for i in range(len(batch_inds)):
""" Preparing the Patches/Labels """
# batch indices are global indices,
# extract local indices for each image
local_inds = patch_utils.global2local_inds(batch_inds[i],
img_ind_sizes)
# local indices --> image (voxel) indices
img_inds = [
np.array(training_inds[j])[local_inds[j]] for j in range(s)
]
b_patches, b_labels = patch_utils.get_patches_multimg(
all_padded_imgs, img_inds, expr.pars['patch_shape'],
expr.train_stats)
# stitching patches and labels
b_patches = [
b_patches[j] for j in range(len(img_inds))
if len(img_inds[j]) > 0
]
b_patches = np.concatenate(b_patches, axis=0)
b_labels = [
b_labels[j] for j in range(len(img_inds))
if len(img_inds[j]) > 0
]
b_labels = np.concatenate(b_labels)
# converting to hot-one vectors
hot_labels = np.zeros((2, len(b_labels)))
hot_labels[0, b_labels == 0] = 1
hot_labels[1, b_labels == 1] = 1
""" Doing an Optimization Iteration """
# finally we are ready to take
# optimization step
sess.run(model.train_step,
feed_dict={
model.x: b_patches,
model.y_: hot_labels,
model.keep_prob: model.dropout_rate
})
def NN_result_preview(second=False,
image_num=None,
blur=False,
pred=None,
detc=None,
pos_fals=False):
print("Load labels")
if second:
data = auxilary.read_csv(fileName='../csv_files/embedded_2.csv')
D = 22
else:
data = auxilary.read_csv(fileName='../csv_files/embedded.csv')
D = 128
N = len(data)
data_inputs = (data.iloc[:, :D])
inputs = np.zeros([N, D])
inputs = np.array(data_inputs)
labels = np.zeros([N, 1])
labels = np.array(data.iloc[:, D])
if not second:
embeddings, face_name, human_file_path = face_recognition.face_recognition(
dataset_path="../dataset/main_data/*/*",
preview=True,
image_num=image_num,
blur=blur)
else:
embeddings, face_name, human_file_path = facial_landmarks.test_preview(
blur=blur,
dataset_path="../dataset/main_data/*/*",
pred=pred,
detc=detc)
# identicals, similars = NN_results(embeddings, inputs, labels)
identicals, similars = NN.predict_input(embeddings, second=second)
idc_paths, idc_names, sim_paths, sim_names, others_paths, others_names = \
trim_NN_outputs (labels, face_name, identicals, similars, human_file_path, pos_fals = pos_fals)
show_tests.buttons(identicalls=idc_paths,
id_titles=idc_names,
similars=sim_paths,
sim_titles=sim_names,
left_overs=others_paths,
left_titles=others_names,
orig_image_path=human_file_path,
orig_title=face_name,
title1="MATCHING",
title2="SIMILARS",
title3="OTHERS")
def finetune(model, sess, expr, padded_imgs, mask, train_inds):
"""Fine-tuning a given model for
a number of epochs; written mainly
to be used within querying iterations
This function basically does same thin
as `PW_NN.PW_train_epoch_MultiModal`,
but is little handier and more brief:
it only uses indices of a single imagel,
there is no option for saving variables
in tensorboard, and also the images are
given as input arguments, hence no need
to load them separately here
"""
n = len(train_inds)
m = len(padded_imgs)
b = expr.pars['b']
patch_shape = expr.pars['patch_shape']
stats = expr.pars['stats']
for t in range(expr.pars['epochs']):
# batch-ify the data
batch_inds = NN.gen_batch_inds(n, b)
for i in range(len(batch_inds)):
img_inds = train_inds[batch_inds[i]]
patches, labels = patch_utils.\
get_patches(
padded_imgs,
img_inds,
patch_shape,
True,
mask)
# hot-one vector for labels
hot_labels = np.zeros((2, len(labels)))
hot_labels[0, labels == 0] = 1
hot_labels[1, labels == 1] = 1
# normalizing the patches
for j in range(m):
patches[:, :, :,
j] = (patches[:, :, :, j] - stats[j][0]) / stats[j][1]
# perform this iteration
# batch gradient step
sess.run(model.train_step,
feed_dict={
model.x: patches,
model.y_: hot_labels,
model.keep_prob: model.dropout_rate
})
def XOR():
print("\nRunning XOR Neural Network:")
case_0 = [[[0, 0]], [0]]
case_1 = [[[1, 0]], [1]]
case_2 = [[[0, 1]], [1]]
case_3 = [[[1, 1]], [0]]
case_base = [case_0, case_1, case_2, case_3]
nn = NN.NeuralNetwork(layer_sizes=[2, 2, 1], learning_rate=0.1)
nn.Train(10000, case_base, print_interval=100000)
nn.Test(add_bias=False)
nn.PlotError()
def __init__(self, nn, trainingSet, initialLearningRate,
learningModification, learnModFrequency, inputCount,
outputCount):
self.nn = nn
self.trainingSet = NN.PrepareTrainingSet(nn.GetDimensions()[1],
trainingSet)
self.learningRate = float(initialLearningRate)
self.learningModification = learningModification
self.learnModFrequency = learnModFrequency
self.inputCount = inputCount
self.outputCount = outputCount
self.cycles = 0
self.errors = []
def convertNN(neuralNet, formatString):
assert formatString.count("-") == (getLabel(neuralNet)[1:]).count(
"-"), "I can't add/remove a whole layer yet"
l = formatString.split("-")
nbHiddenLayer = 0
nbNeuronPerHiddenLayer = 0
if len(l) >= 3:
nbHiddenLayer = len(l) - 2
nbNeuronPerHiddenLayer = int(l[1])
compareTo = l[1]
for i in l[1:-1]:
assert i == compareTo, "I can't have hidden layers of different sizes"
targetNbNeuronsLayerWithoutBias = [int(s) for s in formatString.split("-")]
actualNbNeuronsLayerWithoutBias = [
int(s) for s in (getLabel(neuralNet)[1:]).split("-")
]
outWeights = copy.deepcopy(neuralNet.weights)
for k in range(len(neuralNet.weights)):
diffEntree = targetNbNeuronsLayerWithoutBias[
k] - actualNbNeuronsLayerWithoutBias[k]
if diffEntree < 0: #reduction de l'entree de la matrice
lastLigne = copy.deepcopy(neuralNet.weights[k][-1, :])
outWeights[k] = copy.deepcopy(
neuralNet.weights[k][:targetNbNeuronsLayerWithoutBias[k] + 1])
outWeights[k][-1, :] = lastLigne
elif diffEntree > 0: #augmentation de l'entree de la matrice
lastLigne = copy.deepcopy(neuralNet.weights[k][-1, :])
for p in range(len(neuralNet.weights[k][-1, :])):
neuralNet.weights[k][-1, p] = 0
outWeights[k] = numpy.pad(
neuralNet.weights[k], ((0, diffEntree), (0, 0)),
mode='constant'
) # def lignes de 0 ont été rajoutées. Mais on veut garder les valeurs de la dernière ligne (biais)
(outWeights[k])[-1, :] = lastLigne
else:
outWeights[k] = copy.deepcopy(neuralNet.weights[k])
diffSortie = targetNbNeuronsLayerWithoutBias[
k + 1] - actualNbNeuronsLayerWithoutBias[k + 1]
if diffSortie < 0: #reduction de la sortie de la matrice
outWeights[k] = copy.deepcopy(
outWeights[k][:, :targetNbNeuronsLayerWithoutBias[k + 1]])
elif diffSortie > 0: #augmentation de la sortie de la matrice
outWeights[k] = numpy.pad(outWeights[k], ((0, 0), (0, diffSortie)),
mode='constant')
out = NN.NeuralNetwork(outWeights[0].shape[0] - 1, nbHiddenLayer,
nbNeuronPerHiddenLayer, outWeights[-1].shape[1])
out.setWeight(copy.deepcopy(outWeights))
return out
def main():
genre_type.find_genre(
) #creating the genre types column iterating through each row
content_rating.find_content_type(
) #creating the content rating type column iterating through each row
readCSV.read_data() #creating feature columns by iteartong row in CSV
normalize.normalize_data(readCSV.numberCritics, readCSV.duration,
readCSV.directorFBLikes, readCSV.actor3FBLikes,
readCSV.actor2FBLikes, readCSV.actor1FBLikes,
readCSV.numberVotedUsers, readCSV.castFBLikes,
readCSV.numberUserReviews, readCSV.budget,
readCSV.imdbScore, readCSV.movieFBLikes)
Dataset.get_dataset(
readCSV.numberCritics, readCSV.duration, readCSV.directorFBLikes,
readCSV.actor3FBLikes, readCSV.actor2FBLikes, readCSV.actor1FBLikes,
readCSV.numberVotedUsers, readCSV.castFBLikes,
readCSV.numberUserReviews, readCSV.budget, readCSV.imdbScore,
readCSV.movieFBLikes, genre_type.action_type,
genre_type.adventure_type, genre_type.fantasy_type,
genre_type.scifi_type, genre_type.thriller_type,
genre_type.comedy_type, genre_type.family_type, genre_type.horror_type,
genre_type.war_type, genre_type.animation_type,
genre_type.western_type, genre_type.romance_type,
genre_type.musical_type, genre_type.documentary_type,
genre_type.drama_type, genre_type.history_type,
genre_type.biography_type, genre_type.mystery_type,
genre_type.crime_type, content_rating.general_type,
content_rating.parental_type, content_rating.parentStrong_type,
content_rating.restrict_type, content_rating.adults_type)
train = Dataset.Data[:3539] #training data as 70%
test = Dataset.Data[3539:] #test data as 30%
trainLabel = readCSV.gross[:3539]
testLabel = readCSV.gross[3539:]
applySVM()
applyNB()
applyLogRegression(train, test, trainLabel, testLabel)
applyLinearRegression(train, test, trainLabel, testLabel)
NN.apply_NN(Dataset.Data, readCSV.gross)
def evaluate(NNWorking, population, inputs, outputs):
'''Tests each citizen in the population against a NN topology with inputs
and outputs to generate an cumulitive fitness measurement, which should be
minimized'''
for citizen in population:
citizen[-1] = 0
NNWorking.SetNNWeights(citizen[:-1]) # Load weights into the NN
for i in range(len(inputs)):
NNWorking.SetStartingNodesValues(inputs[i]) # Load inputs into NN
NNWorking.CalculateNNOutputs() # Run the NN once
# Calculate the fitness value and let the citizen track it
# for j in range(len(NN.GetNNResults())):
# citizen[-1] += ((outputs[i][j] - NN.GetNNResults()[j]))**2
citizen[-1] += NN.calcRelativeError(NNWorking,
inputs, outputs) / len(inputs)
def start_thread(inp, activation, out_activ, outp, learn, thresh, mmntm, logger):
global count
training_inputs = []
training_data = []
count += 1
testNN = NN.main(inp, activation, out_activ, outp, learn, thresh, mmntm)
print ("DONE TRAINING")
for i in inp:
for j in i:
training_inputs.append(random.randint(0,4)) #create random inputs for testing
training_data.append(training_inputs)
training_inputs = []
for x in training_data:
testNN.SetStartingNodesValues(x)
testNN.CalculateNNOutputs()
logger.info(str(x))
logger.info(testNN.GetNNResults())
#################################################
# Title : Male or Female?
# Method : Back-Propagation Neural Networks
# Author : yang xiaolong
#################################################
from NN import *
from autonorm import*
import time
startTime = time.time()
# create a network with two input, one hidden, and one output nodes
n = NN(2, 1, 1)
opts = {'iterations': 50, 'learning rate': 0.25, 'momentum factor': 0.1}
def loadData():
te=[];te_x=[];te_y=[]
fileIn = open('data.txt')
for line in fileIn.readlines():
lineArr = line.strip().split()
te_x.append([float(lineArr[0]),float(lineArr[1])])
te_y.append([float(lineArr[2])])
te.append([[float(lineArr[0]),float(lineArr[1])],[float(lineArr[2])]])
return te_x,te_y,te
tex,tey,te= loadData()
tex1=AutoNorm(tex)
tey1=tey
te1=[]
for i in range(len(tex1)):
te1.append([tex1[i],tey1[i]])
def train(inputs, outputs, size, generations, threshold, cRate, mRate, printFile=False):
"""The train method creates a neural netwrok from the sets of
inputs and outputs. A population vector of size, is initialized
with ranodm weight vectors associated with the weights between
nodes in the neural network and will be the values being trained.
Generations is the max number of generations allowed while
threshold is the accuracy needed. cRate and mRate are the
crossover and mutation rates respectively."""
global hero
global OrigAnswers
OrigAnswers = copy.deepcopy(outputs)
# set up NN
EvaluationNN = GA.create_net(inputs, outputs)
# initialize population of size as random weights of NN
population = GA.generatePopulation(EvaluationNN, inputs, outputs, size)
if printFile:
f = open("DE.csv", "w")
gen = 0
trialV = []
offspringV = []
# evaluate the entire population
GA.evaluate(EvaluationNN, population, inputs, outputs)
# loop until a hero is found or we've reached max generations
while gen <= generations and hero == 0:
for i in range(size):
# mutate with DE/x/1/bin
trialV = mutate(population, i, mRate)
# perform binomial crossover
offspringV = crossover(population[i], trialV, cRate)
# evaluation of offspring
GA.evaluate(EvaluationNN, [offspringV], inputs, outputs)
# selection of better vector
if population[i][-1] > offspringV[-1]:
population[i] = offspringV
population = sorted(population, key=itemgetter(-1))
# check for hero in population
if GA.heroFound(population, threshold):
break
else:
print("Training: {:2.2%}".format(population[0][-1]), "{:2.2%} ".format(gen / generations), end="\r")
if printFile:
f.write("%f," % population[0][-1])
if printFile:
f.write("\n")
gen += 1
# return best hero if max generations is met and hero hasn't been selected.
# hero = sorted(population, key=itemgetter(-1))[0] # default to best in
# population if no hero steps forward
if printFile:
f.close()
if hero == 0:
gen -= 1
hero = sorted(population, key=itemgetter(-1))[0]
EvaluationNN.SetNNWeights(hero[:-1]) # Load hero into NN, prep for usage.
# Evaluate the hero on the inputs and outputs
print("Generations: %d" % gen, " " * 20)
print("Error Relative: {:2.5%}".format(NN.calcRelativeError(EvaluationNN, inputs, OrigAnswers)))
print("Least Squares: %d" % NN.calcLeastSquaresError(EvaluationNN, inputs, OrigAnswers))
print("Loss Squared: %d" % NN.calcLossSquared(EvaluationNN, inputs, OrigAnswers))
# for x in inputs:
# EvaluationNN.SetStartingNodesValues(x)
# EvaluationNN.CalculateNNOutputs()
# print(x, EvaluationNN.GetNNResults(), EvaluationNN.GetNNResultsInt(), OrigAnswers[inputs.index(x)])
print()
return EvaluationNN
trainingdatafile = open('newdata.txt', 'r')
inputdata= readRecords(trainingdatafile)
datasize = len(inputdata)
# Randomly splits the file in 80-20 split for test and train data
trainingbatch , testbatch = train_test_split(inputdata, test_size=0.2)
# Values for the no of hidden, input and output neurons
numHiddenCells = 64
numInputCells = 9
numOutputCells = len(inputdata[0]) - numInputCells
nn = NN.neuralNet(numInputCells, numHiddenCells, numOutputCells)
errTrain = nn.train(trainingbatch)
errPred = 0
for test in testbatch:
output = nn.predict(test[:numInputCells])
#print output, np.argmax(test[numInputCells:]) + 1
if (output != np.argmax(test[numInputCells:]) + 1) :
errPred = errPred + 1
print errTrain, len(trainingbatch)
print errPred, len(testbatch)
#interneuron
NN.Matrix([5,1],[5,1],sigmaR=sigmaR),
NN.Addition([5,1],sigmaR=sigmaR),
NN.ComponentwiseFunction(),
# interneuron 2
NN.Matrix([5,1],[5,1],sigmaR=sigmaR),
NN.Addition([5,1],sigmaR=sigmaR),
NN.ComponentwiseFunction(),
# output
#Matrix([5,1],[2,1],sigmaR=sigmaR),
#Addition([2,1],sigmaR=sigmaR)
NN.Matrix([5,1],[3,1],sigmaR=sigmaR),
NN.Addition([3,1],sigmaR=sigmaR)
])
# a nice simple interface
nn = NN.makeStandardNeuralNet(inputDim=2,outputDim=3,interDim=20,nInter=5,sigmaR=sigmaR)
# simple training set in 2D
n = 100
x = np.zeros([2,1,n])
z = np.zeros([2,1,n])
for i in range(n):
off = np.random.rand() > 0.5
x[:,:,i] = np.random.randn(2,1) + off*3
z[0,:,i] = float(off)
z[1,:,i] = 1.0 - float(off)
n = 200
x = np.zeros([2,1,n])
z = np.zeros([3,1,n])
for i in range(n):
category = np.random.randint(3)
hAx.imshow(I,interpolation='none')
#raise Exception
plt.pause(0.01)
I.shape=I.shape+(1,)
X[:,:,:,:,i] = I.astype(np.double)/255.0
# gender
if f.find('-f')>0:
Z[0,0,i] = 1.0
Z[1,0,i] = 0.0
else:
Z[0,0,i] = 0.0
Z[1,0,i] = 1.0
nCopies = 4
nn = NN.makeStandardConvolutionalNeuralNet(inputShape=I.shape,outputDim=2,nCopies=nCopies,sigmaR=10.0)
nn.setTrainingData(X,Z)
layersToDraw = [l for l in nn if type(l) == type(NN.ComponentwiseFunction())]
nPlots = len(layersToDraw)
hFig.clf()
hFig2 = plt.figure()
nIter = 10000
nPerDraw = 1
nRepeats = nIter/nPerDraw
epsilon = 0.0001
epsilon = 0.00001
L = 1.0e10
for repeat in range(nRepeats):
if __name__ == '__main__':
print "Part 1: Loading Data\n"
X, y = utils.loadData(conf.FILE_X, conf.FILE_Y)
print "Part 2: Loading Parameters\n"
W1, W2 = utils.loadParams(conf.FILE_W1, conf.FILE_W2)
# Unroll parameters
W = np.hstack((W1.flatten(0), W2.flatten(0)))
W = W.reshape((len(W), 1))
print "Part 3: Compute Cost(Feedforward)\n"
LEARN_RATE = 0
J, _ = NN.nnCostFunction(W, conf.INPUT_LAYER_SIZE, conf.HIDDEN_LAYER_SIZE,
conf.NUM_LABELS, X, y, LEARN_RATE)
print ("Cost at parameters (loaded from w1.txt and w2.txt): %f"
"\n(this value should be about 0.287629)\n") % J
print "Part 4: Implement Regularization\n"
LEARN_RATE = 1
J, _ = NN.nnCostFunction(W, conf.INPUT_LAYER_SIZE, conf.HIDDEN_LAYER_SIZE,
conf.NUM_LABELS, X, y, LEARN_RATE)
print ("Cost at parameters (loaded from w1.txt and w2.txt): %f"
"\n(this value should be about 0.383770)\n") % J
print "Part 5: Sigmoid Gradient\n"
g = Sigmoid.dy_dz(Sigmoid.y(np.array([-1, -0.5, 0, 0.5, 1])))
print "Sigmoid gradient evaluated at [-1 -0.5 0 0.5 1]:", g
def func(x,*args):
"""evaluate function L=sum_{samples}[E(pmd)-E(ref)]^2.
This will be called from scipy.optimize.fmin_cg().
The 1st argument x should be 1-D array of variables.
"""
global _valmin
t0= time.time()
#.....write parameters to in.params.????? file
dir= args[0]
if fmethod in ('test','TEST','check_grad') or \
not potential in ('linreg','NN'):
#.....store original file
os.system('cp '+dir+'/'+parfile+' '+dir+'/'+parfile+'.tmp')
write_params(dir+'/'+parfile,x)
#.....run smd in all sample directories
os.chdir(dir)
#print os.getcwd(),dir
if runmode in ('serial','Serial','SERIAL','sequential','single'):
os.system('./serial_run_smd.sh '+parfile)
elif runmode in ('parallel','Parallel','PARALLEL'):
os.system('python ./parallel_run_smd.py '+parfile)
else:
print "{0:*>20}: no such run_mode !!!".format(' Error', runmode)
exit()
os.chdir(cwd)
#.....restore original file
os.system('cp '+dir+'/'+parfile+' '+dir+'/'+parfile+'.current')
os.system('cp '+dir+'/'+parfile+'.tmp'+' '+dir+'/'+parfile)
#.....gather smd results
ergs,frcs=gather_smd_data(dir)
elif potential in ('linreg'):
#.....calc ergs and frcs from bases data and x (variables)
read_bases(dir)
ergs,frcs=calc_ef_from_bases(x,*args)
elif potential in ('NN'):
#.....now it is possible to compute only from bases
ergs,frcs= NN.calc_ef_from_bases(x,*args)
#.....calc function value of L
val= eval_L(ergs,frcs,ergrefs,frcrefs,samples)
#.....output temporal results
output_energy_relation(ergs,ergrefs,samples,sample_dirs, \
fname='out.erg.pmd-vs-dft.tmp')
output_force_relation(frcs,frcrefs,samples,sample_dirs, \
fname='out.frc.pmd-vs-dft.tmp')
print
print ' L value=',val
if penalty in ('ridge','Ridge','RIDGE') and potential in ('linreg'):
p= 0.0
lx= len(x)
for n in range(lx):
p += math.sqrt(x[n]**2)
print ' penalty value=',p*pweight
val += p*pweight
print ' total L value=',val
elif penalty in ('lasso','LASSO') and potential in ('linreg'):
p= 0.0
lx= len(x)
for n in range(lx):
p += abs(x[n])
print ' penalty value=',p*pweight
val += p*pweight
print ' total L value=',val
sys.stdout.flush()
#.....if L value is minimum ever, store this parameter file
if val < _valmin:
_valmin= val
if potential in ('linreg','NN'):
write_params(dir+'/'+parfile+'.min',x)
else:
os.system('cp '+dir+'/'+parfile+'.current' \
+' '+dir+'/'+parfile+'.min')
print ' ===> time func: {0:12.3f} sec'.format(time.time()-t0) \
+', {0:12.3f} sec'.format(time.time()-_init_time)
return val
sample_dirs.sort()
if nsmpl != len(sample_dirs):
print '{0:*>20}: num_samples in in.fitpot is wrong.'.format(' Error')
exit()
read_pos()
#.....initial data
gather_ref_data(maindir)
#.....read bases data if needed
if potential in ('linreg') and not fmethod in ('test','TEST'):
read_bases(maindir)
if regularize:
vars= scale_vars(vars,bmax)
elif potential in ('NN') and not fmethod in ('test','TEST'):
NN.init(maindir,params,sample_dirs,samples,nprcs,fmatch \
,ergrefs,frcrefs,fmethod,parfile,runmode,rcut,pranges \
,vranges)
#.....1st call of func
func(vars,maindir)
if potential in ('linreg') and not fmethod in ('test','TEST'):
ergs,frcs= calc_ef_from_bases(vars,maindir)
elif potential in ('NN') and not fmethod in ('test','TEST'):
ergs,frcs= NN.calc_ef_from_bases(vars)
else:
ergs,frcs= gather_smd_data(maindir)
if fmethod in ('test','TEST') and potential in ('NN'):
NN.init(maindir,params,sample_dirs,samples,nprcs,fmatch \
,ergrefs,frcrefs,fmethod,parfile,runmode \
,rcut,pranges,vranges)
def train(inputs, outputs, size, participants, victors,
generations, threshold, cRate, mRate, printFile=False):
'''The train method takes in a set of inputs and outputs which will be
compared against a hardcoded NN topology. The size, participants, and
victors are with regard to tournament selection and elitism selection
techniques. Generations is the max number of generations allowed while
threshold is the accuracy needed. cRate and mRate are the rate of
crossover and rate of mutation respectively. '''
global hero
global OrigAnswers
EvaluationNN = create_net(inputs, outputs)
population = generatePopulation(EvaluationNN, inputs, outputs, size)
# Test each citizen and determine initial fitness
evaluate(EvaluationNN, population, inputs, outputs)
if printFile: f = open('GA.csv', 'w')
gen = 0
children = []
# loop until a hero is found or we've reached max generations
while gen <= generations and hero == 0:
# Select our parents using tournament selection
parents = tournament(population, participants, victors)
# Have our parents mate (Crossover)
children = mate(parents, cRate)
# Have the children experience the world (Mutate)
for child in children:
mutate(child, mRate)
# Test each child's fitness
evaluate(EvaluationNN, children, inputs, outputs)
# We were to prolific, thus children must fight to the death via draft
# call. Make participants len(children) to have all of them fight
# This might not be a good idea as late generation counts result in not
# keeping the children.
children = tournament(children, participants, victors)
# purging of population is determined by elitism inverted on fitness
# level (cowardace is greater number).
# Take number of children equal to number of tournament victors and
# reintroduce to the population
population = sorted(population + children,
key=itemgetter(-1))[:-victors]
# Determine if a child is a hero (<threshold) and if so, return child
if heroFound(population, threshold):
break
else:
print("Training: {:2.2%}".format(
population[0][-1]), "{:2.2%} ".format(gen / generations), end="\r")
if printFile: f.write('%f,' % population[0][-1])
if printFile: f.write('\n')
gen += 1
# return best hero if max generations is met and hero hasn't been selected.
if printFile: f.close()
if hero == 0:
gen -= 1
hero = sorted(population, key=itemgetter(-1))[0]
EvaluationNN.SetNNWeights(hero[:-1]) # Load hero into NN, prep for usage.
# Evaluate the hero on the inputs and outputs
print('Generations: %d' % gen, ' ' * 20)
print("Error Relative: {:2.5%}".format(NN.calcRelativeError(EvaluationNN, inputs, OrigAnswers)))
print("Least Squares: %d" % NN.calcLeastSquaresError(EvaluationNN, inputs, OrigAnswers))
print("Loss Squared: %d" % NN.calcLossSquared(EvaluationNN, inputs, OrigAnswers))
#for x in inputs:
# EvaluationNN.SetStartingNodesValues(x)
# EvaluationNN.CalculateNNOutputs()
# print(x, EvaluationNN.GetNNResults(), EvaluationNN.GetNNResultsInt(), OrigAnswers[inputs.index(x)])
print()
return EvaluationNN
def costFunc(p):
return NN.nnCostFunction(p, conf.INPUT_LAYER_SIZE,
conf.HIDDEN_LAYER_SIZE, conf.NUM_LABELS,
X, y, conf.PART8_LEARN_RATE)
import weather
import NN as nn
get_pm5_prediction = nn.setup()
# SAMPLE usage, Delete before using it as library
def predictPollution(precipitation_prob, relative_humidity, temp, wind_direction, wind_speed):
O3PreictedVal = 5.03704930e+01 + (precipitation_prob * 9.66895471e-02) + (relative_humidity * -2.99780572e-03) + (temp * -2.26017118e-01) + (wind_direction * -8.96663780e-03) + (wind_speed * 9.98339351e+00)
PM25PredictedVal = 1.36006991e+01 + (temp * -9.32461073e-02) + (wind_direction * -3.35510810e-04) + (wind_speed * -7.50369156e-01)
nnPrediction = get_pm5_prediction(TMP = temp,WDIR = wind_direction,WSPD = wind_speed)
#3.6 is average
if abs(nnPrediction - PM25PredictedVal) > 7.2:
if abs(nnPrediction - 3.6) > abs(PM25PredictedVal - 3.6):
return O3PreictedVal, PM25PredictedVal
else:
return O3PreictedVal, nnPrediction
return O3PreictedVal, (nnPrediction + PM25PredictedVal)/2
def pollutionAPi(lat, lon, offset):
return predictPollution(*weather.get_weather(lat, lon, offset))
#offset -> [0,2], 0 means now, 1 means one hour from now, and 2 means 2 hour from now
print pollutionAPi(51.0123, 0.3, 0)
return tmp
'''
Convert neural network data to game situation.
'''
def NN2board(li):
tmp = ["X" if x==1 else ("O" if x==-1 else " ") for x in li]
tmp = np.array(tmp)
tmp.shape = (7,6)
tmp = tmp.T
return tmp
################################################################
try:
nn = NN.openNN()
print "NN successfully opened"
except:
print "not able to open NN"
print "create new one"
inp = []
out = []
fobj = open("database.txt")
counter=0
for line in fobj:
#print line.rstrip()
#displayStr(line.rstrip())
def train_test():
global cRate, mRate, threshold, generations, size, participants, victors, inFile, algo, dataset, resultsFile
inputs = []
outputs = []
evolve()
resultsFile.write("DATASET: " + dataset + "\n")
#resultsFile.write("ALGORITHM | Generations | Size | Participants | Victors | mRate | cRate | Threshold \n")
#resultsFile.write(" " + str(algo) + " | " + str(generations) + " | " +
# str(size) + " | " + str(participants) + " | " + str(victors) +
# " | " + str(mRate) + " | " + str(cRate) + " | " + str(threshold) + " \n")
dataIn = dataHandler()
inputs = dataIn[0]
outputs = dataIn[1]
testInput = []
testOutput = []
learnrate = 0.3
momentum = 0.5
# Need 20% of inputs for testing
for i in range((int(len(inputs)*0.8)+1), len(inputs)):
x = random.choice(inputs)
testInput.append(x)
testOutput.append(outputs[inputs.index(x)])
del outputs[inputs.index(x)]
del inputs[inputs.index(x)]
resultsFile.write("\nTest inputs: \n")
for i in range(len(testInput)):
resultsFile.write("%s " % testInput[i])
resultsFile.write("\nTest expected outputs: \n")
for i in range(len(testOutput)):
resultsFile.write("%s " % testOutput[i])
# Which algorithm gets chosen to run
if algo in 'G':
print("DOING GA TRAINING...")
resultsFile.write("\nALGORITHM | Generations | Size | Participants | Victors | mRate | cRate | Threshold \n")
resultsFile.write(" " + str(algo) + " | " + str(generations) + " | " + str(size) + " | " + str(participants) + " | " + str(victors) + " | " + str(mRate) + " | " + str(cRate) + " | " + str(threshold) + " \n")
testNN = GA.train(inputs, outputs, size, participants, victors, generations, threshold, cRate, mRate)
elif algo in 'E':
print("DOING ES TRAINING...")
resultsFile.write("\nALGORITHM | Generations | Size | Participants | Victors | mRate | cRate | Threshold \n")
resultsFile.write(" " + str(algo) + " | " + str(generations) + " | " + str(size) + " | " + str(participants) + " | " + str(victors) + " | " + str(mRate) + " | " + str(cRate) + " | " + str(threshold) + " \n")
testNN = ES.train(inputs, outputs, size, participants, victors, generations, threshold, cRate, mRate)
elif algo in 'D':
print("DOING DE TRAINING...")
resultsFile.write("\nALGORITHM | Generations | Size | mRate | cRate | Threshold \n")
resultsFile.write(" " + str(algo) + " | " + str(generations) + " | " + str(size) + " | " + str(mRate) + " | " + str(cRate) + " | " + str(threshold) + " \n")
testNN = DE.train(inputs, outputs, size, generations, threshold, cRate, mRate)
elif algo in 'B':
print("DOING BP TRAINING...")
resultsFile.write("\nALGORITHM | Generations | learnrate | momentum | Threshold \n")
resultsFile.write(" " + str(algo) + " | " + str(generations) + " | " + str(learnrate) + " | " + str(momentum) + " | " + str(threshold) + " \n")
testNN = NN.main(inputs, [['S','S','S'], ['S','S']], ['S'], outputs, generations, learnrate, threshold, momentum)
else:
print("Unrecognized algorithm!")
sys.exit()
# Print test input/expected output - could be made prettier in a table
# Start testing testNN
for x in testInput:
resultsFile.write("\nSet starting node vals\n")
resultsFile.write("%s \n" % testNN.SetStartingNodesValues(x))
testNN.CalculateNNOutputs()
resultsFile.write("\nTest Input: " + str(x) + "\n")
resultsFile.write("\nTest results: %s\n" % testNN.GetNNResults())
resultsFile.write("\nRelative Error: {:2.2%} \n".format(NN.calcRelativeError(testNN, testInput, testOutput)))
resultsFile.write("\nLeast Squares Error: %s \n" % NN.calcLeastSquaresError(testNN, testInput, testOutput))
resultsFile.write("\nLoss Squared Error: %s \n" % NN.calcLossSquared(testNN, testInput, testOutput))
resultsFile.write("\nPercent Misidentified: {:2.2%} \n".format(NN.calcPercentIncorrect(testNN, testInput, testOutput)))
resultsFile.close()
print('Attempting to open {}.'.format(filename))
im = readImage(filename)
if (max(im.size) > 600):
im = im.resize((int(600*(float(im.size[0])/max(im.size))), int(600*(float(im.size[1])/max(im.size)))))
print('Opened.')
print('Attempting to stipple...')
cellSize = 2
# Create a stippled version of the image; limit 6000 px.
lst = stipple(im, cellSize, 0)
while (len(lst) > 6000):
cellSize += 1
lst = stipple(im, cellSize, 0)
lst = stipple(im, cellSize, 8)
print('There are {} points.'.format(len(lst)))
print('Stippled!')
print('Attempting TSP with naive NN...')
lst = NN.tsp(lst)
print('Now converting to list of segments...')
segSet = createSegSet(lst)
## Let's make sure all of our segments share a point...
print('Correcting any overlaps...')
drawSegSet(segSet, im.size, 'start.jpg')
segSet = correct(segSet, im)
drawSegSet(segSet, im.size, 'end.jpg')
print('Done.')
