def learning(learning_path, synapse):
#potentials of output neurons
potential_lists = []
for i in range(par.kSecondLayerNuerons_):
potential_lists.append([])
#time series
time_array = np.arange(1, par.kTime_+1, 1)
layer2 = []
# creating the hidden layer of neurons
for i in range(par.kSecondLayerNuerons_):
a = neuron()
layer2.append(a)
for epoch in range(1):
for wave_file in learning_path:
#for wave_file in ["sounddata\F1\F1SYB01_が.wav"]:
resemble_print(str(wave_file) + " " + str(epoch))
#音声データの読み込み
splited_sig_array, samplerate = wav_split(str(wave_file))
resemble_print(str(wave_file))
splited_sig_array = remove_silence(splited_sig_array)
print(len(splited_sig_array))
#スパイクの連結
#spike_train = wav_split2spike(splited_sig_array, samplerate)
#spike_connected = np.array(connect_spike(spike_train))
#for spike_train in spike_connected:
for signal in splited_sig_array:
#Generating melspectrum
f_centers, mel_spectrum = get_log_melspectrum(signal, samplerate)
#Generating spike train
spike_train = np.array(encode(np.log10(mel_spectrum)))
#calculating threshold value for the image
var_threshold = threshold(spike_train)
# resemble_print var_threshold
# synapse_act = np.zeros((par.kSecondLayerNuerons_,par.kFirstLayerNuerons_))
# var_threshold = 9
# resemble_print var_threshold
# var_D = (var_threshold*3)*0.07
var_D = 0.15 * par.kScale_
for x in layer2:
x.initial(var_threshold)
#flag for lateral inhibition
flag_spike = 0
img_win = 100
active_potential = []
for index1 in range(par.kSecondLayerNuerons_):
active_potential.append(0)
#Leaky integrate and fire neuron dynamics
for time in time_array:
for second_layer_position, second_layer_neuron in enumerate(layer2):
active = []
if(second_layer_neuron.t_rest < time):
second_layer_neuron.P = (second_layer_neuron.P
+ np.dot(
synapse[second_layer_position], spike_train[:, time]
)
)
#resemble_print("synapse : " + str(synapse[second_layer_position]))
if(second_layer_neuron.P > par.kPrest_):
second_layer_neuron.P -= var_D
active_potential[second_layer_position] = second_layer_neuron.P
potential_lists[second_layer_position].append(second_layer_neuron.P)
# Lateral Inhibition
if(flag_spike==0):
max_potential = max(active_potential)
if(max_potential > var_threshold):
flag_spike = 1
winner_neuron = np.argmax(active_potential)
img_win = winner_neuron
resemble_print("winner is " + str(winner_neuron))
for s in range(par.kSecondLayerNuerons_):
if(s != winner_neuron):
layer2[s].P = par.kMinPotential_
#Check for spikes and update weights
for second_layer_position, second_layer_neuron in enumerate(layer2):
neuron_status = second_layer_neuron.check()
if(neuron_status == 1):
second_layer_neuron.t_rest = time + second_layer_neuron.t_ref
second_layer_neuron.P = par.kPrest_
for first_layer_position in range(par.kFirstLayerNuerons_):
#前シナプスの計算
for back_time in range(-2, par.kTimeBack_-1, -1): #-2 → -20
if 0 <= time + back_time < par.kTime_ + 1:
if spike_train[first_layer_position][time + back_time] == 1:
# resemble_print "weight change by" + str(update(synapse[j][h], rl(t1)))
synapse[second_layer_position][first_layer_position] = update(
synapse[second_layer_position][first_layer_position], rl(back_time)
)
resemble_print("back : " + str(second_layer_position) + "-" + str(first_layer_position) + " : " + str(synapse[second_layer_position][first_layer_position]))
#後シナプスの計算
for fore_time in range(2, par.kTimeFore_+1, 1): # 2 → 20
if 0 <= time + fore_time<par.kTime_+1:
if spike_train[first_layer_position][time + fore_time] == 1:
# resemble_print "weight change by" + str(update(synapse[j][h], rl(t1)))
synapse[second_layer_position][first_layer_position] = update(
synapse[second_layer_position][first_layer_position], rl(fore_time)
)
resemble_print("fron : " + str(second_layer_position) + "-" + str(first_layer_position) + " : " + str(synapse[second_layer_position][first_layer_position]))
if(img_win!=100):
for first_layer_position in range(par.kFirstLayerNuerons_):
if sum(spike_train[first_layer_position]) == 0:
synapse[img_win][first_layer_position] -= 0.06 * par.kScale_
if(synapse[img_win][first_layer_position]<par.kMinWait_):
synapse[img_win][first_layer_position] = par.kMinWait_
return potential_lists, synapse, layer2
for j in range(par.m):
synapse[i][j] = random.uniform(0, 0.4 * par.scale)
for k in range(par.epoch):
for i in range(322, 323):
print i, " ", k
img = cv2.imread("mnist1/" + str(i) + ".png", 0)
#Convolving image with receptive field
pot = rf(img)
#Generating spike train
train = np.array(encode(pot))
#calculating threshold value for the image
var_threshold = threshold(train)
# print var_threshold
# synapse_act = np.zeros((par.n,par.m))
# var_threshold = 9
# print var_threshold
# var_D = (var_threshold*3)*0.07
var_D = 0.15 * par.scale
for x in layer2:
x.initial(var_threshold)
#flag for lateral inhibition
f_spike = 0
def learning(learning_or_classify):
#1 = learning, 0 = classify
#learning_or_classify = 0
print learning_or_classify
if(learning_or_classify == 0):
print "Starting classify..."
elif(learning_or_classify == 1):
print "Starting learning..."
else:
print "Error in argument, quitting"
quit()
if(learning_or_classify == 0):
par.epoch = 1
#potentials of output neurons
pot_arrays = []
pot_arrays.append([]) #because 0th layer do not require neuron model
for i in range(1,par.num_layers):
pot_arrays_this = []
for j in range(0,par.num_layer_neurons[i]):
pot_arrays_this.append([])
pot_arrays.append(pot_arrays_this)
print "created potential arrays for each layer..."
Pth_array = []
Pth_array.append([]) #because 0th layer do not require neuron model
for i in range(1,par.num_layers):
Pth_array_this = []
for j in range(0,par.num_layer_neurons[i]):
Pth_array_this.append([])
Pth_array.append(Pth_array_this)
print "created potential threshold arrays for each layer..."
train_all = []
for i in range(0,par.num_layers):
train_this = []
for j in range(0,par.num_layer_neurons[i]):
train_this.append([])
train_all.append(train_this)
print "created spike trains for each layer..."
#synapse matrix initialization
synapse = [] #synapse[i] is the matrix for weights from layer i to layer i+1, assuming index from 0
for i in range(0,par.num_layers-1):
synapse_this = np.zeros((par.num_layer_neurons[i+1],par.num_layer_neurons[i]))
synapse.append(synapse_this)
if(learning_or_classify == 1):
for layer in range(0,par.num_layers-1):
for i in range(par.num_layer_neurons[layer+1]):
for j in range(par.num_layer_neurons[layer]):
synapse[layer][i][j] = random.uniform(0,0.4*par.scale)
else:
for layer in range(0,par.num_layers-1):
for i in range(par.num_layer_neurons[layer+1]):
#for j in range(par.num_layer_neurons[layer]):
filename = "weights/layer_"+str(layer)+"_neuron_"+str(i)+".dat"
with open(filename,"rb") as f:
synapse[layer][i] = pickle.load(f)
print "created synapse matrices for each layer..."
#this contains neurons of all layers except first
layers = [] #layers[i] is the list of neurons from layer i, assuming index from 0
layer_this = []
layers.append(layer_this) #0th layer is empty as input layer do not require neuron model
#time series
time = np.arange(1, par.T+1, 1)
# creating each layer of neurons
for i in range(1,par.num_layers):
layer_this = []
for i in range(par.num_layer_neurons[i]):
a = neuron()
layer_this.append(a)
layers.append(layer_this)
print "created neuron for each layer..."
for k in range(par.epoch):
for i in range(1,7):
print "Epoch: ",str(k),", Image: ", str(i)
if(learning_or_classify == 1):
img = cv2.imread("training_images/" + str(i) + ".png", 0)
else:
img = cv2.imread("training_images/" + str(i) + ".png", 0)
#Convolving image with receptive field
pot = rf(img)
#print pot
#training layers i and i+1, assuming 0 indexing, thus n layers require n-1 pairs of training
for layer in range(0,par.num_layers-1):
print "Layer: ", str(layer)
#Generating spike train when the first layer
#else take the spike train from last layer
if(layer == 0):
train_all[layer] = np.array(encode(pot))
train = np.array(encode(pot))
else:
train_all[layer] = np.asarray(train_this_layer)
train = np.array(np.asarray(train_this_layer))
#print train[1]
#calculating threshold value for the image
var_threshold = threshold(train)
#print "var_threshold is ", str(var_threshold)
# print var_threshold
# synapse_act = np.zeros((par.n,par.m))
# var_threshold = 9
# print var_threshold
# var_D = (var_threshold*3)*0.07
var_D = 0.15*par.scale
for x in layers[layer+1]:
x.initial(var_threshold)
#flag for lateral inhibition
f_spike = 0
img_win = 100
active_pot = []
train_this_layer = []
for index1 in range(par.num_layer_neurons[layer+1]):
active_pot.append(0)
train_this_layer.append([])
#print synapse[layer].shape, train.shape
#Leaky integrate and fire neuron dynamics
for t in time:
#print "Time: ", str(t)
for j, x in enumerate(layers[layer+1]):
active = []
if(x.t_rest<t):
x.P = x.P + np.dot(synapse[layer][j], train[:,t])
if(x.P>par.Prest):
x.P -= var_D
active_pot[j] = x.P
#pot_arrays[layer+1][j].append(x.P)
#Pth_array[layer+1][j].append(x.Pth)
# Lateral Inhibition
# Occurs in the training of second last and last layer
#if(f_spike==0 and layer == par.num_layers - 2 and learning_or_classify == 1):
if(f_spike==0 ):
high_pot = max(active_pot)
if(high_pot>var_threshold):
f_spike = 1
winner = np.argmax(active_pot)
img_win = winner
#print "winner is " + str(winner)
for s in range(par.num_layer_neurons[layer+1]):
if(s!=winner):
layers[layer+1][s].P = par.Pmin
#Check for spikes and update weights
for j,x in enumerate(layers[layer+1]):
pot_arrays[layer+1][j].append(x.P)
Pth_array[layer+1][j].append(x.Pth)
s = x.check()
train_this_layer[j].append(s)
if(learning_or_classify == 1):
if(s==1):
x.t_rest = t + x.t_ref
x.P = par.Prest
for h in range(par.num_layer_neurons[layer]):
for t1 in range(-2,par.t_back-1, -1):
if 0<=t+t1<par.T+1:
if train[h][t+t1] == 1:
# print "weight change by" + str(update(synapse[j][h], rl(t1)))
synapse[layer][j][h] = update(synapse[layer][j][h], rl(t1))
for t1 in range(2,par.t_fore+1, 1):
if 0<=t+t1<par.T+1:
if train[h][t+t1] == 1:
# print "weight change by" + str(update(synapse[j][h], rl(t1)))
synapse[layer][j][h] = update(synapse[layer][j][h], rl(t1))
for j in range(par.num_layer_neurons[layer+1]):
train_this_layer[j].append(0)
#if(img_win!=100 and layer == par.num_layers - 2 ):
if(img_win!=100 ):
for p in range(par.num_layer_neurons[layer]):
if sum(train[p])==0:
synapse[layer][img_win][p] -= 0.06*par.scale
if(synapse[layer][img_win][p]<par.w_min):
synapse[layer][img_win][p] = par.w_min
#print train_this_layer
#print synapse[0][0]
train_all[par.num_layers-1] = np.asarray(train_this_layer)
results_each_layer = 1
if(results_each_layer):
for layer in range(par.num_layers-1,par.num_layers):
for i in range(par.num_layer_neurons[layer]):
print "Layer"+ str(layer) + ", Neuron"+str(i+1)+": "+str(sum(train_all[layer][i]))
#print classification results
# if(learning_or_classify == 0):
plot = 0
if (plot == 1):
for layer in range(par.num_layers-1,par.num_layers):
ttt = np.arange(0,len(pot_arrays[layer][0]),1)
#plotting
for i in range(par.num_layer_neurons[layer]):
axes = plt.gca()
axes.set_ylim([-20,50])
plt.plot(ttt,Pth_array[layer][i], 'r' )
plt.plot(ttt,pot_arrays[layer][i])
plt.show()
#Reconstructing weights to analyse training
reconst = 1
if(learning_or_classify != 1):
reconst = 0
if(reconst == 1):
for layer in range(par.num_layers-1):
siz_x = int(par.num_layer_neurons[layer]**(.5))
siz_y = siz_x
for i in range(par.num_layer_neurons[layer+1]):
reconst_weights(synapse[layer][i],i+1,layer,siz_x,siz_y)
#Dumping trained weights of last layer to file
dump = 1
if(learning_or_classify != 1):
dump = 0
if(dump == 1):
for layer in range(par.num_layers-1):
for i in range(par.num_layer_neurons[layer+1]):
filename = "weights/"+"layer_"+str(layer)+"_neuron_"+str(i)+".dat"
with open(filename,'wb') as f:
#f.write(str(synapse[layer][i]))
pickle.dump(synapse[layer][i],f)
synapse[i][j] = random.uniform(0, 0.4 * par.kScale_)
for epoch in range(par.kEpoch_):
for image_num in range(3221, 3222):
print(str(image_num) + " " + str(epoch))
img = cv2.imread("mnist1/" + str(image_num) + ".png", 0)
print("mnist1/" + str(image_num) + ".png")
#Convolving image with receptive field
pot = rf(img)
#Generating spike train
spike_train = np.array(encode(pot))
#calculating threshold value for the image
var_threshold = threshold(spike_train)
# print var_threshold
# synapse_act = np.zeros((par.kSecondLayerNuerons_,par.kFirstLayerNuerons_))
# var_threshold = 9
# print var_threshold
# var_D = (var_threshold*3)*0.07
var_D = 0.15 * par.kScale_
for x in layer2:
x.initial(var_threshold)
#flag for lateral inhibition
flag_spike = 0
def winner_take_all(synapse, wave_file):
#potentials of output neurons
potential_lists = []
for i in range(par.kSecondLayerNuerons_):
potential_lists.append([])
#time series
time_array = np.arange(1, par.kTime_ + 1, 1)
layer2 = []
# creating the hidden layer of neurons
for i in range(par.kSecondLayerNuerons_):
a = neuron()
layer2.append(a)
neuron_spiked = np.zeros(par.kSecondLayerNuerons_)
for epoch in range(1):
resemble_print(str(wave_file) + " " + str(epoch))
#音声データの読み込み
#splited_sig_array, samplerate = wav_split(str(wave_file))
#resemble_print(wave_file)
splited_sig_array, samplerate = wav_split(str(wave_file))
resemble_print(str(wave_file))
#スパイクの連結
#spike_train = wav_split2spike(splited_sig_array, samplerate)
#spike_connected = np.array(connect_spike(spike_train))
#for spike_train in spike_connected:
for signal in splited_sig_array:
#Generating melspectrum
f_centers, mel_spectrum = get_log_melspectrum(signal, samplerate)
#Generating spike train
spike_train = np.array(encode(np.log10(mel_spectrum)))
#calculating threshold value for the image
var_threshold = threshold(spike_train)
# resemble_print var_threshold
# synapse_act = np.zeros((par.kSecondLayerNuerons_,par.kFirstLayerNuerons_))
# var_threshold = 9
# resemble_print var_threshold
# var_D = (var_threshold*3)*0.07
var_D = 0.15 * par.kScale_
for x in layer2:
x.initial(var_threshold)
#flag for lateral inhibition
flag_spike = 0
img_win = 100
active_potential = []
for index1 in range(par.kSecondLayerNuerons_):
active_potential.append(0)
#Leaky integrate and fire neuron dynamics
for time in time_array:
for second_layer_position, second_layer_neuron in enumerate(
layer2):
active = []
if (second_layer_neuron.t_rest < time):
second_layer_neuron.P = (
second_layer_neuron.P +
np.dot(synapse[second_layer_position],
spike_train[:, time]))
#resemble_print("synapse : " + str(synapse[second_layer_position]))
if (second_layer_neuron.P > par.kPrest_):
second_layer_neuron.P -= var_D
active_potential[
second_layer_position] = second_layer_neuron.P
potential_lists[second_layer_position].append(
second_layer_neuron.P)
# Lateral Inhibition
if (flag_spike == 0):
max_potential = max(active_potential)
if (max_potential > var_threshold):
flag_spike = 1
winner_neuron = np.argmax(active_potential)
img_win = winner_neuron
neuron_spiked[winner_neuron] += 1
resemble_print("winner is " + str(winner_neuron))
for s in range(par.kSecondLayerNuerons_):
if (s != winner_neuron):
layer2[s].P = par.kMinPotential_
#勝ったニューロンの特定
resemble_print(neuron_spiked)
return neuron_spiked