max_iter=500,
hidden_layer_sizes=(20, 10, 3),
random_state=1)
svc = SVC(kernel='linear')
csp = CSP(n_components=8, reg=None, log=True, norm_trace=False)
x, y = data.load_single(cut=True)
y = [np.where(yi == 1)[1] for yi in y]
xt, yt = data.load_single(cut=True, study=False)
yt = [np.where(yi == 1)[1] for yi in yt]
subs = len(x)
splits = 10
for name, classifier in [('LDA', lda), ('SVM', svc), ('MLP', mlp)]:
acc = 0
acc2 = 0
for sub in range(18):
n = x[sub].shape[0]
for tr, val in util.kfold(n, splits, shuffle=True):
clf = Pipeline([('CSP', csp), ('LDA', lda)])
#clf.set_params(CSP__reg=0.5)
clf.fit(x[sub][tr], y[sub][tr])
a = clf.score(x[sub][val], y[sub][val])
a2 = clf.score(xt[sub], yt[sub])
acc += a
acc2 += a2
print("{} got {}/{}".format(name, acc / (splits * subs),
acc2 / (splits * subs)))
m_save = model.get_config()
model.summary()
avgacc = 0
wlist = np.zeros((18, ), dtype=np.object)
for i in range(n_subs):
n = x[i].shape[0]
acc = 0
model = Model.from_config(m_save)
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
tr, val = util.kfold(n, splits)[0]
# fit with next kfold data
h = model.fit(x[i][tr], y[i][tr], batch_size=64, epochs=50, verbose=0)
_, acc = model.evaluate(x[i][val], y[i][val], verbose=0)
print("subject {}, avg accuracy {}".format(i + 1 if i + 1 < 10 else i + 2,
acc))
layers = model.layers
print(len(layers))
print(layers[3])
w0 = model.layers[3].get_weights()[0]
print(w0.shape)
wlist[i] = w0
model.add(Flatten())
model.add(Dense(mset["dense_nodes"], activation='tanh'))
model.add(Dense(3, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()
print(mset)
w_save = model.get_weights()
avgacc = 0
for i in range(n_subs):
n = x[i].shape[0]
acc = 0
for tr, val in util.kfold(n, splits):
# reset to initial weights
model.set_weights(w_save)
# fit with next kfold data
model.fit(x[i][tr], y[i][tr],
batch_size=64, epochs=90, verbose=0)
loss, accuracy = model.evaluate(x[i][val], y[i][val],
verbose=0)
acc += accuracy
acc /= splits
avgacc += acc
print("subject {}, avg accuracy {} over {} splits".format(i + 1 if i + 1 < 10 else i + 2, acc, splits))