def forward_loss(self, input, target):
self.input = input
self.target = target
self.logsoftmax_module = LogSoftmax()
self.logsoftmax_out = self.logsoftmax_module.forward(self.input)
self.batch_size, self.target_num = self.input.shape
# print('target:', self.target)
# print('batch_size:', self.batch_size)
# print('target_num:', self.target_num)
self.target_one_hot = get_one_hot(self.target, self.target_num)
# print('target_one_hot:', self.target_one_hot)
nll_log = -self.logsoftmax_out * self.target_one_hot
# print('nll_log:', nll_log)
return 1.0 / self.batch_size * np.sum(nll_log)
def main():
#set hyperparameter at here
hiddenlayerlist = [[
16, 32, 16
]] #change the number of hidden layer, and nodes in the layer
ss = 1e-2 #step Size
numofiter = 300 #iterations
size = 2500 #input size
dim = 2 #input dimension
margin = 0 #change Margin at here, change this value to 0 to make the data not linear separable
output_unit = 1
algorithm = input('Select algorithm: (input ebp, r+, r-, ir+ or ir-)')
# algorithm = 'r+'
modeltype = input(
'Classification or Regression? (input c, r, mnist or bc)')
modeltype = 'mnist'
if modeltype == 'c':
#generate the input and output for classification
inputdata, outputdata = generatedata(size, dim, margin)
#plot to viaualize if it is 1D
print('Training Data Plot: ')
plt.figure(1)
if dim == 1:
plt.scatter(inputdata[:size // 2, 0],
np.ones((size // 2, 1)),
color='r')
plt.scatter(inputdata[size // 2:, 0],
np.ones((size // 2, 1)),
color='b')
plt.legend(['Label 1', 'Label 0'], loc='upper right')
elif dim == 2:
plt.scatter(inputdata[:size // 2, 0],
inputdata[:size // 2, 1],
color='r')
plt.scatter(inputdata[size // 2:, 0],
inputdata[size // 2:, 1],
color='b')
plt.legend(['Label 1', 'Label 0'], loc='upper right')
network = net(inputdata, outputdata, size, ss, numofiter, dim,
hiddenlayerlist, modeltype, algorithm, output_unit, [],
0)
network.backpropagation()
output = network.forwardewithcomputedW(inputdata)
#plot network computed result
output = np.append(inputdata, output, axis=1)
print('Network computed output: ')
plt.figure(4)
if dim == 1:
output1 = output[output[:, -1] == 1]
output2 = output[output[:, -1] == 0]
plt.scatter(output1[:, 0],
np.ones((np.shape(output1)[0], 1)),
color='r')
plt.scatter(output2[:, 0],
np.ones((np.shape(output2)[0], 1)),
color='b')
plt.legend(['Label 1', 'Label 0'], loc='upper right')
if dim == 2:
output1 = output[output[:, -1] == 1]
output2 = output[output[:, -1] == 0]
plt.scatter(output1[:, 0], output1[:, 1], color='r')
plt.scatter(output2[:, 0], output2[:, 1], color='b')
plt.legend(['Label 1', 'Label 0'], loc='upper right')
plt.show()
elif modeltype == 'r':
#generate the input and output for regression
inputdata, outputdata = generatedataForRegression(size, dim)
network = net(inputdata, outputdata, size, ss, numofiter, dim,
hiddenlayerlist, modeltype, output_unit, [], 0)
network.backpropagation()
if dim == 2:
fig = plt.figure(figsize=(10, 10))
ax = plt.axes(projection='3d')
X = np.arange(-4, 4, 0.1)
Y = np.arange(-4, 4, 0.1)
X, Y = np.meshgrid(X, Y)
a = X.flatten()
b = Y.flatten()
testx = np.append(np.reshape(a, (len(a), 1)),
np.reshape(b, (len(b), 1)),
axis=1)
outputy = np.reshape(network.forwardewithcomputedW(testx),
np.shape(X))
ax.plot_surface(X,
Y,
outputy,
rstride=1,
cstride=1,
cmap=cm.coolwarm,
linewidth=0,
antialiased=False)
elif modeltype == 'mnist':
train_images, train_labels, test_images, test_labels = get_mnist()
# size = train_images.shape[0]
size = 60000
numofiter = 1000
dim = 28**2
hiddenlayerlist = [[1000]] # 2500, 2000, 1500, 1000, 500
output_unit = 10
ss = 5e-2
print('Algorithm: ' + algorithm + '\nModel type: ' + modeltype +
'\nIterations: ' + str(numofiter) + '\nLearning rate: ' +
str(ss))
# get_one_hot(train_labels[: size, :], 10)
# train_labels[: size, :].flatten()
network = net(train_images[:size, :],
get_one_hot(train_labels[:size, :], 10), size, ss,
numofiter, dim, hiddenlayerlist, modeltype, algorithm,
output_unit, [], 5000)
network.backpropagation()
# load the saved model
filename = 'wb_' + modeltype + '_' + algorithm + '_' + str(
numofiter) + '.npz'
wb_ini = np.load(filename)['arr_0'].tolist()
network = net(train_images[:size, :],
get_one_hot(train_labels[:size, :], 10), size, ss,
numofiter, dim, hiddenlayerlist, modeltype, algorithm,
output_unit, wb_ini, 0)
# test the accuracy
tst_size = 10000
tst_imgs = train_images[:tst_size]
tst_lbls = train_labels[:tst_size].flatten()
tst_out_raw = network.forwardewithcomputedW(tst_imgs)
tst_out_cls = np.argmax(tst_out_raw, axis=1)
accuracy = sum(tst_out_cls == tst_lbls) / tst_size
print('test accuracy: ' + str(accuracy))
# set_trace()
elif modeltype == 'bc':
data = np.genfromtxt("breastCancerData.csv", delimiter=",")
label = np.genfromtxt("breastCancerLabels.csv", delimiter=",")
MinMaxscaler = sklearn.preprocessing.MinMaxScaler()
data = np.float32(MinMaxscaler.fit_transform(data))
#Split Train and Test Data
trainD, testD, trainT, testT = train_test_split(data,
label,
random_state=6)
size = np.shape(trainD)[0]
numofiter = 1000
dim = 9
hiddenlayerlist = [[80, 100, 50]]
output_unit = 1
network = net(trainD, np.reshape(trainT, (len(trainT), 1)), size, ss,
numofiter, dim, hiddenlayerlist, modeltype, algorithm,
output_unit, [], 0)
network.backpropagation()
output = network.forwardewithcomputedW(testD)
accuracy = sum(
output == np.reshape(testT, (len(testT), 1))) / len(testT)
print('test accuracy: ' + str(accuracy[0]))
print(info_list)
return info_list
def process_image(img):
global id
# resization
img = misc.imresize(img, [aim_size, aim_size, 3])
print('{} finished.'.format(id))
id += 1
return img
def dump_file(obj, dump_filename):
with open(dump_filename, 'wb') as fout:
pickle.dump(obj, fout)
if __name__ == '__main__':
info_list = get_info_list()
result_list = []
for i, d in enumerate(info_list):
if os.path.exists(d[3]):
print(d[1], d[2])
if int(d[1]) < year:
continue
# one_hot = utils.get_one_hot([d[2]])
result_list.append(
[utils.get_one_hot([d[2]]),
process_image(misc.imread(d[3]))])
dump_file(result_list, dump_filename)
def al(dataset_name='seed4', FOIT_type='cross-all', rounds=10, batch_size=50):
data, label = utils.load_source_data(dataset_name=dataset_name,
FOIT_type=FOIT_type)
_, number_label, _ = utils.get_number_of_label_n_trial(dataset_name)
# data, label = utils.load_session_data_label(dataset_name, 0) # as unlabelled data
cd_count = 16 if dataset_name == 'seed4' else 9 if dataset_name == 'seed3' else print(
'Wrong dataset_name')
iteration_number = 3 if FOIT_type == 'cross-subject' else 15
accs = [([]) for i in range(iteration_number)]
times = [([]) for i in range(iteration_number)]
for ite in range(iteration_number):
session_id = -1
sub_id = -1
if FOIT_type == 'cross-subject':
session_id = ite
sub_id = 14
elif FOIT_type == 'cross-session':
session_id = 2
sub_id = ite
elif FOIT_type == 'cross-all':
session_id = 1
sub_id = ite
else:
print('Wrong FOIT type!')
# print("Ite: ", ite)
cd_data, cd_label, ud_data, ud_label = utils.pick_one_data(
dataset_name,
session_id=session_id,
cd_count=cd_count,
sub_id=sub_id)
cd_data, cd_label = shuffle(cd_data, cd_label, random_state=0)
ud_data, ud_label = shuffle(ud_data, ud_label, random_state=0)
cd_data_min, cd_data_max = np.min(cd_data), np.max(cd_data)
cd_data = utils.normalization(cd_data) # labelled data
ud_data = utils.normalization(ud_data) # test data
if FOIT_type == 'cross-all':
data_ite, label_ite = data.copy(), label.copy()
for i in range(len(data)):
data_ite[i], label_ite[i] = shuffle(data_ite[i],
label_ite[i],
random_state=0)
# data_ite, label_ite = shuffle(data, label, random_state=0)
for i in range(len(data)):
data_ite[i] = utils.norm_with_range(data_ite[i], cd_data_min,
cd_data_max)
# data_ite = utils.normalization(data_ite)
elif FOIT_type == 'cross-session':
data_ite, label_ite = data[ite], label[ite]
for i in range(len(data_ite)):
data_ite[i], label_ite[i] = shuffle(data_ite[i],
label_ite[i],
random_state=0)
# data_ite[i] = utils.normalization(data_ite[i])
data_ite[i] = utils.norm_with_range(data_ite[i], cd_data_min,
cd_data_max)
# data_ite = utils.normalization(data_ite)
else:
data_ite, label_ite = data[ite], label[ite]
for i in range(len(data_ite)):
data_ite[i], label_ite[i] = shuffle(data_ite[i],
label_ite[i],
random_state=0)
# data_ite, label_ite = shuffle(data_ite, label_ite, random_state=0)
for i in range(len(data_ite)):
# data_ite[i] = utils.normalization(data_ite[i])
data_ite[i] = utils.norm_with_range(data_ite[i], cd_data_min,
cd_data_max)
# data_ite, label_ite = data.copy(), label.copy()
# for i in range(len(data)):
# data_ite[i], label_ite[i] = shuffle(data_ite[i], label_ite[i], random_state=0)
# for i in range(len(data)):
# data_ite[i] = utils.norm_with_range(data_ite[i], cd_data_min, cd_data_max)
# baseline
clf = svm.LinearSVC(max_iter=30000)
clf = CalibratedClassifierCV(clf, cv=5)
since = time.time()
clf.fit(cd_data, cd_label.squeeze())
time_baseline = time.time() - since
scoreA = utils.test(clf, ud_data, ud_label.squeeze())
accs[ite].append(scoreA)
times[ite].append(time_baseline)
# select the data from the reservoir iteratively
s_data_all, s_label_all = utils.stack_list(data_ite, label_ite)
L_S_data = None
L_S_label = None
for i in range(rounds):
# print("Rounds: ", i)
# print(type(s_data_all))
# print(s_data_all.shape)
s_data_all_predict_proba = clf.predict_proba(s_data_all)
s_label_all_proba = utils.get_one_hot(s_label_all.squeeze(),
number_label)
confidence = np.zeros((s_label_all_proba.shape[0], 1))
for i in range(s_label_all_proba.shape[0]):
confidence[i] = s_label_all_proba[i].dot(
s_data_all_predict_proba[i].T)
# confidence[i] = log_loss(s_label_all_proba[i], s_data_all_predict_proba[i])
indices = np.argsort(confidence,
axis=0) # take the minimum topK indices
topK_indices = indices[:batch_size]
S_data = None
S_label = None
for i in topK_indices:
one_data = s_data_all[i]
one_label = s_label_all[i]
if S_data is not None:
S_data = np.vstack((S_data, one_data))
S_label = np.vstack((S_label, one_label))
else:
S_data = one_data
S_label = one_label
for i in range(len(s_data_all) - 1, -1, -1):
if i in topK_indices:
s_data_all = np.delete(s_data_all, i, axis=0)
s_label_all = np.delete(s_label_all, i, axis=0)
if L_S_data is None:
L_S_data = cd_data.copy()
L_S_label = cd_label.copy()
else:
pass
L_S_data = np.vstack((L_S_data, S_data))
L_S_label = np.vstack((L_S_label, S_label))
L_S_data, L_S_label = shuffle(L_S_data, L_S_label, random_state=0)
clf.fit(L_S_data, L_S_label.squeeze())
time_updated_time = time.time() - since
times[ite].append(time_updated_time)
scoreTMP = utils.test(clf, ud_data, ud_label.squeeze())
accs[ite].append(scoreTMP)
ResultTime = []
ResultAcc = []
ResultStd = []
for i in range(rounds + 1):
tmpTime = []
tmpAcc = []
for j in range(iteration_number):
tmpTime.append(times[j][i])
tmpAcc.append(accs[j][i])
ResultTime.append(np.mean(tmpTime))
ResultAcc.append(np.mean(tmpAcc))
ResultStd.append(np.std(tmpAcc))
print("Time: ", ResultTime)
print("Accs: ", ResultAcc)
print("Stds: ", ResultStd)
temp_clf = joblib.load(path)
clf_sources.append(temp_clf)
score = utils.test(temp_clf, utils.normalization(ud_data),
ud_label.squeeze())
accs.append(score)
# print('Accs of classifiers: {}'.format(accs))
accs = utils.normalization(accs)
print('Accs of classifiers, normalized: {}'.format(accs))
'''
c)
'''
rho = 0.5
s_data_all, s_label_all = utils.stack_list(sub_data_ses, sub_label_ses)
s_data_all_predict_proba = clf.predict_proba(
utils.normalization(s_data_all))
s_label_all_proba = utils.get_one_hot(s_label_all.squeeze(), number_label)
confidence = np.zeros((s_label_all_proba.shape[0], 1))
for i in range(s_label_all_proba.shape[0]):
confidence[i] = s_label_all_proba[i].dot(s_data_all_predict_proba[i].T)
subs_data_0, subs_data_1, subs_data_2, subs_data_3 = [], [], [], []
conf_0, conf_1, conf_2, conf_3 = [], [], [], []
subs_label_0, subs_label_1, subs_label_2, subs_label_3 = [], [], [], []
for i in range(len(s_data_all)):
temp_label = s_label_all[i][0]
eval('subs_data_' + str(temp_label)).append(s_data_all[i])
eval('conf_' + str(temp_label)).append(confidence[i])
eval('subs_label_' + str(temp_label)).append(s_label_all[i])
indices = []
for i in range(4):
indices.append(np.argsort(eval('conf_' + str(i)), axis=0)[::-1])
# indices.append(np.argsort(eval('conf_'+str(i)), axis=0))
import numpy as np
from simple_rnn_model import SimpleRNNModel, params, SimpleRNNModel1
from utils import get_one_hot
training_X = np.load("./AG/X_train.npy")
training_y = np.load("./AG/y_train.npy")
training_num = len(training_X)
# shuffle_ids = np.array(np.arange(training_num))
# np.random.shuffle(shuffle_ids)
# training_X = training_X[shuffle_ids]
# training_y = training_y[shuffle_ids]
test_X = np.load("./AG/X_test.npy")
test_y = np.load("./AG/y_test.npy")
nb_classes = 4
training_Y = get_one_hot(training_y, nb_classes)
test_Y = get_one_hot(test_y, nb_classes)
model = SimpleRNNModel1(params, 60, nb_classes)
model.model.fit(x=training_X,
y=training_Y,
batch_size=64,
epochs=30,
callbacks=[model.early_stopping],
verbose=2,
validation_data=(test_X[:500], test_Y[:500]),
shuffle=True)
model.model.save_weights(filepath="./models/rnn")
# batch_size = 64
# params["batch_size"] = batch_size
# model = SimpleRNNModel(params, 60, 4)
def runModel(datagen, model, optimizer, class_wts, process, batch_size,
n_batches, loss_wts):
'''
process : 'trn', 'val' or 'tst'
'''
running_loss = 0
pred_list = []
label_list = []
soft_pred_list = []
all_file_list = []
with trange(n_batches, desc=process, ncols=100) as t:
for m in range(n_batches):
data, labels, filenames = datagen.__next__()
labels_one_hot = utils.get_one_hot(labels).cuda()
if process == 'trn':
optimizer.zero_grad()
model.train()
pred, aux_pred = model.forward(data)
pred = F.softmax(pred, 1)
aux_pred = F.softmax(aux_pred, 1)
loss = 0
for i in range(2):
loss += loss_wts[0] * utils.weightedBCE(class_wts[i],
pred[:, i],
(labels_one_hot
[:, i]))\
+ loss_wts[1] * utils.weightedBCE(class_wts[i],
aux_pred[:, i],
(labels_one_hot
[:, i]))
loss.backward()
if torch.isnan(loss):
pdb.set_trace()
optimizer.step()
elif process == 'val' or process == 'tst':
model.eval()
with torch.no_grad():
pred = F.softmax(model.forward(data), 1)
loss = utils.weightedBCE(class_wts[0], pred[:, 0],
labels_one_hot[:, 0])\
+ utils.weightedBCE(class_wts[1], pred[:, 1],
labels_one_hot[:, 1])
running_loss += loss
hard_pred = torch.argmax(pred, 1)
pred_list.append(hard_pred.cpu())
soft_pred_list.append(pred.detach().cpu())
label_list.append(labels.cpu())
all_file_list += filenames
t.set_postfix(loss=running_loss.item() /
(float(m + 1) * batch_size))
t.update()
finalLoss = running_loss / (float(m + 1) * batch_size)
# if process != 'trn':
# pred_list, soft_pred_list, label_list = utils.test_time_aug(
# all_file_list,
# soft_pred_list,
# label_list, 3)
acc = utils.globalAcc(pred_list, label_list)
if not isinstance(pred_list, torch.Tensor):
f1 = sklearn.metrics.f1_score(torch.cat(label_list),
torch.cat(pred_list),
labels=None)
else:
f1 = sklearn.metrics.f1_score(label_list, pred_list, labels=None)
auroc, auprc, fpr_tpr_arr, precision_recall_arr = utils.AUC(
soft_pred_list, label_list)
metrics = Metrics(finalLoss, acc, f1, auroc, auprc, fpr_tpr_arr,
precision_recall_arr)
utils.save_preds(soft_pred_list, pred_list, label_list, all_file_list,
args.savename, process)
return metrics
def FOIT(dataset_name='seed4',
rho=1,
clf_name='lr',
threshold=0.6,
with_balance=True,
FOIT_type='cross-all'):
# time and accs
c0_acc = []
c0u_acc = []
foit_acc = []
time_c0 = []
time_c0u = []
time_foit = []
cd_count = 16 if dataset_name == 'seed4' else 9 if dataset_name == 'seed3' else print(
'Wrong dataset_name')
iteration_number = 3 if FOIT_type == 'cross-subject' else 15
number_trial, number_label, labels = utils.get_number_of_label_n_trial(
dataset_name)
data, label = utils.load_source_data(dataset_name=dataset_name,
FOIT_type=FOIT_type)
for ite in range(iteration_number):
# The data of 15th sub is taken as the cd and ud data in cross-subject for each session (iteration_number=3)
# The data of 3rd session is taken as the cd and ud data in cross-session for each subject (iteration_number=14)
# For each iteration, the data of one sub from session 2 is taken as the cd and ud data in cross-all situation
# print("Iteration: {}".format(ite))
'''
Parameters
'''
session_id = -1
sub_id = -1
accs_number = -1
if FOIT_type == 'cross-subject':
session_id = ite
sub_id = 14
accs_number = 14
elif FOIT_type == 'cross-session':
session_id = 2
sub_id = ite
accs_number = 2
elif FOIT_type == 'cross-all':
session_id = 1
sub_id = ite
accs_number = 15
else:
print('Wrong FOIT type!')
'''
Data
'''
# data
cd_data, cd_label, ud_data, ud_label = utils.pick_one_data(
dataset_name,
session_id=session_id,
cd_count=cd_count,
sub_id=sub_id)
cd_data, cd_label = shuffle(cd_data, cd_label, random_state=0)
ud_data, ud_label = shuffle(ud_data, ud_label, random_state=0)
cd_data_min, cd_data_max = np.min(cd_data), np.max(cd_data)
cd_data = utils.normalization(cd_data)
ud_data = utils.normalization(ud_data)
if FOIT_type == 'cross-all':
data_ite, label_ite = data.copy(), label.copy()
for i in range(len(data)):
data_ite[i], label_ite[i] = shuffle(data_ite[i],
label_ite[i],
random_state=0)
# data_ite, label_ite = shuffle(data, label, random_state=0)
for i in range(len(data)):
data_ite[i] = utils.norm_with_range(data_ite[i], cd_data_min,
cd_data_max)
# data_ite = utils.normalization(data_ite)
elif FOIT_type == 'cross-session':
data_ite, label_ite = data[ite], label[ite]
for i in range(len(data_ite)):
data_ite[i], label_ite[i] = shuffle(data_ite[i],
label_ite[i],
random_state=0)
# data_ite[i] = utils.normalization(data_ite[i])
data_ite[i] = utils.norm_with_range(data_ite[i], cd_data_min,
cd_data_max)
# data_ite = utils.normalization(data_ite)
else:
data_ite, label_ite = data[ite], label[ite]
for i in range(len(data_ite)):
data_ite[i], label_ite[i] = shuffle(data_ite[i],
label_ite[i],
random_state=0)
# data_ite, label_ite = shuffle(data_ite, label_ite, random_state=0)
for i in range(len(data_ite)):
# data_ite[i] = utils.normalization(data_ite[i])
data_ite[i] = utils.norm_with_range(data_ite[i], cd_data_min,
cd_data_max)
# data_ite = utils.normalization(data_ite)
# print(len(data_ite), len(label_ite[0]), len(label_ite[0][0]))
'''
A)
'''
if clf_name == 'svm':
clf = svm.LinearSVC(max_iter=30000)
# clf = svm.SVC(probability=True, max_iter=10000)
elif clf_name == 'lr':
clf = LogisticRegression(max_iter=30000)
else:
print('Unexcepted clf name, using LR as the baseline now.')
clf = svm.LinearSVC(max_iter=30000)
clf = CalibratedClassifierCV(clf, cv=5)
since = time.time()
clf.fit(cd_data, cd_label.squeeze())
time_baseline = time.time() - since
# print('Baseline training complete in {:.4f}'.format(time_baseline))
scoreA = utils.test(clf, ud_data, ud_label.squeeze())
# print('Baseline score: {}'.format(scoreA))
time_c0.append(time_baseline)
c0_acc.append(scoreA)
'''
B)
'''
accs = []
clf_sources = []
for i in range(accs_number):
if FOIT_type == 'cross-subject':
path = 'models/' + dataset_name + '/csu/sesn' + str(
ite) + '/lr' + str(i) + '.m'
elif FOIT_type == 'cross-session':
path = 'models/' + dataset_name + '/csn/sub' + str(
ite) + '/lr' + str(i) + '.m'
else:
path = 'models/' + dataset_name + '/csun/lr' + str(i) + '.m'
temp_clf = joblib.load(path)
clf_sources.append(temp_clf)
score = utils.test(temp_clf, ud_data, ud_label.squeeze())
accs.append(score)
if FOIT_type == 'cross-session':
pass
else:
accs = utils.normalization(accs)
# print('Accs of classifiers, normalized: {}'.format(accs))
'''
C)
'''
s_data_all, s_label_all = utils.stack_list(data_ite, label_ite)
s_data_all_predict_proba = clf.predict_proba(s_data_all)
s_label_all_proba = utils.get_one_hot(s_label_all.squeeze(),
number_label)
confidence = np.zeros((s_label_all_proba.shape[0], 1))
for i in range(s_label_all_proba.shape[0]):
confidence[i] = s_label_all_proba[i].dot(
s_data_all_predict_proba[i].T)
if with_balance:
## divide into 4 categories
data_ite_divided = []
label_ite_divided = []
conf_divided = []
for i in range(number_label):
data_ite_divided.append([])
label_ite_divided.append([])
conf_divided.append([])
for i in range(len(s_data_all)):
temp_label = s_label_all[i][0]
data_ite_divided[temp_label].append(s_data_all[i])
label_ite_divided[temp_label].append(s_label_all[i])
conf_divided[temp_label].append(confidence[i])
indices = []
for i in range(number_label):
indices.append(np.argsort(conf_divided[i], axis=0)[::-1])
topK_indices = [
indices[i][:int(rho * len(cd_label) / 4)]
for i in range(len(indices))
]
S_data = None
S_label = None
for i in range(len(topK_indices)):
for j in topK_indices[i]:
# temp_conf = conf_divided[i][j[0]]
one_data = data_ite_divided[i][j[0]]
one_label = label_ite_divided[i][j[0]]
if S_data is not None:
S_data = np.vstack((S_data, one_data))
S_label = np.vstack((S_label, one_label))
else:
S_data = one_data
S_label = one_label
elif with_balance is False:
indices = np.argsort(confidence, axis=0)[::-1]
topK_indices = indices[:int(rho * len(cd_label))]
S_data = None
S_label = None
for i in topK_indices:
one_data = s_data_all[i]
one_label = s_label_all[i]
if S_data is not None:
S_data = np.vstack((S_data, one_data))
S_label = np.vstack((S_label, one_label))
else:
S_data = one_data
S_label = one_label
else:
print('Unexcepted value of with balance!')
raise EnvironmentError
'''
D)
'''
L_S_data = cd_data.copy()
L_S_label = cd_label.copy()
L_S_data = np.vstack((L_S_data, S_data))
L_S_label = np.vstack((L_S_label, S_label))
L_S_data, L_S_label = shuffle(L_S_data, L_S_label, random_state=0)
# L_S_data = utils.normalization(L_S_data) # to decide
clf.fit(L_S_data, L_S_label.squeeze())
time_updated_baseline = time.time() - since
# print('Updated baseline training complete in {:.4f}s'.format(time_updated_baseline))
time_c0u.append(time_updated_baseline)
scoreD = utils.test(clf, ud_data, ud_label.squeeze())
# print('Updated model score: {}'.format(scoreD))
c0u_acc.append(scoreD)
'''
E)
'''
weight = (len(accs) + 1) / 2
proba_result_all = clf.predict_proba(ud_data) * weight
if FOIT_type == 'cross-session':
weight_for_clfs = utils.decide_which_clf_to_use(scoreD, accs)
for j in range(len(weight_for_clfs)):
proba_result_all += clf_sources[j].predict_proba(
utils.normalization(ud_data)) * weight_for_clfs[j]
else:
for i in range(len(clf_sources)):
if accs[i] > threshold:
proba_result_all += clf_sources[i].predict_proba(
ud_data) * accs[i]
corrects = np.sum(
np.argmax(proba_result_all, axis=1) == ud_label.squeeze())
time_ensemble = time.time() - since
time_foit.append(time_ensemble)
scoreE = corrects / len(ud_label)
# print('Ensembled model score: {}'.format(scoreE))
foit_acc.append(scoreE)
# print(c0_acc)
# print(c0u_acc)
# print(foit_acc)
print('Mean acc and std of A: {} {}'.format(np.mean(c0_acc),
np.std(c0_acc)))
print('Mean acc and std of D: {} {}'.format(np.mean(c0u_acc),
np.std(c0u_acc)))
print('Mean acc and std of E: {} {}'.format(np.mean(foit_acc),
np.std(foit_acc)))
print("Time cost for training baseline: ", np.mean(time_c0))
print("Time cost for training updated baseline: ", np.mean(time_c0u))
print("Time cost for training FOIT: ", np.mean(time_foit))
return avatar_list
def process_image(img):
global id
# resization
img = misc.imresize(img, [aim_size, aim_size, 3])
print('{} finished.'.format(id))
id += 1
return img
def dump_file(obj, dump_filename):
with open(dump_filename, 'wb') as fout:
pickle.dump(obj, fout)
if __name__ == '__main__':
avatar_list = get_avatar_with_tag(avatar_tag_path)
result_list = []
for i, each in enumerate(avatar_list):
if os.path.exists(each[3]):
if int(each[1]) < 2005:
continue
# tag's one-hot, image-bytes
result_list.append([
utils.get_one_hot(each[2]),
process_image(misc.imread(each[3]))
])
dump_file(result_list, dump_filename)
def train(self, graphs_train, graphs_test=None, graphs_valid=None) -> None:
"""
Trains model with training set in multiple epochs.
Args:
graphs_train: A list of graphs as train set.
graphs_test: A list of graphs as test set.
graphs_valid: A list of graphs as validation set.
"""
# Enrich graphs with adj list
for graph in graphs_train:
graph[utils.AE.ADJ_LIST] = utils.graph_to_adjacency_lists(
graph[utils.T.EDGES], self.config['tie_fwd_bkwd'] == 1)[0]
# Extract graph sizes
graph_sizes = []
for graph in graphs_train:
graph_sizes.append(len(graph[utils.T.NODES]))
best_epoch_loss = sys.maxsize
best_epoch_accuracy = 0
best_epoch_count = 0
if graphs_valid:
# Extract valid labels
y_valid = []
for graph in graphs_valid:
y_valid.append(graph[utils.L.LABEL_0])
y_valid = np.array(y_valid)
if graphs_test:
# Extract test labels
y_test = []
for graph in graphs_test:
y_test.append(graph[utils.L.LABEL_0])
y_test = np.array(y_test)
# Run epochs
for epoch in range(0, self.config['num_epochs']):
# Training
# ############################################
epoch_start_time = time.time()
# Partition into batches
batch_size = self.config['batch_size']
lst = list(zip(graphs_train, graph_sizes))
np.random.shuffle(lst)
batches = [
lst[i * batch_size:(i + 1) * batch_size]
for i in range((len(lst) + batch_size - 1) // batch_size)
]
# Run batches
training_losses = []
training_good_predicted_all = []
epoch_instances_per_secs = []
for batch in batches:
start_time = time.time()
batch_graphs, batch_graph_sizes = zip(*batch)
batch_graphs = list(batch_graphs)
batch_graph_sizes = list(batch_graph_sizes)
# Extract train labels
y_train = []
for graph in batch_graphs:
y_train.append(graph[utils.L.LABEL_0])
y_train = np.array(y_train)
# Build feed dict
# 1. Graph info
feed_dict = self._graphs_to_batch_feed_dict(
batch_graphs, batch_graph_sizes, True)
# Run batch
result = self._run_batch(feed_dict)
end_time = time.time()
# Log
instances_per_sec = len(batch_graphs) / (end_time - start_time)
epoch_instances_per_secs.append(instances_per_sec)
training_losses.append(result[0])
# Evaluate predictions on train set
predictions = result[1]
training_good_predicted = list(
np.argmax(predictions, axis=1) == y_train)
training_good_predicted_all += training_good_predicted
training_accuracy = np.sum(training_good_predicted_all) / len(
training_good_predicted_all)
training_loss = np.sum(training_losses)
epoch_instances_per_sec = np.mean(epoch_instances_per_secs)
epoch_end_time = time.time()
epoch_time = epoch_end_time - epoch_start_time
# Logging
# summary = tf.Summary()
# summary.value.add(tag='train_accuracy', simple_value=training_accuracy)
# summary.value.add(tag='train_loss', simple_value=training_loss)
# Testing
# ############################################
if graphs_valid and graphs_test:
# Make predictions on valid set
predictions = self.predict(graphs_valid)
valid_loss = np.sum(predictions - utils.get_one_hot(
y_valid, self.config['prediction_cell']['output_dim']))
valid_accuracy = np.sum(
np.argmax(predictions, axis=1) == y_valid) / len(
predictions)
# Make predictions on test set
predictions = self.predict(graphs_test)
test_loss = np.sum(predictions - utils.get_one_hot(
y_test, self.config['prediction_cell']['output_dim']))
test_accuracy = np.sum(
np.argmax(predictions, axis=1) == y_test) / len(
predictions)
# Logging
# summary.value.add(tag='valid_accuracy', simple_value=valid_accuracy)
# summary.value.add(tag='test_accuracy', simple_value=test_accuracy)
# print('epoch: %i, instances/sec: %.2f, epoch_time: %.2fs, train_loss: %.8f, train_accuracy: %.4f, valid_accuracy: %.4f, test_accuracy: %.4f' % (epoch, epoch_instances_per_sec, epoch_time, training_loss, training_accuracy, valid_accuracy, test_accuracy))
if valid_accuracy > best_epoch_accuracy:
best_epoch_accuracy = valid_accuracy
best_epoch_count += 1
if 'save_best_model_interval' in self.config and best_epoch_count >= self.config[
'save_best_model_interval']:
self.state.backup_best_weights()
best_epoch_count = 0
elif graphs_test:
# Make predictions on test set
predictions = self.predict(graphs_test)
test_loss = np.sum(predictions - utils.get_one_hot(
y_test, self.config['prediction_cell']['output_dim']))
test_accuracy = np.sum(
np.argmax(predictions, axis=1) == y_test) / len(
predictions)
# Logging
# summary.value.add(tag='test_accuracy', simple_value=test_accuracy)
# summary.value.add(tag='test_loss', simple_value=test_loss)
# print('epoch: %i, instances/sec: %.2f, epoch_time: %.2fs, train_loss: %.8f, train_accuracy: %.4f, test_accuracy: %.4f' % (epoch, epoch_instances_per_sec, epoch_time, training_loss, training_accuracy, test_accuracy))
if training_loss < best_epoch_loss:
best_epoch_loss = training_loss
best_epoch_count += 1
if 'save_best_model_interval' in self.config and best_epoch_count >= self.config[
'save_best_model_interval']:
self.state.backup_best_weights()
best_epoch_count = 0
else:
# Logging
# print('epoch: %i, instances/sec: %.2f, epoch_time: %.2fs, loss: %.8f' % (epoch, epoch_instances_per_sec, epoch_time, training_loss))
if training_loss < best_epoch_loss:
best_epoch_loss = training_loss
best_epoch_count += 1
if 'save_best_model_interval' in self.config and best_epoch_count >= self.config[
'save_best_model_interval']:
self.state.backup_best_weights()
best_epoch_count = 0
# Logging
#self.train_writer.add_summary(summary, epoch)
self.state.restore_best_weights()
def _graphs_to_batch_feed_dict(self, graphs: list, graph_sizes: list,
is_training: bool) -> dict:
"""Creates feed dict that is the format the tf model is trained with."""
num_edge_types = self.config['num_edge_types']
batch_data = {
# Graph model
'adjacency_lists':
[[] for _ in range(self.config['num_edge_types'])],
'embeddings_to_graph_mappings': [],
'labels': [],
'embeddings_in': [],
'node_values': [],
}
if self.with_aux_in:
batch_data['aux_in'] = []
for graph_idx, graph in enumerate(graphs):
num_nodes = graph_sizes[graph_idx]
if self.config['use_edge_bias'] == 1:
batch_data['num_incoming_edges_dicts_per_type'] = []
# Aux in
if self.with_aux_in:
if utils.I.AUX_IN_0 in graph:
batch_data['aux_in'].append(graph[utils.I.AUX_IN_0])
# Labels
if utils.L.LABEL_0 in graph:
batch_data['labels'].append(graph[utils.L.LABEL_0])
# Graph model: Adj list
adj_lists = graph[utils.AE.ADJ_LIST]
for idx, adj_list in adj_lists.items():
if idx >= self.config['num_edge_types']:
continue
batch_data['adjacency_lists'][idx].append(adj_list)
if self.config['use_edge_bias'] == 1:
# Graph model: Incoming edge numbers
num_incoming_edges_dicts_per_type = action[
utils.AE.
NUMS_INCOMING_EDGES_BY_TYPE] if action else utils.graph_to_adjacency_lists(
[], self.config['tie_fwd_bkwd'],
self.config['edge_type_filter'])[0]
num_incoming_edges_per_type = np.zeros(
(num_nodes, num_edge_types))
for (e_type, num_incoming_edges_per_type_dict
) in num_incoming_edges_dicts_per_type.items():
for (node_id, edge_count
) in num_incoming_edges_per_type_dict.items():
num_incoming_edges_per_type[node_id,
e_type] = edge_count
batch_data['num_incoming_edges_dicts_per_type'].append(
num_incoming_edges_per_type)
graph_mappings_all = np.full(num_nodes, graph_idx)
batch_data['embeddings_to_graph_mappings'].append(
graph_mappings_all)
node_types = utils.get_one_hot(
np.array(graph[utils.T.NODES]),
self.config['hidden_size_orig']).astype(float)
batch_data['embeddings_in'].append(node_types)
if self.config['use_node_values'] == 1:
node_values = np.array(
graph[utils.T.NODE_VALUES]).astype(float).reshape(-1, 1)
node_values = node_values / 2147483647
batch_data['node_values'].append(node_values)
# Build feed dict
feed_dict = {}
# Graph model: Adj list
for idx, adj_list in enumerate(
self.ggnn_layers[0].placeholders['adjacency_lists']):
feed_dict[adj_list] = np.array(batch_data['adjacency_lists'][idx])
if len(feed_dict[adj_list]) == 0:
feed_dict[adj_list] = np.zeros((0, 2), dtype=np.int32)
else:
feed_dict[adj_list] = feed_dict[adj_list][0]
if self.config['use_edge_bias'] == 1:
# Graph model: Incoming edge numbers
feed_dict[self.ggnn_layers[0].placeholders['num_incoming_edges_per_type']] \
= np.concatenate(batch_data['num_incoming_edges_dicts_per_type'], axis=0)
# Is training
feed_dict[self.cells[0].placeholders['is_training']] = is_training
# Aux in
if self.with_aux_in:
feed_dict[
self.cells[0].placeholders['aux_in']] = batch_data['aux_in']
# Labels
if 'labels' in self.cells[0].placeholders:
feed_dict[
self.cells[0].placeholders['labels']] = utils.get_one_hot(
np.array(batch_data['labels']),
self.config['prediction_cell']['output_dim'])
# Embeddings
feed_dict[self.placeholders['embeddings_in']] = np.concatenate(
batch_data['embeddings_in'], axis=0)
feed_dict[self.cells[0].placeholders['embeddings_to_graph_mappings']] \
= np.concatenate(batch_data['embeddings_to_graph_mappings'], axis=0)
# Node values
if self.config['use_node_values'] == 1:
feed_dict[self.placeholders['node_values']] = np.concatenate(
batch_data['node_values'], axis=0)
return feed_dict
