def val(path, model, test_loader, device, criterion, epoch, batch_size):
model.eval()
sum_running_loss = 0.0
with torch.no_grad():
for batch_idx, data in enumerate(tqdm(test_loader)):
task = data['task'].to(device).float()
task_label = data['task_label'].to(device).float()
# All black
# init = data['init']
# init[:] = 0
# init = init.to(device).float()
# Normal
init = data['init'].to(device).float()
# print("init shape", init.shape)
label = data['label'].to(device).float()
# model prediction
prediction = model(subx=task_label, mainx=init)
# loss
loss_mse = criterion(prediction, label.data)
# accumulate loss
sum_running_loss += loss_mse.item() * init.size(0)
# visualize the sum testing result
visualize_sum_testing_result(path, init, prediction, task_label,
label.data, batch_idx, epoch,
batch_size)
if batch_idx == 0:
prediction_output = prediction.cpu().detach().numpy()
label_output = label.cpu().detach().numpy()
init_output = init.cpu().detach().numpy()
else:
prediction_output = np.append(
prediction.cpu().detach().numpy(),
prediction_output,
axis=0)
label_output = np.append(label.cpu().detach().numpy(),
label_output,
axis=0)
init_output = np.append(init.cpu().detach().numpy(),
init_output,
axis=0)
sum_running_loss = sum_running_loss / len(test_loader.dataset)
print('\nTesting phase: epoch: {} Loss: {:.4f}\n'.format(
epoch, sum_running_loss))
auc_path = os.path.join(path, "epoch_" + str(epoch))
auc(['flow'], [2, 4, 10, 100], [[label_output, prediction_output]],
auc_path)
return sum_running_loss, prediction_output, label_output, init_output
def test_step(images, labels, tp, tn, fn, fp):
output = model(images, training=False)
t_loss = loss_object(labels, output)
test_loss(t_loss)
test_accuracy(labels, output)
aucvalue = auc(labels,
output,
weights=None,
num_thresholds=200,
name=None,
summation_method='trapezoidal')
test_auc(aucvalue)
# print("output: ", output)
# print("labels: ", labels)
y_pred = tf.argmax(output, 1)
y_true = tf.argmax(labels, 1)
# print("y_pred： ", y_pred)
# print("y_true： ", y_true)
pred = tf.cast(y_pred, tf.int32)
true = tf.cast(y_true, tf.int32)
# print("pred: ", pred)
# print("true: ", true)
for i in range(len(pred)):
if pred[i] == true[i] and true[i] == 1:
tp += 1
elif pred[i] == true[i] and true[i] == 0:
tn += 1
elif pred[i] == 1 and true[i] == 0:
fp += 1
else:
fn += 1
return tp, tn, fp, fn, output
def cal_auc(ans, i, preds):
preds['auc'] = 0.0
for index, pred in preds.iterrows():
truth = ans.loc[index]
a = auc(pred[0:28], truth)
preds.at[index, 'auc'] = round(a, 4)
cols = preds.columns.tolist()
cols = cols[-1:] + cols[:-1]
preds = preds[cols]
# preds.to_csv('../auc/%s-%03d.csv' % (PREFIX, i))
return preds
def test_step(images, labels):
output = model(images, training=False)
t_loss = loss_object(labels, output)
test_loss(t_loss)
test_accuracy(labels, output)
aucvalue = auc(labels,
output,
weights=None,
num_thresholds=200,
name=None,
summation_method='trapezoidal')
test_auc(aucvalue)
def train_step(images, labels):
with tf.GradientTape() as tape:
output = model(images, training=True)
loss = loss_object(labels, output)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
train_loss(loss)
train_accuracy(labels, output)
aucvalue = auc(labels,
output,
weights=None,
num_thresholds=200,
name=None,
summation_method='trapezoidal')
train_auc(aucvalue)
def evaluate_c16_main():
eva_score = get_testset_score()
#print "##################################"
#print eva_score
gt_score = get_ground_truth()
#print "++++++++++++++++++++++++++++++++++++"
#print gt_score
#print "================================="
save_testset_score(eva_score, gt_score)
fig_name = "c16_auc_mean_cnt_%d.jpg" % (MEAN_COUNT)
fig_path = os.path.join(dir_c16_outs, fig_name)
#print gt_score.shape
#print eva_score.shape
auc_score = auc.auc(gt_score, eva_score, fig_path)
print("Meat count:%d C16 AUC is :%0.4lf" % (MEAN_COUNT, auc_score))
def conv1d_auc(i):
pred_data = pd.read_csv('../con/%03d.csv' % i, index_col='user_id')
pred_data = scaler.transform(pred_data)
pred = model.predict(reshape(pred_data))
pred_df = pd.DataFrame(pred)
pred_df['auc'] = 0.0
ans = pd.read_csv('../public/label-%03d.csv' % i, index_col='user_id')
pi = 0
for index, ans_row in ans.iterrows():
a = auc.auc(pred[pi], ans_row)
pred_df.at[pi, 'auc'] = a
pi += 1
cols = pred_df.columns.tolist()
cols = cols[-1:] + cols[:-1]
pred_df = pred_df[cols]
pred_df.to_csv('../conv1d-auc/%s-%03d.csv' %
(datetime.now().isoformat(timespec='minutes'), i))
def make_learning_curve(X,
Y,
min_repeat,
max_repeat,
ebar,
max_point_num,
debug=False,
useRPMat=False):
if debug:
print "ENTER MLC"
"""x, y, e = make_learning_curve(a, X, Y, min_repeat,
max_repeat, ebar, max_point_num)
% Make the learning curve
% Inputs:
% X -- data matrix
% Y -- labels
% Returns:
% x -- Number of samples
% y -- Performance (AUC)
% e -- error bar
"""
"""print 'X'
print X.shape
print X.sum()
print 'Y'
print Y.shape
print Y.sum()"""
#die
if debug:
print "Casting X to 64 bit"
assert Y.any(axis=0).all(),\
("Some classes have no example, at least one should be present "
"for each class")
X = N.cast['float64'](X)
Y = N.cast['float64'](Y.copy())
# Verify dimensions and set target values
p, n = X.shape
if len(Y.shape) == 1:
Y = (N.ones((1, 1)) * Y).T
pp, cn = Y.shape
sep_num = cn
if pp != p:
raise Exception('Size mismatch. X has ' + str(p) +
' examples but Y has ' + str(pp) + ' labels')
#
# If only 2 classes, the 2nd col is the same as the first but opposite
if cn == 2 and N.all(Y.sum(axis=1)):
Y = (N.ones((1, 1)) * Y[:, 0]).T
sep_num = 1
#
Y[Y == 0] = -1
if debug:
time.sleep(2)
print "Creating the data matrices"
# Create the data matrices (Y at this stage is still multi-column)
D = data_struct(X, Y)
feat_num = D.X.shape[1]
K = None
#if not pd_check.pd_check(D.X):
#die
# if debug:
# time.sleep(2)
# print "kernelizing"
# K = kernelize.kernelize(D);
#
if not pd_check.pd_check(D.X) and D.X.shape[0] < 2000:
D = kernelize.kernelize(D)
# Load random splits (these are the same for everyone)
RP = None
if useRPMat and os.path.exists(rp_path) and os.path.isfile(rp_path):
RP = io.loadmat(rp_path, struct_as_record=False)['RP']
#print RP
rp, mr = RP.shape
if rp < p:
if debug:
print 'make_learning_curve::warning: RP too small'
RP = None
else:
max_repeat = min(max_repeat, mr)
RP = N.ceil(N.cast['float64'](RP) / (float(rp) / float(p)))
RP = RP.astype(int)
if debug:
print 'make_learning_curve: using RP of dim ' + \
str(rp) + \
'x' + \
str(mr) + \
' min=' + \
str(RP.min()) + \
' max=' + \
str(RP.max()) + \
', max_repeat=' + \
str(max_repeat)
else:
print 'make_learning_curve::warning: no RP file found\n'
if debug:
time.sleep(2)
print "Computing sample sizes"
# Sample sizes scaled in log2
m = N.floor(math.log(p, 2))
x = 2.**N.arange(0, int(m) + 1)
if x[-1] != p:
x = N.hstack((x, [p]))
x = x[0:-1] # need to remove the last point
#print 'warning: this is a likely place i could have messed things up'
if max_point_num is None:
point_num = x.shape[0]
else:
point_num = min(x.shape[0], max_point_num)
#
# Loop over the sample sizes
x = x[0:point_num]
x = N.cast['uint32'](x)
y = N.zeros(x.shape)
e = N.zeros(x.shape)
for k in xrange(0, point_num):
if debug:
print '-------------------- Point %d ----------------------' % k
#
A = N.zeros((sep_num, 1))
E = N.zeros((sep_num, 1))
e[k] = N.Inf
# Loop over number of "1 vs all" separations
for j in xrange(0, sep_num):
if debug:
print ' sep %d -- ' % j
""
repnum = 0
area = []
# Loop over repeats (floating number of repeats)
while repnum < min_repeat or \
(E[j] > ebar and repnum < max_repeat):
if debug:
print 'repeat %d **' % repnum
print 'min_repeat: ' + str(min_repeat)
print 'max_repeat: ' + str(max_repeat)
pass
#
if RP is None:
rp = randperm(p)
else:
rp = RP[0:p, repnum] - 1
tr_idx = rp[0:x[k]]
te_idx = rp[x[k]:]
if debug:
print "Obtaining sub arrays"
time.sleep(2)
if pd_check.pd_check(D): # kernelized version
if debug:
print 'pd_check ok, using kernelized version'
Dtr = D.subdim(tr_idx, tr_idx, [j])
Dte = D.subdim(te_idx, tr_idx, [j])
#elif x[k] < feat_num: # kernelized too (for speed reason)
# if debug:
# print 'x[k] < feat_num, using kernelized version'
# Dtr = K.subdim(tr_idx, tr_idx, [j]);
# Dte = K.subdim(te_idx, tr_idx, [j]);
else: # primal version
if debug:
print 'using non-kernelized version'
Dtr = data_struct(D.X[tr_idx, :], D.Y[tr_idx, j])
Dte = data_struct(D.X[te_idx, :], D.Y[te_idx, j])
if debug:
time.sleep(2)
print "Training classifier"
d, m = train.train(Dtr)
if debug:
time.sleep(2)
print "Computing test values"
#print 'Dte.Y'
#print Dte.Y
#assert False
d1 = test.test(m, Dte)
assert d1.X.shape[0] != 0, "d1.X.shape[0] == 0"
assert repnum == len(area), "repnum == len(area)"
#print 'target'
#print d1.Y
#print d1.Y.shape
#print d1.Y.sum()
#assert False
if debug:
time.sleep(2)
print "Computing auc"
area.append(auc.auc(d1.X, d1.Y, dosigma=False)[0])
if debug:
time.sleep(2)
print "done"
repnum += 1
E[j] = N.asarray(area).std() / N.sqrt(repnum)
# repnum loop
assert not N.any(N.isnan(area))
A[j] = N.asarray(area).mean()
if N.isnan(A[j]):
assert False, "Invalid area: " + str(area)
#
#end % for j=1:sep_num
e[k] = E.mean()
y[k] = A.mean()
assert not N.isnan(y[k])
if debug:
print '==> ' + \
str(repnum) + \
' repeats, auc=' + \
str(y[k]) + \
'+-' + \
str(e[k]) + \
' -----------------'
#
# % Loop over k
# Add point with 0 examples
x = N.concatenate((N.asarray([0]), x))
P = 0.5
y = N.concatenate((N.asarray([P]), y))
e = N.concatenate((N.asarray([N.sqrt(P * (1 - P) / p)]), e))
return x, y, e
model.load_weights('model_save/deep_fm_sample-ep001-loss0.184-val_loss0.172.h5')
# model = load_model('model_save/deep_fm_sample-ep001-loss0.192-val_loss0.176.h5')
data = pd.read_csv("./data/sample/validation.txt")
# 1.Label Encoding for sparse features,and do simple Transformation for dense features
for feat in sparse_features:
lbe = LabelEncoder()
data[feat] = lbe.fit_transform(data[feat])
# 2.count #unique features for each sparse field
sparse_feature_dim = {feat: data[feat].nunique()
for feat in sparse_features}
# 3.generate input data for model
model_input = [data[feat].values for feat in sparse_feature_dim]
pred = model.predict(model_input, batch_size, 1)
label = data[target].values.flatten().tolist()
pred = pred.flatten().tolist()
with open('data/pctr', 'w') as fw:
for i in range(len(pred)):
if i % 10000 == 0:
print('label: %f, pred: %f' % (label[i], pred[i]))
to_write = str(i+1)+','+str(label[i])+','+str(pred[i])+'\n'
fw.write(to_write)
fw.close()
AUC = auc.auc(label, pred)
print('auc: %f' % AUC)
print("demo done")
s = "/valid/fluorescence_valid.txt"
sn = None
sp = "/valid/networkPositions_valid.txt"
TV, _, PV = read(s, sn, sp, Knormal)
print "reading test..."
s = "/test/fluorescence_test.txt"
sn = None
sp = "/test/networkPositions_test.txt"
TT, _, PT = read(s, sn, sp, Knormal)
[RN2, RT, RV, RN3] = learnAndPredict(TN1, CN1, [TN2, TT, TV, TN3])
suff = np.random.randint(10000)
f = open("./res_ver" + str(VER) + ".csv", 'w')
f.write("NET_neuronI_neuronJ,Strength\n")
for i in range(1000):
for j in range(1000):
f.write("valid_" + str(i + 1) + "_" + str(j + 1) + "," +
str(RV[i][j]) + "\n")
for i in range(1000):
for j in range(1000):
f.write("test_" + str(i + 1) + "_" + str(j + 1) + "," +
str(RT[i][j]) + "\n")
f.close()
print "Wrote solution of VER ==", str(VER)
RN2_ = RN2.flatten().tolist()
a = auc.auc(CN2.flatten().tolist(), RN2_)
RN3_ = RN3.flatten().tolist()
a2 = auc.auc(CN3.flatten().tolist(), RN3_)
print("RES: %.2f learning (%.2f, %.2f)" %
((a + a2) * 50, a * 100, a2 * 100))
batch[key + 1].values for key in np.hstack((
np.arange(0, 6),
np.arange(6, len(sparse_feature_dim))))
]
labels.extend(batch[0].values)
pred = model.predict_on_batch(X)
preds.extend(pred.flatten().tolist())
cnt += batch_size
if cnt % (batch_size * 100):
print(pred[0])
print('calculating auc ......')
AUC = auc.auc(labels, preds)
print('auc: %f' % AUC)
elif mode == 'pred':
# model.load_weights('model_save/deep_fm_fn_bs10000-ep001-loss0.155-val_loss0.153.h5') # auc: 0.714774
#model.load_weights('model_save/deep_fm_fn_bs15000-ep001-loss0.156-val_loss0.152.h5') # auc: 0.717083
#model.load_weights('model_save/deep_fm_fn-ep002-loss0.154-val_loss0.154-bs15000-ee20-hz[128, 128].h5') # auc: 0.718581
#model.load_weights('model_save/deep_fm_fn-ep020-loss0.153-val_loss0.153-bs15000-ee20-hz[5, 600].h5') # auc: 0.719317
model.load_weights(
'model_save/deep_fm_fn-ep043-loss0.152-val_loss0.152-bs15000-ee20-hz[3, 600].h5'
) # auc: 0.722419
ctr = []
reader = pd.read_csv(
'/home/yezhizi/Documents/python/2018DM_Project/track2/KDD_Track2_solution.csv',
chunksize=chunk_size)
def auc_scorer(estimator, X, y):
predicted = estimator.predict_proba(X)[:,1]
return auc.auc(y, predicted)
pred = model.predict_on_batch(X)
preds.extend(pred.flatten().tolist())
cnt += batch_size
if cnt % (batch_size * 100):
print(pred[0])
now_time = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
print(now_time)
print('calculating auc ......')
print('labels； %d' % len(click))
print('preds: %d' % len(preds))
AUC = auc.auc(np.array(click, dtype=np.float) / np.array(imp, dtype=np.float), preds)
print('auc: %f' % AUC)
AUC = auc.scoreClickAUC(click, imp, preds)
print('scoreClickAUC: %f' % AUC)
now_time = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
print(now_time)
# writing preds to csv
with open('data/' + model_name + '.csv', 'w') as fw:
for i in range(len(preds)):
if i % 1000000 == 0:
print('label: %f, pred: %f' % (click[i]/imp[i], preds[i]))
to_write = str(preds[i]) + '\n'
fw.write(to_write)
fw.close()
print "reading valid..."
s = "/valid/fluorescence_valid.txt"
sn = None
sp = "/valid/networkPositions_valid.txt"
TV, _, PV = read(s,sn,sp, Knormal)
print "reading test..."
s = "/test/fluorescence_test.txt"
sn = None
sp = "/test/networkPositions_test.txt"
TT, _, PT = read(s,sn,sp, Knormal)
[RN2, RT, RV, RN3] = learnAndPredict(TN1, CN1, [TN2, TT, TV, TN3])
suff = np.random.randint(10000)
f = open("./res_ver"+str(VER)+".csv", 'w')
f.write("NET_neuronI_neuronJ,Strength\n")
for i in range (1000):
for j in range (1000):
f.write("valid_" +str(i+1)+"_"+str(j+1)+","+str(RV[i][j])+"\n")
for i in range (1000):
for j in range (1000):
f.write("test_" +str(i+1)+"_"+str(j+1)+","+str(RT[i][j])+"\n")
f.close()
print "Wrote solution of VER ==", str(VER)
RN2_ = RN2.flatten().tolist()
a = auc.auc(CN2.flatten().tolist(),RN2_)
RN3_ = RN3.flatten().tolist()
a2 = auc.auc(CN3.flatten().tolist(),RN3_)
print ("RES: %.2f learning (%.2f, %.2f)" % ((a+a2)*50, a*100, a2*100 ))
sp = "/" + "small" + "/networkPositions_" + name + ".txt"
print name
TSmall5, CSmall5, PSmall5 = read(s, sn, sp, Knormal)
gc.collect()
name = "iNet1_Size100_CC06inh"
s = "/" + "small" + "/fluorescence_" + name + ".txt"
sn = "/" + "small" + "/network_" + name + ".txt"
sp = "/" + "small" + "/networkPositions_" + name + ".txt"
print name
TSmall6, CSmall6, PSmall6 = read(s, sn, sp, Knormal)
gc.collect()
[RS4, RS6] = learnAndPredict(TSmall5, CSmall5, [TSmall4, TSmall6])
RS4_ = RS4.flatten().tolist()
a = auc.auc(CSmall4.flatten().tolist(), RS4_)
RS6_ = RS6.flatten().tolist()
a2 = auc.auc(CSmall6.flatten().tolist(), RS6_)
print("RES: %.2f Small4, %.2f Small6" % (a * 100, a2 * 100))
f = open("res_small_4_6.csv", 'w')
f.write("NET_neuronI_neuronJ,Strength\n")
for i in range(NN):
for j in range(NN):
f.write("valid_" + str(i + 1) + "_" + str(j + 1) + "," +
str(RS4[i][j]) + "\n")
for i in range(NN):
for j in range(NN):
f.write("test_" + str(i + 1) + "_" + str(j + 1) + "," +
str(RS6[i][j]) + "\n")
f.close()
print "Wrote solution to ./res_Small_4_6.csv"
import pandas as pd
from auc import auc
DATA_PATH = '../public/'
i = 1
ans = pd.read_csv(DATA_PATH + 'label-%03d.csv' % i, index_col='user_id')
preds = pd.read_csv('../predict-%03d.csv' % i, index_col='user_id')
preds['auc'] = 0.0
for index, pred in preds.iterrows():
truth = ans.loc[index]
a = auc(pred[0:28], truth)
preds.at[index, 'auc'] = round(a, 4)
cols = preds.columns.tolist()
cols = cols[-1:] + cols[:-1]
preds = preds[cols]
preds.to_csv('../auc-%03d.csv' % i)
sn = "/" + "small" + "/network_" + name + ".txt"
sp = "/" + "small" + "/networkPositions_" + name + ".txt"
print name
TSmall5, CSmall5, PSmall5 = read(s, sn, sp, Knormal)
gc.collect()
name = "iNet1_Size100_CC06inh"
s = "/" + "small" + "/fluorescence_" + name + ".txt"
sn = "/" + "small" + "/network_" + name + ".txt"
sp = "/" + "small" + "/networkPositions_" + name + ".txt"
print name
TSmall6, CSmall6, PSmall6 = read(s, sn, sp, Knormal)
gc.collect()
[RS4, RS6] = learnAndPredict(TSmall5, CSmall5, [TSmall4, TSmall6])
RS4_ = RS4.flatten().tolist()
a = auc.auc(CSmall4.flatten().tolist(), RS4_)
RS6_ = RS6.flatten().tolist()
a2 = auc.auc(CSmall6.flatten().tolist(), RS6_)
print ("RES: %.2f Small4, %.2f Small6" % (a * 100, a2 * 100))
f = open("res_small_4_6.csv", "w")
f.write("NET_neuronI_neuronJ,Strength\n")
for i in range(NN):
for j in range(NN):
f.write("valid_" + str(i + 1) + "_" + str(j + 1) + "," + str(RS4[i][j]) + "\n")
for i in range(NN):
for j in range(NN):
f.write("test_" + str(i + 1) + "_" + str(j + 1) + "," + str(RS6[i][j]) + "\n")
f.close()
print "Wrote solution to ./res_Small_4_6.csv"