print(len(x_train_str), len(y_train))
X_train_s, X_test_s, y_train_s, y_test_s = train_test_split(x_train_str,
y_train,
random_state=42,
test_size=0.2)
print('X_train_s.shape', len(X_train_s), len(X_train_s[0]))
for x in X_train_s:
print(len(x))
sys.exit(1)
print('y_train_s.shape', len(y_train_s), y_train_s[0].shape)
clf.fit(X_train_s, y_train_s)
print("predicting")
y_pred_scores_s = clf.predict_proba(X_test_s)
y_pred_scores_s[np.where(y_pred_scores_s == 0)] = -10
y_pred_s = y_pred_scores_s > -0.25
print(
'f1 micro:',
f1_score(y_true=y_test_s,
y_pred=y_pred_s[:, :y_test_s.shape[1]],
average='micro'))
print(
'f1 macro:',
f1_score(y_true=y_test_s,
y_pred=y_pred_s[:, :y_test_s.shape[1]],
average='macro'))
print(
def svm_train():
train_x, train_y, apps, train_y2 = get_data_set(train_path)
test_x, test_y, apps, test_y2 = get_data_set(test_path)
pred_x, _, apps, _ = get_data_set(pred_path)
# with open(CHANNEL_MODEL + 'svm_label.pkl', 'wb') as f:
# pickle.dump(label_dic, f)
logging.info('train {} test{}'.format(len(train_x), len(test_x)))
t = time.time()
data_set = train_x + test_x + pred_x
vec = TfidfVectorizer(ngram_range=(1, 3),
min_df=10,
max_df=0.9,
use_idf=1,
smooth_idf=1,
sublinear_tf=1)
#vec=HashingVectorizer(ngram_range=(1, 3))
vec.fit_transform(data_set)
#
with open(project_path + 'tfidf.pkl', 'wb') as f:
pickle.dump(vec, f)
# with open(CHANNEL_MODEL + 'tfidf.pkl', 'rb') as f:
# vec = pickle.load(f)
trn_term_doc = vec.transform(train_x)
print(label1_label2)
time.sleep(20)
tfidf_time = time.time()
logging.info('time spend {}'.format(tfidf_time - t))
logging.info('begin svm ')
lin_clf = svm.LinearSVC(C=1)
lin_clf = CalibratedClassifierCV(lin_clf)
clf = HierarchicalClassifier(
base_estimator=lin_clf,
class_hierarchy=label1_label2,
)
clf.fit(trn_term_doc, train_y)
print(clf.classes_)
logging.info('end svm ')
# with open(project_path + 'svm_model.pkl', 'wb') as f:
# pickle.dump(lin_clf, f)
train_preds = clf.predict(trn_term_doc)
train_preds_prob = clf.predict_proba(trn_term_doc)
print(len((clf.classes_)), train_preds_prob.shape)
time.sleep(20)
for reg, prob in zip(train_preds, train_preds_prob):
print(reg, list(prob.argsort()[-1:][::-1]))
time.sleep(20)
from sklearn.metrics import classification_report
logging.info('train {} accuracy_score {}, \n {}'.format(
'train', accuracy_score(train_y, train_preds),
classification_report(train_y, train_preds)))
t2 = time.time()
logging.info('time spend {}'.format(t2 - t))
test_term_doc = vec.transform(test_x)
test_preds_1 = clf.predict_proba(test_term_doc)
test_preds = clf.predict(test_term_doc)
logging.info('train {} accuracy_score {}, \n {}'.format(
'train', accuracy_score(train_y, train_preds),
classification_report(test_y2, test_preds)))
dic_lab = {}
for k, v in label_dic2.items():
dic_lab[v] = k
test_preds = []
test_preds = []
for prob in test_preds_1:
test_preds.append(list(prob.argsort()[-2:][::-1]))
test_y_name = []
test_preds_name = []
for real, pred in zip(test_y2, test_preds):
prd = pred[0]
print(real, pred)
for pr in pred:
if real == clf.classes_[pr]:
prd = pr
test_y_name.append(real)
test_preds_name.append(clf.classes_[prd])
logging.info('{} model on {} data accuracy_score {} top2 test\n {}'.format(
"train", test_path, accuracy_score(test_y_name, test_preds_name),
classification_report(test_y_name, test_preds_name)))
def main():
subexp_name = 'YoungJaeShinSamples/4'
anno_incomplete_file = '../data/data_jan2019_anno/anno_all_incomplete_YoungJaeShinSamples_4_useryoungjae.db'
anno_complete_file = '../data/data_jan2019_anno/anno_all_YoungJaeShinSamples_4_useryoungjae.db'
# anno_complete_file = '../data/data_jan2019_anno/anno_all_YoungJaeShinSamples_4_usertest0123.db'
thickthin_anno_file = '../data/data_jan2019_anno/anno_thickthin_v2_YoungJaeShinSamples_4_useryoungjae.db'
# thickthin_anno_file = '../data/data_jan2019_anno/anno_thickthin_v2_YoungJaeShinSamples_4_usertest0123.db'
data_path = os.path.join('../data/data_jan2019', subexp_name)
result_path = os.path.join('../results/data_jan2019_script/mat',
subexp_name)
clf_path = os.path.join(
'../results/data_jan2019_script/thickthinglue_clf_complete',
subexp_name)
if not os.path.exists(clf_path):
os.makedirs(clf_path)
# merge the annotation
if not os.path.exists(anno_complete_file):
os.system('cp %s %s' % (anno_incomplete_file, anno_complete_file))
# update the annotation into the all annotation
thickthin_oriname_newflakeids = readthickthindb(thickthin_anno_file)
n_thickthin_flakes = len(thickthin_oriname_newflakeids)
print(n_thickthin_flakes)
updatedb(anno_complete_file, thickthin_oriname_newflakeids)
# get the train/val split
split_name = os.path.join(clf_path, 'train_val_split.p')
if os.path.exists(split_name):
to_load = pickle.load(open(split_name, 'rb'))
train_names = to_load['train_names']
train_labels = to_load['train_labels']
val_names = to_load['val_names']
val_labels = to_load['val_labels']
else:
itemname_labels = readdb(anno_complete_file)
train_names, train_labels, val_names, val_labels = split_trainval(
itemname_labels)
to_save = dict()
to_save['train_names'] = train_names
to_save['train_labels'] = train_labels
to_save['val_names'] = val_names
to_save['val_labels'] = val_labels
pickle.dump(to_save, open(split_name, 'wb'))
# load flakes
flake_save_name = os.path.join(clf_path, 'train_val_data.p')
if os.path.exists(flake_save_name):
to_load = pickle.load(open(flake_save_name, 'rb'))
train_flakes = to_load['train_flakes']
train_feats = to_load['train_feats']
val_flakes = to_load['val_flakes']
val_feats = to_load['val_feats']
else:
img_names = os.listdir(data_path)
img_names.sort()
img_flakes = Parallel(n_jobs=8)(delayed(load_one_image)(
os.path.join(data_path, img_names[i]),
os.path.join(result_path, img_names[i][:-4] + '.p'))
for i in range(len(img_names)))
# pickle.dump(img_flakes, open(flake_save_name, 'wb'))
# load corresponding flakes
train_flakes, train_feats = locate_flakes(train_names, img_flakes,
img_names)
val_flakes, val_feats = locate_flakes(val_names, img_flakes, img_names)
to_save = dict()
to_save['train_flakes'] = train_flakes
to_save['train_feats'] = train_feats
# to_save['train_names'] = train_names
# to_save['train_labels'] = train_labels
to_save['val_flakes'] = val_flakes
to_save['val_feats'] = val_feats
# to_save['val_names'] = val_names
# to_save['val_labels'] = val_labels
pickle.dump(to_save, open(flake_save_name, 'wb'))
print('loading done')
# normalize data
mean_feat = np.mean(train_feats, axis=0, keepdims=True)
std_feat = np.std(train_feats, axis=0, keepdims=True)
norm_fea = {}
norm_fea['mean'] = mean_feat
norm_fea['std'] = std_feat
pickle.dump(norm_fea, open(os.path.join(clf_path, 'normfea.p'), 'wb'))
train_feats -= mean_feat
train_feats = train_feats / std_feat
# train_feats = train_feats / np.linalg.norm(train_feats, 2, axis=1, keepdims=True)
val_feats -= mean_feat
val_feats = val_feats / std_feat
# val_feats = val_feats / np.linalg.norm(val_feats, 2, axis=1, keepdims=True)
# run classifier
# method = 'linearsvm'
# method = 'ridge'
# method = 'rbfkernelsvm'
method = hyperparams['clf_method']
print(method)
C = hyperparams['C']
# C = 10
Cs = [
0.001,
0.01,
0.1,
0.25,
0.5,
0.75,
1,
2.5,
5,
]
from sklearn.decomposition import TruncatedSVD
from sklearn.pipeline import make_pipeline
from sklearn_hierarchical_classification.classifier import HierarchicalClassifier
from sklearn_hierarchical_classification.constants import ROOT
from sklearn_hierarchical_classification.metrics import h_fbeta_score, multi_labeled
from sklearn_hierarchical_classification.tests.fixtures import make_digits_dataset
class_hierarchy = {
ROOT: ["A", "2"],
"A": ["0", "1"],
}
for C in Cs:
clf_save_path = os.path.join(
clf_path, 'feanorm_weighted_classifier-%s-%f.p' % (method, C))
if os.path.exists(clf_save_path):
clf = pickle.load(open(clf_save_path, 'rb'))
else:
if method == 'hielinearsvm':
# clf = LinearSVC(random_state=0, tol=1e-5, C=C, max_iter=5e4, class_weight='balanced')
# clf = LinearSVC(random_state=0, tol=1e-5, C=C, max_iter=9e4, class_weight={0:1, 1:5, 2:1})#, multi_class='crammer_singer')
# clf.fit(train_feats, train_labels)
base_estimator = make_pipeline(
# TruncatedSVD(n_components=24),
SVC(gamma='auto', kernel="linear", probability=True,
C=C), )
# elif method == 'ridge':
# clf = RidgeClassifier(random_state=0, alpha=C)
# clf.fit(train_feats, train_labels)
elif method == 'hierbfsvm':
# clf = SVC(kernel='rbf', C=C, tol=1e-5, gamma='auto')
# clf.fit(train_feats, train_labels)
base_estimator = make_pipeline(
# TruncatedSVD(n_components=24),
SVC(gamma='auto', kernel="rbf", probability=True, C=C), )
else:
raise NotImplementedError
clf = HierarchicalClassifier(
base_estimator=base_estimator,
class_hierarchy=class_hierarchy,
)
train_labels_str = [str(_) for _ in train_labels]
clf.fit(train_feats, train_labels_str)
pickle.dump(clf, open(clf_save_path, 'wb'))
train_pred_cls = clf.predict(train_feats)
train_pred_cls = [int(_) for _ in train_pred_cls]
train_pred_scores = clf.predict_proba(train_feats)
val_pred_cls = clf.predict(val_feats)
val_pred_cls = [int(_) for _ in val_pred_cls]
val_pred_scores = clf.predict_proba(val_feats)
clf_vis_path = os.path.join(clf_path, subexp_name, 'vis',
'feanorm_weighted_%s-%f' % (method, C))
if not os.path.exists(clf_vis_path):
os.makedirs(clf_vis_path)
from sklearn.metrics import accuracy_score, average_precision_score, confusion_matrix, precision_recall_curve
train_acc = accuracy_score(train_labels, train_pred_cls)
val_acc = accuracy_score(val_labels, val_pred_cls)
train_conf = confusion_matrix(train_labels, train_pred_cls)
train_conf = train_conf / np.sum(train_conf, 1, keepdims=True)
val_conf = confusion_matrix(val_labels, val_pred_cls)
val_conf = val_conf / np.sum(val_conf, 1, keepdims=True)
# # val_acc = accuracy_score(, test_pred_cls)
# print('train acc: %.4f' % (train_acc))
print(train_conf)
print(val_conf)
# vis_error(val_pred_cls, val_pred_scores, val_labels, val_flakes, clf_vis_path, val_names, 'val')
# vis_error(train_pred_cls, train_pred_scores, train_labels, train_flakes, clf_vis_path, train_names, 'train')
# calculate map:
uniquelabels = [0, 1, 2]
train_aps = []
val_aps = []
fig = plt.figure()
ax = fig.add_subplot(111)
legends = ['thick', 'thin', 'glue']
for l in uniquelabels:
l_train_labels = [_ == l for _ in train_labels]
l_val_labels = [_ == l for _ in val_labels]
# if method == 'linearsvm':
# clf = LinearSVC(random_state=0, tol=1e-5, C=C)
# clf.fit(train_feats, l_train_labels)
# elif method == 'ridge':
# clf = RidgeClassifier(random_state=0, alpha=C)
# clf.fit(train_feats, l_train_labels)
# elif method == 'rbfkernelsvm':
# clf = SVC(kernel='rbf', C=C, tol=1e-5, gamma='auto')
# clf.fit(train_feats, l_train_labels)
# else:
# raise NotImplementedError
# l_train_pred_scores = clf.decision_function(train_feats)
# l_val_pred_scores = clf.decision_function(val_feats)
train_aps.append(
average_precision_score(l_train_labels, train_pred_scores[:,
l]))
val_aps.append(
average_precision_score(l_val_labels, val_pred_scores[:, l]))
# print(val_pred_scores[:, l])
# print(np.array(l_val_labels, dtype=np.uint8))
precision_l, recall_l, _ = precision_recall_curve(
np.array(l_val_labels, dtype=np.uint8), val_pred_scores[:, l])
print(l, getPrecisionAtRecall(precision_l, recall_l, 0.90),
getPrecisionAtRecall(precision_l, recall_l, 0.95))
ax.plot(recall_l, precision_l, label=legends[l])
plt.legend()
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.ylim([0.0, 1.05])
plt.xlim([0.0, 1.0])
plt.savefig(os.path.join(clf_path,
'feanorm_weighted_%s-%f.png' % (method, C)),
dpi=300)
plt.close(fig)
# train_labels = (np.array(train_labels) + 1)//2
# val_labels = (np.array(val_labels) + 1 ) // 2
# train_ap = average_precision_score(train_labels, train_pred_scores)
# val_ap = average_precision_score(val_labels, pred_scores)
print(train_aps)
print(val_aps)
print('%s-%f: train: %.4f, val: %.4f, ap train: %4f, ap val: %4f' %
(method, C, train_acc, val_acc, np.mean(train_aps),
np.mean(val_aps)))
base_estimator = make_pipeline(vectorizer, bclf)
clf = HierarchicalClassifier(base_estimator=base_estimator,
class_hierarchy=class_hierarchy,
algorithm="lcn",
training_strategy="siblings",
preprocessing=True,
mlb=mlb,
use_decision_function=True)
print("training classifier")
clf.fit(X_train_raw, y_train[:, :])
print("predicting")
y_pred_scores = clf.predict_proba(X_dev_raw)
y_pred_scores[np.where(y_pred_scores == 0)] = -10
y_pred = y_pred_scores > -0.25
if train == 1:
print('f1 micro:',
f1_score(y_true=y_dev, y_pred=y_pred, average='micro'))
print('f1 macro:',
f1_score(y_true=y_dev, y_pred=y_pred, average='macro'))
print(classification_report(y_true=y_dev, y_pred=y_pred))
else:
import networkx as nx
graph = nx.DiGraph(hierarchy)
print_results("submission_baseline.txt", hierarchy, y_pred > -0.25,
mlb, ids, graph)