def train_DeepFM():
X, y, sparse_list, dense_list = feature_engineering.get_NN_data(
use_over_sampler=True)
data = pd.DataFrame(y)
dnn_feature_columns = linear_feature_columns = sparse_list + dense_list
feature_names = inputs.get_feature_names(linear_feature_columns +
dnn_feature_columns)
# kFold cv
models = []
scores = []
kf = StratifiedKFold(n_splits=5, random_state=42)
for i, (tdx, vdx) in enumerate(kf.split(X, y)):
print(f'Fold : {i}')
X_train, X_val, y_train, y_val = X.loc[tdx], X.loc[vdx], y.loc[
tdx], y.loc[vdx]
y_true = y_val
X_train = {name: X_train[name] for name in feature_names}
X_val = {name: X_val[name] for name in feature_names}
model = DeepFM(linear_feature_columns,
dnn_feature_columns,
dnn_hidden_units=(128, 64),
dnn_use_bn=True,
task='binary',
dnn_dropout=0.5)
best_param_path = '/Users/a_piao/PycharmProjects/BankMarketing/workspace/DeepFM/best_param_DeepFM.py_%d.h5' % i
# best_param_path = 'best_param_%s_%d.h5' % (os.path.basename(__file__), i)
if os.path.exists(best_param_path):
model.load_weights(best_param_path)
else:
model.compile("adam",
"binary_crossentropy",
metrics=['binary_crossentropy'])
# tb = TensorBoard(log_dir="/Users/a_piao/PycharmProjects/NightLife_recommend/workspace/DeepFM/log", write_images=1, histogram_freq=1)
es = EarlyStopping(monitor='val_binary_crossentropy',
mode='min',
patience=20)
mc = ModelCheckpoint(best_param_path,
monitor='val_binary_crossentropy',
mode='min',
save_best_only=False,
verbose=False,
save_weights_only=True)
# model.fit(X_train, y_train, batch_size=512, epochs=1000, verbose=2, validation_split=0.2, callbacks=[es, mc])
model.fit(X_train,
y_train,
validation_data=(X_val, y_val),
batch_size=1024,
epochs=1000,
verbose=2,
callbacks=[es, mc])
# model.fit(X_train, y_train, batch_size=1024, epochs=100, verbose=2, callbacks=[es, mc])
model.load_weights(best_param_path)
y_pred = model.predict(X_val, batch_size=64).flatten()
auc = roc_auc_score(y_true, y_pred)
print("AUC score at %d floder: %f" % (i, auc))
scores.append(auc)
models.append(model)
data.loc[vdx, 'y_pred'] = y_pred
# print(data['y_pred'].value_counts())
# exit()
mean_score = np.mean(scores)
oof = roc_auc_score(data['y'], data['y_pred'])
print("5-floder total mean_score:", mean_score)
print("5-floder oof auc score:", oof)
print("----train %s finish!----" % 'DeepFM')
cal_roc_curve(data['y'], data['y_pred'], 'DeepFM')
return data['y_pred']
test_model_input = {name: test[name].values for name in sparse_features}
# Instantiate a FeatureMetas object, add your features' meta information to it
feature_metas = FeatureMetas()
for feat in sparse_features:
feature_metas.add_sparse_feature(name=feat,
one_hot_dim=data[feat].nunique(),
embedding_dim=32)
# a warning need to be fixed see https://stackoverflow.com/questions/35892412/tensorflow-dense-gradient-explanation
# Instantiate a model and compile it
model = DeepFM(feature_metas=feature_metas,
linear_slots=sparse_features,
fm_slots=sparse_features,
dnn_slots=sparse_features)
model.compile(optimizer="adam",
loss="binary_crossentropy",
metrics=['binary_crossentropy'])
# Train the model
history = model.fit(x=train_model_input,
y=train[target].values,
batch_size=128,
epochs=3,
verbose=2,
validation_split=0.2)
# Testing
pred_ans = model.predict(test_model_input, batch_size=256)
print("test LogLoss", round(log_loss(test[target].values, pred_ans), 4))
print("test AUC", round(roc_auc_score(test[target].values, pred_ans), 4))