def read_dataset(is_merged):
debug = DEBUG
if debug:
filename_train = '../input/debug{}/{}_debug{}.feather'.format(
debug, 'train_translated', debug)
filename_test = '../input/debug{}/{}_debug{}.feather'.format(
debug, 'test_translated', debug)
else:
filename_train = '../input/{}.feather'.format('train_translated')
filename_test = '../input/{}.feather'.format('test_translated')
print_doing('reading train, test and merge')
if is_merged:
df = read_train_test(filename_train,
filename_test,
'.feather',
is_merged=True)
if debug: print(df.head())
else:
train_df, test_df = read_train_test(filename_train,
filename_test,
'.feather',
is_merged=False)
if debug:
print(train_df.head())
print(test_df.head())
print_memory()
if is_merged:
return df
else:
return train_df, test_df
def predict_sub(model_lgb, testdex, test, subfilename):
print_doing('predicting')
lgpred = model_lgb.predict(test)
lgsub = pd.DataFrame(lgpred, columns=["deal_probability"], index=testdex)
lgsub['deal_probability'].clip(0.0, 1.0, inplace=True)
print('saving submission file to', subfilename)
lgsub.to_csv(subfilename, index=True, header=True)
print('done')
def read_dataset_origin(dataset):
filename_train = '../input/train.csv'
filename_test = '../input/test.csv'
print_doing('reading train, test and merge')
df = read_train_test(filename_train,
filename_test,
'.feather',
is_merged=1)
print_memory()
print(df.head())
return df
def prepare_training(mat_filename, dir_feature, predictors, is_textadded):
print_header('Load features')
df, y, len_train, traindex, testdex = load_train_test(['item_id'], TARGET,
DEBUG)
del len_train
gc.collect()
df = drop_col(df, REMOVED_LIST)
# add features
print_doing('add tabular features')
for feature in predictors:
dir_feature_file = dir_feature + feature + '.pickle'
if not os.path.exists(dir_feature_file):
print('can not find {}. Please check'.format(dir_feature_file))
else:
if feature in df:
print('{} already added'.format(feature))
else:
print_doing_in_task('adding {}'.format(feature))
df = add_feature(df, dir_feature_file)
print_memory()
if is_textadded:
# add text_feature
print_doing_in_task('add text features')
ready_df, tfvocab = get_text_matrix(mat_filename, 'all', 2, 0)
# stack
print_doing_in_task('stack')
X = hstack([
csr_matrix(df.loc[traindex, :].values),
ready_df[0:traindex.shape[0]]
]) # Sparse Matrix
testing = hstack([
csr_matrix(df.loc[testdex, :].values), ready_df[traindex.shape[0]:]
])
print_memory()
print_doing_in_task('prepare vocab')
tfvocab = df.columns.tolist() + tfvocab
for shape in [X, testing]:
print("{} Rows and {} Cols".format(*shape.shape))
print("Feature Names Length: ", len(tfvocab))
else:
tfvocab = df.columns.tolist()
testing = hstack([csr_matrix(df.loc[testdex, :].values)])
X = hstack([csr_matrix(df.loc[traindex, :].values)]) # Sparse Matrix
return X, y, testing, tfvocab, df.columns.tolist(), testdex
def do_dataset(dataset):
train_df, test_df = read_dataset(False, DEBUG)
len_train = len(train_df)
if dataset=='train':
df = train_df
del test_df; gc.collect()
else:
df = test_df
del train_df; gc.collect()
if DEBUG:
storename = '../processed_features_debug{}/{}_debug{}.h5'.format(DEBUG, dataset, DEBUG)
featuredir = '../processed_features_debug{}/'.format(DEBUG)
else:
storename = '../processed_features/{}.h5'.format(dataset)
featuredir = '../processed_features/'
temp = add_dataset_to_hdf5(storename, df)
if DEBUG: print(temp.isnull().sum(axis=0))
files = glob.glob(featuredir + '*.pickle')
for file in files:
if 'text_feature_kernel' not in file:
print(file)
filename = file
print ('\n>> doing', filename)
df = load_pickle(filename)
if DEBUG:
print(df.tail())
print_doing('extract')
if DEBUG: print(df.head()); print(df.tail())
if dataset=='train':
df_new = df.iloc[:len_train]
if DEBUG:
print('train: ', df.head())
print(df_new.isnull().sum(axis=0))
else:
df_new = df.iloc[len_train:]
if DEBUG:
print('test: ', df.tail())
print(df_new.isnull().sum(axis=0))
print('merging...')
temp = add_dataset_to_hdf5(storename, df_new)
if DEBUG: print(temp.isnull().sum(axis=0))
print_memory()
def read_dataset():
debug = DEBUG
if debug:
filename_train = '../input/debug{}/{}_debug{}.feather'.format(
debug, 'train', debug)
filename_test = '../input/debug{}/{}_debug{}.feather'.format(
debug, 'test', debug)
else:
filename_train = '../input/{}.feather'.format('train')
filename_test = '../input/{}.feather'.format('test')
print_doing('reading train, test and merge')
df = read_train_test(filename_train,
filename_test,
'.feather',
is_merged=1)
print_memory()
print(df.head())
return df
def read_dataset_deal_probability(seed):
debug = DEBUG
if debug:
filename_train = '../input/debug{}/{}_debug{}.feather'.format(
debug, 'train', debug)
filename_test = '../input/debug{}/{}_debug{}.feather'.format(
debug, 'test', debug)
else:
filename_train = '../input/{}.feather'.format('train')
filename_test = '../input/{}.feather'.format('test')
print_doing('reading train, test and merge')
train_df, test_df = read_train_test(filename_train,
filename_test,
'.feather',
is_merged=0)
df = find_df_local_valid_and_make_deal_prob_nan(train_df, test_df, seed)
print_memory()
print(df.head())
return df
def get_svdtruncated_vectorizer(todir):
print_doing('doing svdtruncated text feature')
filename = todir + 'text_feature_kernel.pickle'
savename = todir + 'truncated_text_feature_kernel.pickle'
if os.path.exists(savename):
print('done already...')
with open(savename, "rb") as f:
svd_matrix, vocab = pickle.load(f)
with open(filename, "rb") as f:
tfid_matrix, tfvocab = pickle.load(f)
else:
with open(filename, "rb") as f:
tfid_matrix, tfvocab = pickle.load(f)
svdT = TruncatedSVD(n_components=400)
print_doing_in_task('truncated svd')
svd_matrix = svdT.fit_transform(tfid_matrix)
print_doing_in_task('convert to sparse')
svd_matrix = sparse.csr_matrix(svd_matrix, dtype=np.float32)
vocab = []
for i in range(np.shape(svd_matrix)[1]):
vocab.append('lsa' + str(i + 1))
with open(savename, "wb") as f:
pickle.dump((svd_matrix, vocab),
f,
protocol=pickle.HIGHEST_PROTOCOL)
print('---- before truncate')
print(tfid_matrix.shape), print('len of feature:', len(tfvocab))
print('---- after truncate')
print(svd_matrix.shape), print('len of feature:', len(vocab))
if DEBUG:
print(tfid_matrix)
print('\n')
print(svd_matrix)
del svd_matrix, vocab, tfid_matrix, tfvocab
gc.collect()
print_memory()
def do_dataset(dataset):
train_df, test_df = read_dataset(False)
len_train = len(train_df)
if dataset == 'train':
df = train_df
del test_df
gc.collect()
else:
df = test_df
del train_df
gc.collect()
if DEBUG:
storename = '../processed_features_debug{}/{}_debug{}.h5'.format(
DEBUG, dataset, DEBUG)
featuredir = '../processed_features_debug{}/'.format(DEBUG)
else:
storename = '../processed_features/{}.h5'.format(dataset)
featuredir = '../processed_features/'
add_dataset_to_hdf5(storename, df)
files = glob.glob(featuredir + '*.pickle')
for file in files:
if 'train' not in file and 'test' not in file:
if 'text_feature_kernel' in file:
print(file)
filename = file
print('\n>> doing', filename)
if DEBUG:
if '_en' in file:
savename = '../processed_features_debug{}/{}_text_dense_en_debug{}.pickle'.format(
DEBUG, dataset, DEBUG)
else:
savename = '../processed_features_debug{}/{}_text_dense_debug{}.pickle'.format(
DEBUG, dataset, DEBUG)
else:
if '_en' in file:
savename = '../processed_features/{}_text_dense_en.h5'.format(
dataset)
else:
savename = '../processed_features/{}_text_dense.h5'.format(
dataset)
if os.path.exists(savename):
print('done already')
else:
mat = load_pickle(filename)
mat = mat.todense()
if DEBUG: print(mat.shape, np.sum(mat))
print_doing('extract')
if DEBUG:
print(mat[0:5, 0:7])
print(mat[-5:, 0:7])
if dataset == 'train':
mat = mat[:len_train, :]
if DEBUG: print(mat.shape, np.sum(mat))
if DEBUG: print('train: ', print(mat[0:5, 0:7]))
else:
mat = mat[len_train:, :]
print(mat.shape, np.sum(mat))
if DEBUG: print(mat.shape, np.sum(mat))
if DEBUG: print('test: ', print(mat[-5:, 0:7]))
print(mat)
print('merging...')
save_file(mat, savename, '.pickle')
print_memory()
else:
print(file)
filename = file
print('\n>> doing', filename)
df = load_pickle(filename)
print_doing('extract')
if DEBUG:
print(df.head())
print(df.tail())
if dataset == 'train':
df = df.iloc[:len_train]
if DEBUG: print('train: ', df.head())
else:
df = df.iloc[len_train:]
if DEBUG: print('test: ', df.tail())
print('merging...')
add_dataset_to_hdf5(storename, df)
print_memory()
def DO(mat_filename, storename,num_leaves,max_depth, option, boosting_type):
frac = FRAC
print('------------------------------------------------')
print('start...')
print('fraction:', frac)
print('prepare predictors, categorical and target...')
predictors = PREDICTORS
print (predictors)
categorical = get_categorical(predictors)
target = TARGET
subfilename = yearmonthdate_string + '_' + str(len(predictors)) + \
'features_' + boosting_type + '_cv_' + str(int(100*frac)) + \
'percent_full_%d_%d'%(num_leaves,max_depth) + '_OPTION' + str(option) + '.csv.gz'
modelfilename = yearmonthdate_string + '_' + str(len(predictors)) + \
'features_' + boosting_type + '_cv_' + str(int(100*frac)) + \
'percent_full_%d_%d'%(num_leaves,max_depth) + '_OPTION' + str(option)
print('----------------------------------------------------------')
print('SUMMARY:')
print('----------------------------------------------------------')
print('predictors:',predictors)
print('number of predictors: {} \n'.format(len(predictors)))
print('categorical', categorical)
print('number of predictors: {} \n'.format(len(categorical)))
print('taget {} \n'.format(target))
print('submission file name: {} \n'.format(subfilename))
print('model file name: {} \n'.format(modelfilename))
# print('fraction:', frac)
# print('option:', option)
print('----------------------------------------------------------')
train_df = read_processed_h5(storename, predictors+target, categorical)
print(train_df.info())
print(train_df.head())
train_df["price"] = np.log(train_df["price"]+0.001)
train_df["price"].fillna(-999,inplace=True)
# train_df["price"] = train_df["price"].astype('float')
# train_df["image_top_1"].fillna(-999,inplace=True)
print(train_df.head()); print(train_df.info())
# train_df = train_df.sample(frac=frac, random_state = SEED)
print_memory('afer reading train:')
print(train_df.head())
print("train size: ", len(train_df))
gc.collect()
print_doing('cleaning train...')
train_df_array = train_df[predictors].values
train_df_labels = train_df[target].values.astype('int').flatten()
del train_df; gc.collect()
print_memory()
print_doing('reading text matrix')
train_mat_text, tfvocab = get_text_matrix(mat_filename, 'train', DEBUG, train_df_array.shape[0])
print_memory()
print_doing('stack two matrix')
train_df_array = hstack([csr_matrix(train_df_array),train_mat_text])
print_memory()
new_predictors = tfvocab
predictors = predictors + new_predictors
del train_mat_text; gc.collect()
print('----------------------------------------------------------')
print("Training...")
start_time = time.time()
params = {
'boosting_type': boosting_type,
'objective': 'regression',
'metric': 'rmse',
'learning_rate': 0.02,
'num_leaves': num_leaves, # we should let it be smaller than 2^(max_depth)
'max_depth': max_depth, # -1 means no limit
'subsample': 0.9, # Subsample ratio of the training instance.
'subsample_freq': 1, # frequence of subsample, <=0 means no enable
'feature_fraction': 0.9, # Subsample ratio of columns when constructing each tree.
# 'min_child_weight': 0, # Minimum sum of instance weight(hessian) needed in a child(leaf)
# 'subsample_for_bin': 200000, # Number of samples for constructing bin
# 'min_split_gain': 0, # lambda_l1, lambda_l2 and min_gain_to_split to regularization
# 'reg_alpha': 10, # L1 regularization term on weights
# 'reg_lambda': 0, # L2 regularization term on weights
'nthread': 4,
'verbose': 0
}
print('>> prepare dataset...')
dtrain_lgb = lgb.Dataset(train_df_array, label=train_df_labels,
feature_name=predictors,
categorical_feature=categorical)
del train_df_array, train_df_labels; gc.collect()
print_memory()
print(params)
print('>> start cv...')
cv_results = lgb.cv(params,
dtrain_lgb,
categorical_feature = categorical,
num_boost_round=20000,
metrics='rmse',
seed = SEED,
shuffle = False,
nfold=10,
show_stdv=True,
early_stopping_rounds=100,
verbose_eval=50)
print('[{}]: model training time'.format(time.time() - start_time))
print_memory()
# print (cv_results)
print('--------------------------------------------------------------------')
num_boost_rounds_lgb = len(cv_results['rmse-mean'])
print('num_boost_rounds_lgb=' + str(num_boost_rounds_lgb))
print ('>> start trainning... ')
model_lgb = lgb.train(
params, dtrain_lgb,
num_boost_round=num_boost_rounds_lgb,
feature_name = predictors,
categorical_feature = categorical)
del dtrain_lgb
gc.collect()
print('--------------------------------------------------------------------')
print('>> save model...')
# save model to file
# if not DEBUG:
model_lgb.save_model(modelfilename+'.txt')