def feat_importance_ip(row_id_str, ds_id, hdfs_feat_dir, local_score_file,
score_file_IT, sp_master, spark_rdd_compress,
spark_driver_maxResultSize, sp_exe_memory, sp_core_max,
zipout_dir, zipcode_dir, zip_file_name, mongo_tuples,
jobname, uploadtype):
# zip func in other files for Spark workers ================= ================
zip_file_path = ml_util.ml_build_zip_file(zipout_dir,
zipcode_dir,
zip_file_name,
prefix='zip_feature_util')
print "INFO: zip_file_path=", zip_file_path
# get_spark_context
sc = ml_util.ml_get_spark_context(sp_master, spark_rdd_compress,
spark_driver_maxResultSize,
sp_exe_memory, sp_core_max, jobname,
[zip_file_path])
'''
SparkContext.setSystemProperty('spark.rdd.compress', config.get('spark', 'spark_rdd_compress'))
SparkContext.setSystemProperty('spark.driver.maxResultSize', config.get('spark', 'spark_driver_maxResultSize'))
#SparkContext.setSystemProperty('spark.kryoserializer.buffer.mb', config.get('spark', 'spark_kryoserializer_buffer_mb'))
SparkContext.setSystemProperty('spark.executor.memory', args.exe_memory)
SparkContext.setSystemProperty('spark.cores.max', args.core_max)
sc = SparkContext(args.sp_master, 'feature_importance_2ways:'+str(args.row_id))
'''
t0 = time()
# get folder list (labels) from hdfs data_out/<id>/metadata ==============
dirFile_loc = os.path.join(hdfs_feat_dir, "metadata")
dirFolders = sc.textFile(dirFile_loc)
hash_Folders = dirFolders.collect()
print "INFO: dirFile_loc=", dirFile_loc, ", hash_Folders=", hash_Folders
folder_list = [x.encode('UTF8') for x in hash_Folders]
print "INFO: folder_list=", folder_list
# get feature seq : ngram hash mapping ==================================
key = "dic_seq_hashes" #{"123":"136,345"}
jstr_filter = '{"rid":' + row_id_str + ',"key":"' + key + '"}'
jstr_proj = '{"value":1}'
# get parent dataset's data
if ds_id != row_id_str:
jstr_filter = '{"rid":' + ds_id + ',"key":"' + key + '"}'
doc = query_mongo.find_one_t(mongo_tuples, jstr_filter, jstr_proj)
dic_list = doc['value']
dic_all_columns = dic_list
feature_count = len(dic_list)
#print "INFO: feature_count=",feature_count
#print "dic_list=",dic_list #{u'123,345':u'136'}
#print "INFO: dic_all_columns=",dic_all_columns # {1: u'8215,8216'}
# end
# get hash : raw string mapping ==================================
key = "dic_hash_str" #{"123":"openFile"}
jstr_filter = '{"rid":' + row_id_str + ',"key":"' + key + '"}'
jstr_proj = '{"value":1}'
# get parent dataset's data
if ds_id != row_id_str:
jstr_filter = '{"rid":' + ds_id + ',"key":"' + key + '"}'
doc = query_mongo.find_one_t(mongo_tuples, jstr_filter, jstr_proj)
dic_hash_str = doc['value']
'''
# get folder list (labels) from hdfs data_out/<id>/libsvm ==============
libsvm_loc = os.path.join(hdfs_feat_dir , "libsvm_data")
# based on label, divide RDD into arrays
f_rdd = sc.textFile(libsvm_loc).map(lambda x: libsvm2tuple_arr(x))
arr_libsvm=sorted(f_rdd.collect(), key=lambda x:x[0]) # sorted by label
'''
# filename for featured data
libsvm_data_file = os.path.join(hdfs_feat_dir, "libsvm_data")
print "INFO: libsvm_data_file=", libsvm_data_file
print "INFO: feature_count=",feature_count\
# get sample array from hdfs
arr_libsvm = zip_feature_util.get_sample_tuple_arr(sc, libsvm_data_file)
# sorted by label
arr_libsvm = sorted(arr_libsvm, key=lambda x: x[0])
# convert libsvm to features_list, row_list, col_list, sample count, col_num
lbl_flag = -1
row_num_training = 0
sparse_mtx_list = [] # for feat impor calculation
features_list = [] # for csc_matrix
row_list = [] # for csc_matrix
col_list = [] # for csc_matrix
sample_numbers = [] # for csc_matrix
feature_arr = None
for idx, i in enumerate(arr_libsvm):
#print "idx=",idx,",l=",i[0],",d=",i[1:]
if lbl_flag != i[0]:
if feature_arr and len(feature_arr) > 0:
features_list.append(np.array(feature_arr))
row_list.append(np.array(row_arr))
col_list.append(np.array(col_arr))
sample_numbers.append(cnt)
row_arr = []
col_arr = []
feature_arr = []
cnt = 0
lbl_flag += 1
for j in i[1:]:
row_arr.append(cnt)
col_arr.append(j[0] - 1)
feature_arr.append(j[1])
cnt += 1
# for last part
if len(feature_arr) > 0:
features_list.append(np.array(feature_arr))
row_list.append(np.array(row_arr))
col_list.append(np.array(col_arr))
sample_numbers.append(cnt)
#print ",features_list=",features_list
#print ",row_list=",row_list
#print ",col_list=",col_list
print "INFO: sample_numbers=", sample_numbers
col_num = len(dic_list)
print "INFO: column number: ", col_num #, ",len(max_feat_list)=",len(max_feat_list)
for i in range(0, len(features_list)):
#print "i=",i
#print "features_list=",features_list[i]
#print "row_list=",row_list[i]
#print "col_list=",col_list[i]
#print "sample_numbers=",sample_numbers[i]
sparse_mtx = csc_matrix((features_list[i], (row_list[i], col_list[i])),
shape=(sample_numbers[i], col_num))
sparse_mtx_list.append(sparse_mtx)
#print sparse_mtx_list[0]
print "INFO: sparse_mtx_list[0].shape=", sparse_mtx_list[0].shape
#print sparse_mtx_list[1]
print "INFO: sparse_mtx_list[1].shape=", sparse_mtx_list[1].shape
exclusive_feature_set_mal = []
exclusive_feature_set_clean = []
dic_feature_cnt_mal = {}
dic_feature_cnt_clean = {}
dic_score = {}
dic_cnt_mal = {}
dic_cnt_clean = {}
dic_IT_grain = {}
####################################################
####feature importance algorithms: 2 methods ####### # Only for 2 classes ???
####################################################
if len(sample_numbers) == 2:
###################################################
################## calculate probability ############
###################################################
print "INFO: =======Feature Importance(probability) ================ "
for j in range(0, col_num):
curr_col_dirty = sparse_mtx_list[0].getcol(j)
sum_col = curr_col_dirty.sum(0)
cnt_mal = sum_col.tolist()[0][0]
curr_col_clean = sparse_mtx_list[1].getcol(j)
sum_col = curr_col_clean.sum(0)
cnt_clean = sum_col.tolist()[0][0]
percnt_mal = cnt_mal / float(sample_numbers[0])
percnt_clean = cnt_clean / float(sample_numbers[1])
score_j = (percnt_mal + 1 - percnt_clean) / 2
dic_score[j + 1] = score_j
dic_cnt_clean[j + 1] = cnt_clean
dic_cnt_mal[j + 1] = cnt_mal
sorted_score = sorted(dic_score.items(),
key=operator.itemgetter(1),
reverse=True)
#print "sorted_score:", sorted_score
#print "dic_cnt_clean", dic_cnt_clean
#print "dic_cnt_mal", dic_cnt_mal
############output result########################
if os.path.exists(local_score_file):
try:
os.remove(local_score_file)
except OSError, e:
print("Error: %s - %s." % (e.local_score_file, e.strerror))
for ii in range(0, len(sorted_score)):
(feat, score) = sorted_score[ii]
#print feat, score, dic_all_columns[feat]
if dic_hash_str:
description_str = feats2strs(dic_all_columns[str(feat)],
dic_hash_str)
else:
description_str = "N/A"
print "Warning: No mapping found for feature number"
str01 = str(feat) + "\t" + str(
score) + "\t" + description_str + "\n"
with open(local_score_file, "a") as f:
f.write(str01)
########################################################
##################Information Gain (entropy)############
########################################################
print "INFO: =======Information Gain================ "
for j in range(0, col_num):
cnt_mal = dic_cnt_mal[j + 1]
cnt_clean = dic_cnt_clean[j + 1]
total_samples = sample_numbers[0] + sample_numbers[1]
p0 = float(sample_numbers[0]) / total_samples
p1 = 1 - p0
if p0 == 0 or p1 == 0:
parent_entropy = 0
else:
parent_entropy = 0 - p0 * np.log2(p0) - p1 * np.log2(p1)
if cnt_clean + cnt_mal == 0:
information_gain = 0
elif total_samples - cnt_clean - cnt_mal == 0:
information_gain = 0
else:
p0 = float(cnt_mal) / (cnt_clean + cnt_mal)
p1 = 1 - p0
if p0 == 0 or p1 == 0:
child_left_entropy = 0
else:
child_left_entropy = 0 - p0 * np.log2(p0) - p1 * np.log2(
p1)
p0 = float(sample_numbers[0] - cnt_mal) / (total_samples -
cnt_clean - cnt_mal)
p1 = 1 - p0
if p0 == 0 or p1 == 0:
child_right_entropy = 0
else:
child_right_entropy = 0 - p0 * np.log2(p0) - p1 * np.log2(
p1)
weighted_child_entropy = child_left_entropy * float(
cnt_clean +
cnt_mal) / total_samples + child_right_entropy * float(
total_samples - cnt_clean - cnt_mal) / total_samples
information_gain = parent_entropy - weighted_child_entropy
dic_IT_grain[j + 1] = information_gain
sorted_IT_gain = sorted(dic_IT_grain.items(),
key=operator.itemgetter(1),
reverse=True)
if os.path.exists(score_file_IT):
try:
os.remove(score_file_IT)
except OSError, e:
print("Error: %s - %s." % (e.score_file_IT, e.strerror))
def feat_importance_comb(row_id_str, ds_id, num_to_show, w_FIRM, w_IT, w_Prob,
mongo_tuples, FIRM_score_file, IT_score_file,
Prob_score_file, score_file_combine):
human_verified = dict()
all_verified = dict()
print "INFO: ======= Combine all feature importance info ================"
# get feature importance voting data from db
all_verified, human_verified = exec_sqlite.get_dict(row_id_str)
#print "INFO: human_verified dict=",human_verified
#print "INFO: all_verified dict=",all_verified
############begin##################
with open(FIRM_score_file, 'r') as f:
FIRM_score = f.readlines()
with open(IT_score_file, 'r') as f:
IT_score = f.readlines()
with open(Prob_score_file, 'r') as f:
Prob_score = f.readlines()
# create file for one table here ======================
dir_name = os.path.dirname(FIRM_score_file)
coef_filename = os.path.join(dir_name,
row_id_str + '_score_coef_comb.json')
#print "INFO: combined fname=",coef_filename
# get data from mongo
key = "coef_arr"
jstr_filter = '{"rid":' + row_id_str + ',"key":"' + key + '"}'
jstr_proj = '{"value":1}'
# get parent dataset's data
if ds_id != row_id_str:
jstr_filter = '{"rid":' + ds_id + ',"key":"' + key + '"}'
doc = query_mongo.find_one_t(mongo_tuples, jstr_filter, jstr_proj)
coef_arr = doc['value']
#print "INFO: len(coef_arr)=",len(coef_arr)
# get sample count ===========
key = "feat_sample_count_arr"
jstr_filter = '{"rid":' + row_id_str + ',"key":"' + key + '"}'
jstr_proj = '{"value":1}'
# get parent dataset's data
# need to chk if new feature
if ds_id != row_id_str:
jstr_filter = '{"rid":' + ds_id + ',"key":"' + key + '"}'
doc = query_mongo.find_one_t(mongo_tuples, jstr_filter, jstr_proj)
feat_sample_count_arr = None
if not doc is None:
feat_sample_count_arr = doc['value']
combine_with_coef(row_id_str, coef_arr, FIRM_score, IT_score, Prob_score,
coef_filename, feat_sample_count_arr)
featurelist_FIRM = []
featurelist_IT = []
featurelist_Prob = []
dic_all_columns = {}
print "INFO: num to_show=", num_to_show
for i in range(0, num_to_show):
####FIRM####
str_in = FIRM_score[i]
feature_id, score, descpt = str_in.split('\t', 3)
featurelist_FIRM.append(feature_id)
if not feature_id in dic_all_columns:
dic_all_columns[feature_id] = descpt
####IT####
str_in = IT_score[i]
feature_id, score, descpt = str_in.split('\t', 3)
featurelist_IT.append(feature_id)
if not feature_id in dic_all_columns:
dic_all_columns[feature_id] = descpt
####Prob####
str_in = Prob_score[i]
feature_id, score, descpt = str_in.split('\t', 3)
featurelist_Prob.append(feature_id)
if not feature_id in dic_all_columns:
dic_all_columns[feature_id] = descpt
list_i = [i + 1 for i in range(0, num_to_show)]
zipped_FIRM = zip(featurelist_FIRM, list_i)
zipped_IT = zip(featurelist_IT, list_i)
zipped_Prob = zip(featurelist_Prob, list_i)
FIRM_dict = dict(zipped_FIRM)
IT_dict = dict(zipped_IT)
Prob_dict = dict(zipped_Prob)
list_combine = featurelist_FIRM + featurelist_IT + featurelist_Prob
list_unique = OrderedDict.fromkeys(list_combine).keys()
#print list_unique, len(list_unique)
#human_verified = {'188':7, '218':6} #####get human_verified from database, all click_number > 5 are human_verified###
score_combine = {}
for i in range(0, len(list_unique)):
feat_id = list_unique[i]
if feat_id in human_verified:
print "INFO: found feat_id=", feat_id
continue
score = 0
if feat_id in FIRM_dict:
score = score + w_FIRM * FIRM_dict[feat_id]
else:
score = score + w_FIRM * (num_to_show + 1)
if feat_id in IT_dict:
score = score + w_IT * IT_dict[feat_id]
else:
score = score + w_IT * (num_to_show + 1)
if feat_id in Prob_dict:
score = score + w_Prob * Prob_dict[feat_id]
else:
score = score + w_Prob * (num_to_show + 1)
score = score / float(3)
#print feat_id, score
#############add human feedback##########
if feat_id in all_verified:
click_number = all_verified[
feat_id] #####get click number from database###
else:
click_number = 0
#print feat_id, click_number, score
#click_number = 3
if click_number == 1:
score = score - 2
elif click_number == 2:
score = score - 4
elif click_number == 3:
score = score - 10
elif click_number == 4:
score = score - 20
elif click_number == 5:
score = 0
if score < 0:
score = 0
#print "***=",feat_id, click_number, score
score_combine[feat_id] = score
if os.path.exists(score_file_combine):
try:
os.remove(score_file_combine)
except OSError, e:
print("Error: %s - %s." % (e.score_file_combine, e.strerror))
def train(row_id_str, ds_id, hdfs_feat_dir, local_out_dir, ml_opts_jstr
, sp_master, spark_rdd_compress, spark_driver_maxResultSize, sp_exe_memory, sp_core_max
, zipout_dir, zipcode_dir, zip_file_name
, mongo_tuples, labelnameflag, fromweb, src_filename
, jobname ):
# create zip files for Spark workers ================= ================
zip_file_path = ml_build_zip_file(zipout_dir, zipcode_dir, zip_file_name, prefix='zip_feature_util')
print "INFO: zip_file_path=",zip_file_path
#data_folder = hdfs_feat_dir + "/"
#local_out_dir = local_out_dir + "/"
#if os.path.exists(local_out_dir):
# shutil.rmtree(local_out_dir) # to keep smaplelist file
if not os.path.exists(local_out_dir):
os.makedirs(local_out_dir)
# init Spark context ====
sc=ml_util.ml_get_spark_context(sp_master
, spark_rdd_compress
, spark_driver_maxResultSize
, sp_exe_memory
, sp_core_max
, jobname
, [zip_file_path])
# start here =================================================================== ===============
t0 = time()
### Need to check if PCA available here ===========================
libsvm_data_file = os.path.join(hdfs_feat_dir , src_filename) # need to set k numb in filename somehow
print "INFO: libsvm_data_file=", libsvm_data_file
#samples_rdd = MLUtils.loadLibSVMFile(sc, libsvm_data_file).cache()
# load sample RDD from text file
# format (LabeledPoint,hash) from str2LabeledPoint_hash()
feature_count=0
samples_rdd, feature_count = zip_feature_util.get_sample_rdd(sc, libsvm_data_file, feature_count, '')
# get label as a list
labels_list_all = samples_rdd.map(lambda p: int(p[0].label)).collect()
total_sample_count=len(labels_list_all)
parsedData =samples_rdd.map(lambda p: p[0].features).cache()
#for i in parsedData.collect(): #p.features: pyspark.mllib.linalg.SparseVector
# print "pd=",type(i),",i=",i
t1 = time()
print 'INFO: running time: %f' %(t1-t0)
t0 = t1
###############################################
########## build learning model ###############
###############################################
### get the parameters###
print "INFO: ============Learning Algorithm and Parameters============="
para_dict = json.loads(ml_opts_jstr)
flag_model = para_dict['learning_algorithm'] # kmeans
iteration_num = eval(para_dict['iterations'])
k=2
if 'k' in para_dict:
k = eval(para_dict['k'])
print "INFO: Learning Algorithm:", flag_model
print "INFO: iterations=", iteration_num
#print "training_sample_number=", training_sample_number
### generate label names (family names) #####
### connect to database to get the column list which contains all column number of the corresponding feature####
if labelnameflag == 1:
key = "dic_name_label"
jstr_filter='{"rid":'+row_id_str+',"key":"'+key+'"}'
jstr_proj='{"value":1}'
# get parent dataset's data
if ds_id != row_id_str:
jstr_filter='{"rid":'+ds_id+',"key":"'+key+'"}'
doc=query_mongo.find_one_t(mongo_tuples, jstr_filter, jstr_proj)
dic_list = doc['value']
label_dic = {}
for i in range(0, len(dic_list)):
for key in dic_list[i]:
label_dic[dic_list[i][key]] = key.encode('UTF8')
print "INFO: label_dic:", label_dic
else:
label_dic = {}
label_set = set(labels_list_all)
for label_value in label_set:
label_dic[int(label_value)] = str(int(label_value))
print "INFO: generated label_dic:", label_dic
labels_list = []
for key in sorted(label_dic):
labels_list.append(label_dic[key])
print "INFO: labels_list=", labels_list
### build model ###
if flag_model == "kmeans":
print "=================== Kmeans ============"
model = KMeans.train(parsedData, k, maxIterations=iteration_num)
t_cost= model.computeCost(parsedData)
print "INFO: cost for training set =", str(t_cost)
clusterCenters=model.clusterCenters
print "INFO: clusterCenters t=", type(clusterCenters) #list
elif flag_model == "gaussian_mixture_model": # didn't work some native lib issue
print "=================== Gaussian_Mixture_Model ============"
model = GaussianMixture.train(parsedData, k, maxIterations=iteration_num)
print "INFO: model.weights =", model.weights
else:
print "INFO: Training model selection error: no valid ML model selected!"
return
### Save model
save_dir = config.get('app', 'HADOOP_MASTER')+config.get('app', 'HDFS_MODEL_DIR')+'/'+row_id_str
try:
hdfs.ls(save_dir)
#print "find hdfs folder"
hdfs.rmr(save_dir)
#print "all files removed"
except IOError as e:
print "ERROR: I/O error({0}): {1}".format(e.errno, e.strerror)
except:
print "ERROR: Unexpected error:", sys.exc_info()[0]
print "INFO: model saved at hdfs=",save_dir
print "INFO: model type=",type(model)," model=",model
model.save(sc, save_dir)
###load model if needed
#sameModel = SVMModel.load(sc, save_dir)
###
# (true label, keams label, features list, hash)
all_data=samples_rdd.map(lambda t: ( t[0].label, model.predict(t[0].features), t[0].features, t[1] ) ).collect()
true_label_arr = np.asarray([int(x) for x,_,_,_ in all_data])
labels_kmeans = np.asarray([int(x) for _,x,_,_ in all_data])
hash_list = np.asarray([x for _,_,_,x in all_data])
print "INFO: all_data len=",len(all_data),"all_data t=",type(labels_list_all)
print "INFO: true_label_arr.shape=",true_label_arr.shape,"labels_kmeans.shape=",labels_kmeans.shape
print "INFO: true_label_arr t=",type(true_label_arr),"labels_kmeans t=",type(labels_kmeans)
mtx_center=np.asarray(clusterCenters)
features_array_reduced=np.asarray([x.toArray() for _,_,x,_ in all_data])
print "INFO: mtx_center t=",type(mtx_center),"mtx_center.shape=",mtx_center.shape
print "INFO: features_array_reduced t=",type(features_array_reduced),"features_array_reduced.shape",features_array_reduced.shape
#Adjusted Mutual Information between two clusterings
amis=adjusted_mutual_info_score(labels_list_all,labels_kmeans)
print "INFO: Adjusted_mutual_info_score=", amis
#Similarity measure between two clusterings
ars=adjusted_rand_score(labels_list_all,labels_kmeans)
print "INFO: Adjusted_rand_score=", ars
accuracy=0.0
t1 = time()
print 'INFO: training run time: %f' %(t1-t0)
t0 = t1
###############################################
########## plot histogram ######
###############################################
n_clusters=k
plot_col_num = int(math.ceil(math.sqrt(n_clusters)))
figsize = (4*plot_col_num, 3*int(math.ceil(n_clusters*1.0/plot_col_num)))
print "INFO: n_clusters=",n_clusters,",label_dic=",label_dic
print "INFO: plot_col_num=",plot_col_num,",figsize=",figsize,",local_out_dir=",local_out_dir
# kmeans histogram
_, p_true = ml_plot_kmeans_histogram_subfigures(true_label_arr, labels_kmeans, n_clusters, names = label_dic
, plot_col_num = plot_col_num, figsize=figsize, folder = local_out_dir, rid=row_id_str)
# normalized kmeans histogram
_, p_true_norm = ml_plot_kmeans_histogram_subfigures(true_label_arr, labels_kmeans, n_clusters, names = label_dic
, plot_col_num = plot_col_num, figsize=figsize, normalize = True, folder = local_out_dir, rid=row_id_str)
####plot "reverse" histogram with labels ####
num_bars = max(true_label_arr) + 1
figsize = (4*plot_col_num, 3*int(math.ceil(num_bars*1.0/plot_col_num)))
_, p_cluster = ml_plot_kmeans_histogram_subfigures(labels_kmeans, true_label_arr, num_bars, names = label_dic
, plot_col_num = plot_col_num, figsize=figsize, reverse = True, folder = local_out_dir, rid=row_id_str)
#### plot dot figures ####
# dot plot for Kmeans ===========
filename=os.path.join(local_out_dir ,row_id_str+'_cluster.png')
filename_3d=os.path.join(local_out_dir ,row_id_str+'_cluster_3d.json')
ml_plot_kmeans_dot_graph_save_file(features_array_reduced, labels_kmeans, mtx_center, n_clusters, figsize=(10,7), filename=filename
, title='KMeans', filename_3d=filename_3d)
# dot plot for True Labels ===========
filename=os.path.join(local_out_dir ,row_id_str+'_cluster_tl.png')
filename_3d=os.path.join(local_out_dir ,row_id_str+'_cluster_3d_tl.json')
ml_plot_kmeans_dot_graph_save_file(features_array_reduced, true_label_arr, mtx_center, n_clusters, figsize=(10,7), filename=filename
, title='True Labels', filename_3d=filename_3d)
dataset_info={"training_fraction":1, "class_count":n_clusters,"dataset_count":total_sample_count}
# only update db for web request
if fromweb=="1":
#print "database update"
str_sql="UPDATE atdml_document set "+"accuracy = '" \
+"', status = 'learned', processed_date ='"+str(datetime.datetime.now()) \
+"', total_feature_numb='"+str(feature_count) \
+"', perf_measures='{}" \
+"', dataset_info='"+json.dumps(dataset_info) \
+"' where id="+row_id_str
ret=exec_sqlite.exec_sql(str_sql)
print "INFO: Data update done! ret=", str(ret)
else:
print "INFO: accuracy = '"+str(accuracy*100)+"%"
print 'INFO: Finished!'
return 0
def train(row_id_str, ds_id, hdfs_feat_dir, local_out_dir, ml_opts_jstr, excluded_feat_cslist
, sp_master, spark_rdd_compress, spark_driver_maxResultSize, sp_exe_memory, sp_core_max
, zipout_dir, zipcode_dir, zip_file_name
, mongo_tuples, labelnameflag, fromweb
, training_fraction, jobname, model_data_folder ):
# zip func in other files for Spark workers ================= ================
zip_file_path = ml_build_zip_file(zipout_dir, zipcode_dir, zip_file_name, prefix='zip_feature_util')
print "INFO: zip_file_path=",zip_file_path
# ML model filename ====
model_fname=os.path.join(model_data_folder, row_id_str+'.pkl')
print "INFO: model_data_folder=",model_data_folder
# create out folders and clean up old model files ====
ml_util.ml_prepare_output_dirs(row_id_str,local_out_dir,model_data_folder,model_fname)
# init Spark context ====
sc=ml_util.ml_get_spark_context(sp_master
, spark_rdd_compress
, spark_driver_maxResultSize
, sp_exe_memory
, sp_core_max
, jobname
, [zip_file_path])
t0 = time()
t00 = t0
# check if ml_opts.has_excluded_feat ==1 ===================================
has_excluded_feat=0
if not ml_opts_jstr is None:
ml_opts=json.loads(ml_opts_jstr)
if "has_excluded_feat" in ml_opts:
has_excluded_feat=ml_opts["has_excluded_feat"]
# get excluded feature list from mongo ========== ===
if str(has_excluded_feat) == "1" and excluded_feat_cslist is None:
excluded_feat_cslist=ml_util.ml_get_excluded_feat(row_id_str, mongo_tuples)
print "INFO: excluded_feat_cslist=",excluded_feat_cslist
# source libsvm filename
libsvm_data_file = os.path.join(hdfs_feat_dir , "libsvm_data")
print "INFO: libsvm_data_file=", libsvm_data_file
# load feature count file
feat_count_file=libsvm_data_file+"_feat_count"
feature_count=zip_feature_util.get_feature_count(sc,feat_count_file)
print "INFO: feature_count=",feature_count
# load sample RDD from text file
# also exclude selected features in sample ================ =====
# format (LabeledPoint,hash) from str2LabeledPoint_hash()
#samples_rdd = MLUtils.loadLibSVMFile(sc, libsvm_data_file)
samples_rdd,feature_count = zip_feature_util.get_sample_rdd(sc, libsvm_data_file, feature_count, excluded_feat_cslist)
all_data = samples_rdd.collect()
sample_count=len(all_data)
# 2-D array
features_list = [x.features.toArray() for x,_ in all_data]
# label array
labels_list_all = [x.label for x,_ in all_data]
# hash array
hash_list_all = [x for _,x in all_data]
# convert to np array
labels_list_all = array(labels_list_all)
features_array = np.array(features_list)
hash_list_all=np.array(hash_list_all)
# generate sparse matrix (csr) for all samples
features_sparse_mtx = csr_matrix(features_array)
### randomly split the samples into training and testing data ===============
X_train_sparse, X_test_sparse, labels_train, labels_test, train_hash_list, test_hash_list = \
cross_validation.train_test_split(features_sparse_mtx, labels_list_all, hash_list_all, test_size=(1-training_fraction) )
# X_test_sparse is scipy.sparse.csr.csr_matrix
testing_sample_count = len(labels_test)
training_sample_count=len(labels_train)
training_lbl_cnt_list=Counter(labels_train)
testing_lbl_cnt_list=Counter(labels_test)
print "INFO: training sample count=",training_sample_count,", testing sample count=",testing_sample_count,",sample_count=",sample_count
print "INFO: training label list=",training_lbl_cnt_list,", testing label list=",testing_lbl_cnt_list
print "INFO: train_hash_list count=",len(train_hash_list),", test_hash_list count=",len(test_hash_list)
t1 = time()
print 'INFO: running time: %f' %(t1-t0)
###############################################
###########build learning model################
###############################################
### parse parameters and generate the model ###
(clf, model_name, api, cv, param_dic) = parse_param_and_get_model(ml_opts)
if model_name == "none":
print "ERROR: model name not found!"
return -1
#param_jobj=json.loads(ml_opts_jstr);
#print "param_jobj=",param_jobj
########################################################
##########Grid Search with cross validation#############
########################################################
json2save={}
json2save["rid"]=int(row_id_str)
json2save["key"]="cv_result"
#json2save["param_str"]=ml_opts_jstr
json2save["param_dic"]=param_dic
cv_grid=[]
if api == "centralized":
#########run with Scikit-learn API (for comparison)######
print "INFO: ******************Grid Search with Scikit-learn API************"
t0 = time()
# Set the parameters by cross-validation
#tuned_parameters = [{'C': [0.0001, 0.001, 0.01, 0.1, 1, 10, 100, 1000, 10000, 100000]}]
#tuned_parameters = [{'kernel': ['rbf'], 'gamma': [1e-3, 1e-4], \
# 'C': [1, 10, 100, 1000]}, \
# {'kernel': ['linear'], 'C': [1, 10, 100, 1000]}]
scores = ['accuracy']
json2save["scores"]=scores
#print json2save
for score in scores: # for one item only? score=accuracy
print("INFO: # Tuning hyper-parameters for %s" % score)
#print()
grid = grid_search.GridSearchCV(estimator = clf, param_grid = param_dic, cv=cv, scoring= score)
grid.fit(X_train_sparse, labels_train)
print "INFO: Best parameters set found on development set:"
print "INFO: grid.best_params_=",grid.best_params_
print "INFO: Grid scores on development set:"
for key in grid.best_params_:
print "INFO: best_params["+key+"]=", grid.best_params_[key]
if key.lower()=="regtype":
ml_opts['regularization']=str(grid.best_params_[key]) # add best param to
else:
ml_opts[key.lower()]=str(grid.best_params_[key]) # add best param to
# save best param to db as json string
j_str=json.dumps(ml_opts);
json2save["param_str"]=j_str;
print "INFO: grid_scores_ with params:"
for params, mean_score, scores in grid.grid_scores_:
print "INFO: %0.3f (+/-%0.03f) for %r" % (mean_score, scores.std() * 2, params)
#outstr='%s,%0.3f,%0.03f,%s' % (params,mean_score, scores.std() * 2,"Selected" if params==grid.best_params_ else "")
outj={}
outj["param"]=params
outj["average_accuracy"]="%0.3f" % (mean_score)
outj["std_deviation"]="%0.3f" % (scores.std() * 2)
outj["selected"]="%s" % ("Selected" if params==grid.best_params_ else "")
cv_grid.append(outj)
clf_best = grid.best_estimator_
t1 = time()
############# END run with SKlearn ######
print 'INFO: Grid Search with SKlearn running time: %f' %(t1-t0)
t0 = time()
else:
#############run with SPARK######
print "INFO: ******************Grid Search with SPARK************"
all_comb_list_of_dic = get_all_combination_list_of_dic(param_dic)
print "INFO: Total number of searching combinations=", len(all_comb_list_of_dic)
#print "all_comb_list_of_dic: ", all_comb_list_of_dic
params_rdd = sc.parallelize(all_comb_list_of_dic)
###broad cast clf, traning data, testing data to all workers###
X_broadcast = sc.broadcast(X_train_sparse)
y_broadcast = sc.broadcast(labels_train)
clf_broadcast = sc.broadcast(clf)
### Grid Search with CV in multiple workers ###
models = params_rdd.map(lambda x: learn_with_params(clf_broadcast.value, X_broadcast.value, y_broadcast.value, cv, x)).sortByKey(ascending = False).cache()
(ave_accuracy, (clf_best, p_dic_best, std2)) = models.first()
# output results #
print "INFO: Best parameters set found for ", model_name, " is: "
print "INFO: ",
for key in p_dic_best:
print key, " = ", p_dic_best[key],
if key.lower()=="regtype":
ml_opts['regularization']=str(p_dic_best[key])
else:
ml_opts[key.lower()]=str(p_dic_best[key]) # add best param to
# save best param to db as json string
print ""
j_str=json.dumps(ml_opts);
json2save["param_str"]=j_str;
print "INFO: Average accuracy with CV = ", cv, ": ", ave_accuracy
######## print complete report #######
print "INFO: Grid scores on development set:"
all_results = models.collect()
for i in range(0, len(all_results)):
(ave_accu_i, (clf_i, p_dic_i, std2_i)) = all_results[i]
print "INFO: ",ave_accu_i, " for ", p_dic_i
print "INFO: %0.3f (+/-%0.03f) for " % (ave_accu_i, std2_i), p_dic_i
#outstr='%s,%0.3f,%0.03f,%s' % ( p_dic_i, ave_accu_i, std2_i, "Selected" if p_dic_i==p_dic_best else "")
outj={}
outj["param"]=p_dic_i
outj["average_accuracy"]="%0.3f" % (ave_accu_i)
outj["std_deviation"]="%0.3f" % (std2_i)
outj["selected"]="%s" % ("Selected" if p_dic_i==p_dic_best else "")
cv_grid.append(outj)
print " "
t1 = time()
############# END run with SPARK######
print 'INFO: Grid search with SPARK running time: %f' %(t1-t0)
##################################################################################
#print "cv_grid=",cv_grid
#json2save["cv_grid_title"]='param,average_accuracy,std_deviation,selected'
json2save["cv_grid_data"]=cv_grid
json2save['clf_best']=str(clf_best).replace("\n","").replace(" ","")
cv_result=json.dumps(json2save)
#print "INFO: cv_result=",cv_result
filter='{"rid":'+row_id_str+',"key":"cv_result"}'
upsert_flag=True
## write to mongoDB.myml.dataset_info, ignore doc with duplicated key
# db.dataset_info.createIndex({"rid":1,"key":1},{unique:true})
ret=query_mongo.upsert_doc_t(mongo_tuples,filter,cv_result,upsert_flag)
print "INFO: Upsert count for cv_result: ret=",ret
##################################################################################
##########Retrain with best model for training set and output results#############
##################################################################################
print "INFO: **********Retrain with best model for training set and output results************"
clf_best.fit(X_train_sparse, labels_train)
#### save clf_best for future use ####
#joblib.dump(clf_best, model_data_folder + row_id_str+'.pkl')
joblib.dump(clf_best, model_fname)
### Evaluating the model on testing data
labels_pred = clf_best.predict(X_test_sparse)
accuracy = clf_best.score(X_test_sparse, labels_test)
print "INFO: Accuracy = ", accuracy
######################################the rest of the code is the same as train_sklean.py (replace clf with clf_best)#####################################################################
clf=clf_best
print "INFO: model type=",type(clf)," clf=",clf
# get data from model ================================
coef=None
intercept=None
try:
if type(clf) in ( classes.SVC , classes.NuSVC) :# svm didn't have coef_
col_num=clf.support_vectors_.shape[1]
else: #linear only
# coef_ is only available when using a linear kernel
col_num = len(clf.coef_[0])
coef=clf.coef_[0]
intercept=clf.intercept_[0] # only get 1st item?
#print "**model:clf.coef_[0] =",clf.coef_[0]
except Exception as e:
print "WARNING: Can't get clf.coef_[0]. e=",e,", get total features from meta-data"
col_num = 0 #how to get feature number for sparse array?
print "INFO: total feature # in the model: ", col_num
jfeat_coef_dict={}
# create feature coefficient file ================================
if coef is None:
print "WARNING: model weights not found!"
else:
feat_filename=os.path.join(local_out_dir,row_id_str+"_feat_coef.json")
print "INFO: feat_filename=",feat_filename
# save coef_arr to mongo & create jfeat_coef_dict===
jfeat_coef_dict=ml_util.ml_save_coef_build_feat_coef(row_id_str, mongo_tuples, coef, intercept, feat_filename, ds_id)
#print "INFO: jfeat_coef_dict=", jfeat_coef_dict
print "INFO: jfeat_coef_dict len=", len(jfeat_coef_dict )
# filename for false pred
false_pred_fname=os.path.join(local_out_dir,row_id_str+"_false_pred.json")
print "INFO: false_pred_fname=", false_pred_fname
# build files for false pred & score graph
(score_arr_0, score_arr_1, max_score,min_score)=ml_build_false_pred(X_test_sparse,coef,intercept
, labels_test, labels_pred, test_hash_list, model_name, jfeat_coef_dict, false_pred_fname)
# save pred output
pred_out_arr=[]
for i in range(0,len(labels_test)):
pred_out_arr.append((labels_test[i], labels_pred[i], test_hash_list[i]))
pred_ofname=os.path.join(local_out_dir,row_id_str+"_pred_output.pkl")
print "INFO: pred_ofname=", pred_ofname
ml_util.ml_pickle_save(pred_out_arr,pred_ofname)
###################################################
### generate label names (family names) ###########
### connect to database to get the column list which contains all column number of the corresponding feature####
###################################################
if labelnameflag == 1:
key = "dic_name_label"
jstr_filter='{"rid":'+row_id_str+',"key":"'+key+'"}'
jstr_proj='{"value":1}'
# get parent dataset's data
if ds_id != row_id_str:
jstr_filter='{"rid":'+ds_id+',"key":"'+key+'"}'
doc=query_mongo.find_one_t(mongo_tuples, jstr_filter, jstr_proj)
dic_list = doc['value']
label_dic = {}
for i in range(0, len(dic_list)):
for key in dic_list[i]:
label_dic[dic_list[i][key]] = key.encode('UTF8')
print "INFO: label_dic:", label_dic
else:
label_dic = {}
label_set = set(labels_list_all)
for label_value in label_set:
label_dic[int(label_value)] = str(int(label_value))
print "INFO: ******generated label_dic:", label_dic
labels_list = []
for key in sorted(label_dic):
labels_list.append(label_dic[key])
### generate sample numbers of each family in testing data###
testing_sample_number = len(labels_test)
print "INFO: testing_sample_number=", testing_sample_number
test_cnt_dic = {}
for key in label_dic:
test_cnt_dic[key] = 0
for i in range (0, testing_sample_number):
for key in label_dic:
if labels_test[i] == key:
test_cnt_dic[key] = test_cnt_dic[key] + 1
print "INFO: Number of samples in each label is=", test_cnt_dic
###############################################
###########plot prediction result figure#######
###############################################
pred_fname=os.path.join(local_out_dir,row_id_str+"_1"+".png")
true_fname=os.path.join(local_out_dir,row_id_str+"_2"+".png")
pred_xlabel='Prediction (Single Run)'
true_xlabel='True Labels (Single Run)'
test_cnt_dic=ml_util.ml_plot_predict_figures(labels_pred.tolist(), labels_test.tolist(), labels_list, label_dic, testing_sample_count
, pred_xlabel, pred_fname, true_xlabel, true_fname)
print "INFO: figure files: ", pred_fname, true_fname
print "INFO: Number of samples in each label is=", test_cnt_dic
roc_auc=None
#fscore=None
perf_measures=None
class_count=len(labels_list)
dataset_info={"training_fraction":training_fraction, "class_count":class_count,"dataset_count":sample_count}
#############################################################
###################for 2 class only (plot ROC curve)#########
#############################################################
if len(labels_list) == 2:
# build data file for score graph
score_graph_fname=os.path.join(local_out_dir,row_id_str+"_score_graph.json")
print "INFO: score_graph_fname=", score_graph_fname
ml_build_pred_score_graph(score_arr_0,score_arr_1,model_name, score_graph_fname,max_score,min_score)
do_ROC=True
reverse_label_dic = dict((v,k) for k, v in label_dic.items())
if 'clean' in reverse_label_dic:
flag_clean = reverse_label_dic['clean']
elif 'benign' in reverse_label_dic:
flag_clean = reverse_label_dic['benign']
elif '0' in reverse_label_dic:
flag_clean = 0
else:
print "No ROC curve generated: 'clean' or '0' must be a label for indicating negative class!"
do_ROC=False
if do_ROC:
# calculate fscore ==========
perf_measures=ml_util.calculate_fscore(labels_test, labels_pred)
print "INFO: perf_measures=",perf_measures
confidence_score = clf_best.decision_function(X_test_sparse)
if flag_clean == 0:
scores = [x for x in confidence_score]
s_labels = [x for x in labels_test]
testing_N = test_cnt_dic[0]
testing_P = test_cnt_dic[1]
else:
scores = [-x for x in confidence_score]
s_labels = [1-x for x in labels_test]
testing_N = test_cnt_dic[1]
testing_P = test_cnt_dic[0]
# create ROC data file ======== ====
roc_auc=ml_create_roc_files(row_id_str, scores, s_labels, testing_N, testing_P
, local_out_dir, row_id_str)
perf_measures["roc_auc"]=roc_auc
# only update db for web request
if fromweb=="1":
#print "database update"
str_sql="UPDATE atdml_document set "+"accuracy = '"+str(accuracy*100)+"%" \
+"', status = 'learned', processed_date ='"+str(datetime.datetime.now()) \
+"',ml_opts='"+j_str \
+"', perf_measures='"+json.dumps(perf_measures) \
+"', dataset_info='"+json.dumps(dataset_info) \
+"' where id="+row_id_str
ret=exec_sqlite.exec_sql(str_sql)
print "INFO: Data update done! ret=", str(ret)
else:
print "INFO: accuracy = '"+str(accuracy*100)+"%"
print 'INFO: total running time: %f' %(t1-t00)
print 'INFO: Finished!'
return 0
def feat_importance_firm(row_id_str, ds_id, hdfs_feat_dir, local_score_file,
sp_master, spark_rdd_compress,
spark_driver_maxResultSize, sp_exe_memory,
sp_core_max, zipout_dir, zipcode_dir, zip_file_name,
mongo_tuples, training_fraction, jobname, uploadtype,
description_file):
# zip func in other files for Spark workers ================= ================
zip_file_path = ml_util.ml_build_zip_file(zipout_dir,
zipcode_dir,
zip_file_name,
prefix='zip_feature_util')
print "INFO: zip_file_path=", zip_file_path
# get_spark_context
sc = ml_util.ml_get_spark_context(sp_master, spark_rdd_compress,
spark_driver_maxResultSize,
sp_exe_memory, sp_core_max, jobname,
[zip_file_path])
t0 = time()
# get feature seq mapping from mongo
if uploadtype == "MD5 List IN-dynamic":
### connect to database to get the column list which contains all column number of the corresponding feature
key = "dict_dynamic"
jstr_filter = '{"rid":' + row_id_str + ',"key":"' + key + '"}'
jstr_proj = '{"value":1}'
# get parent dataset's data
if ds_id != row_id_str:
jstr_filter = '{"rid":' + ds_id + ',"key":"' + key + '"}'
doc = query_mongo.find_one_t(mongo_tuples, jstr_filter, jstr_proj)
dic_list = doc['value']
dic_all_columns = {}
max_feature = 0
# reverse dict{hashes:sequence number} ======
for i in range(0, len(dic_list)):
for key in dic_list[i]:
dic_all_columns[eval(dic_list[i][key])] = key
if eval(dic_list[i][key]) > max_feature:
max_feature = eval(dic_list[i][key])
print "INFO: max_feature=", max_feature
#print "dic_all_columns=",dic_all_columns # fid:numb,numb
dirFile_loc = os.path.join(hdfs_feat_dir, "metadata")
dirFolders = sc.textFile(dirFile_loc)
hash_Folders = dirFolders.collect()
#print "INFO: dirFile_loc=",dirFile_loc,", hash_Folders=",hash_Folders
folder_list = [x.encode('UTF8') for x in hash_Folders]
print "INFO: hdfs folder_list=", folder_list #['dirty/', 'clean/']
features_training = []
labels_training = []
names_training = []
row_training = []
col_training = []
max_feat_training = 0
row_num_training = 0
features_testing = []
labels_testing = []
names_testing = []
row_testing = []
col_testing = []
max_feat_testing = 0
row_num_testing = 0
# loop through hdfs folders; TBD
for folder in folder_list:
print "INFO: folder=", folder
label = folder_list.index(folder) + 1
print 'INFO: label=', label
logFile_name = os.path.join(hdfs_feat_dir, folder, mtx_name_list)
#print "logFile_name=",logFile_name
logFile_data = os.path.join(hdfs_feat_dir, folder, mtx_libsvm)
#print "logFile_data=",logFile_data
logNames = sc.textFile(logFile_name).cache()
logData = sc.textFile(logFile_data).cache()
names = logNames.collect()
data = logData.collect()
name_l = [x.encode('UTF8') for x in names]
feature_l = [x.encode('UTF8') for x in data]
name_list = [names.strip() for names in name_l]
feature_list = [features.strip() for features in feature_l]
##########data seperation######
id_perm = data_seperation_random(name_list)
num_names = len(name_list)
print 'INFO: num of samples=', num_names
num_train = int(training_portion * num_names)
print 'INFO: num_train = ', num_train
########generate training data#########
i = 0
#print "INFO: generate training data"
#print "INFO: len(id_perm)=",len(id_perm)
while i < num_train:
#print i, id_perm[i]
features = feature_list[id_perm[i]]
features = features.strip()
feature_array = features.split(' ')
labels_training.append(label)
length = len(feature_array)
j = 0
while j < length:
feature = feature_array[j]
feat, value = feature.split(':', 2)
row_training.append(i + row_num_training)
col_training.append(int(feat) - 1)
features_training.append(int(value))
max_feat_training = max(max_feat_training, int(feat))
j = j + 1
i = i + 1
row_num_training = row_num_training + num_train
i = num_train
########generate testing data#########
while i < num_names:
####for generating testing data folder####
test_file_name = name_list[id_perm[i]]
features = feature_list[id_perm[i]]
features = features.strip()
feature_array = features.split(' ')
labels_testing.append(label)
length = len(feature_array)
j = 0
while j < length:
feature = feature_array[j]
feat, value = feature.split(':', 2)
row_testing.append(i - num_train + row_num_testing)
col_testing.append(int(feat) - 1)
features_testing.append(int(value))
max_feat_testing = max(max_feat_testing, int(feat))
j = j + 1
i = i + 1
row_num_testing = row_num_testing + (num_names - num_train)
# end for loop here ========================
col_num = max(max_feat_training, max_feat_testing)
if max_feat_training < col_num:
for i in range(0, row_num_training):
for j in range(max_feat_training, col_num):
features_training.append(0)
row_training.append(i)
col_training.append(j)
elif max_feat_testing < col_num:
for i in range(0, row_num_testing):
for j in range(max_feat_testing, col_num):
features_testing.append(0)
row_testing.append(i)
col_testing.append(j)
features_training = array(features_training)
row_training = array(row_training)
col_training = array(col_training)
#print "row_training:", row_training
#print "INFO: col_training:", col_training
len_col = len(col_training)
print "INFO: col_num:", col_num
labels_training = array(labels_training)
features_testing = array(features_testing)
row_testing = array(row_testing)
col_testing = array(col_testing)
labels_testing = array(labels_testing)
sparse_mtx = csc_matrix((features_training, (row_training, col_training)),
shape=(row_num_training, col_num))
#print "sparse_mtx.todense(), sparse_mtx.shape=",sparse_mtx.todense(), sparse_mtx.shape
sparse_test = csc_matrix((features_testing, (row_testing, col_testing)),
shape=(row_num_testing, col_num))
#print " sparse_test.todense(), sparse_test.shape=",sparse_test.todense(), sparse_test.shape
clf = svm.LinearSVC()
#clf = svm.SVC(C=0.1, kernel='rbf', degree=3, gamma=0.05, coef0=0.0, shrinking=True, probability=False, tol=0.001, cache_size=200, class_weight=None, verbose=False, max_iter=-1, random_state=None)
#clf = svm.NuSVC(nu=0.3, kernel='rbf', degree=3, gamma=0.05, coef0=0.0, shrinking=True, probability=False, tol=0.001, cache_size=200, verbose=False, max_iter=-1, random_state=None)
#print "labels_training=",labels_training
#print "sparse_mtx=",sparse_mtx
clf.fit(sparse_mtx, labels_training)
#print "INFO: model:intercept=",clf.intercept_
#print "INFO: model:coef=",clf.coef_
labels_pred = clf.predict(sparse_test)
#print "labels_pred:", labels_pred
accuracy = clf.score(sparse_test, labels_testing)
#print "INFO: data folder=", hdfs_feat_dir
print "INFO: accuracy=", accuracy
#####################################################################
##################calculate feature importance with predication labels#######################
#####################################################################
AA = sparse_mtx.todense()
BB = sparse_test.todense()
labels_train_pred = clf.predict(sparse_mtx)
labels_test_pred = labels_pred
#print "###################################################################################"
print "INFO: ======= Calculate feature importance with predication labels =================="
#print "###################################################################################"
dic_importance_label = {}
for j in range(0, col_num): ###for all features in the loop
##############################
#print "====new way with sparse matrix========="
curr_col_train = sparse_mtx.getcol(j)
sum_col = curr_col_train.sum(0)
positive_feature_number = int(sum_col.tolist()[0][0])
labels_value = 3 - labels_train_pred
dot_product = csr_matrix(np.array(labels_value)).dot(curr_col_train)
sum_product = dot_product.sum(1)
labels_positive_sum = int(sum_product.tolist()[0][0])
sum_label_values = sum(labels_value)
labels_negitive_sum = sum_label_values - labels_positive_sum
##############################
#print "====new way with sparse matrix========="
curr_col_test = sparse_test.getcol(j)
sum_col = curr_col_test.sum(0)
positive_feature_number = positive_feature_number + int(
sum_col.tolist()[0][0])
labels_value = 3 - labels_test_pred
dot_product = csr_matrix(np.array(labels_value)).dot(curr_col_test)
sum_product = dot_product.sum(1)
labels_positive_sum = labels_positive_sum + int(
sum_product.tolist()[0][0])
sum_label_values = sum(labels_value)
labels_negitive_sum = labels_negitive_sum + sum_label_values - int(
sum_product.tolist()[0][0])
n_total = row_num_training + row_num_testing
negitive_feature_number = n_total - positive_feature_number
if positive_feature_number == 0:
#print "feature ", j+1, "all 0s!"
dic_importance_label[j + 1] = -100
elif negitive_feature_number == 0:
#print "feature ", j+1, "all 1s!"
dic_importance_label[j + 1] = -200
else:
q_positive = float(labels_positive_sum) / positive_feature_number
q_negitive = float(labels_negitive_sum) / negitive_feature_number
Q = (q_positive - q_negitive) * sqrt(
float(q_positive) * q_negitive / float(n_total) /
float(n_total))
dic_importance_label[j + 1] = Q
sorted_importance = sorted(dic_importance_label.items(),
key=operator.itemgetter(1),
reverse=True)
print "INFO: ======= Feature Importance(FIRM score) ================"
if os.path.exists(local_score_file):
try:
os.remove(local_score_file)
except OSError, e:
print("ERROR: %s - %s." % (e.local_score_file, e.strerror))
def feat_importance_firm(row_id_str, ds_id, hdfs_feat_dir, local_score_file,
sp_master, spark_rdd_compress,
spark_driver_maxResultSize, sp_exe_memory,
sp_core_max, zipout_dir, zipcode_dir, zip_file_name,
mongo_tuples, training_fraction, jobname, uploadtype):
# zip func in other files for Spark workers ================= ================
zip_file_path = ml_util.ml_build_zip_file(zipout_dir,
zipcode_dir,
zip_file_name,
prefix='zip_feature_util')
print "INFO: zip_file_path=", zip_file_path
# get_spark_context
sc = ml_util.ml_get_spark_context(sp_master, spark_rdd_compress,
spark_driver_maxResultSize,
sp_exe_memory, sp_core_max, jobname,
[zip_file_path])
'''
SparkContext.setSystemProperty('spark.rdd.compress', config.get('spark', 'spark_rdd_compress'))
SparkContext.setSystemProperty('spark.driver.maxResultSize', config.get('spark', 'spark_driver_maxResultSize'))
SparkContext.setSystemProperty('spark.executor.memory', args.exe_memory)
SparkContext.setSystemProperty('spark.cores.max', args.core_max)
sc = SparkContext(args.sp_master, 'feature_importance_FRIM:'+str(args.row_id))
'''
t0 = time()
# get folder list (labels) from hdfs data_out/<id>/metadata ==============
dirFile_loc = os.path.join(hdfs_feat_dir, "metadata")
dirFolders = sc.textFile(dirFile_loc)
hash_Folders = dirFolders.collect()
print "INFO: dirFile_loc=", dirFile_loc, ", hash_Folders=", hash_Folders
folder_list = [x.encode('UTF8') for x in hash_Folders]
print "INFO: folder_list=", folder_list #['dirty/', 'clean/']
# get feature seq : ngram hash mapping ==================================
key = "dic_seq_hashes" #{"123":"136,345"}
jstr_filter = '{"rid":' + row_id_str + ',"key":"' + key + '"}'
jstr_proj = '{"value":1}'
# get parent dataset's data
if ds_id != row_id_str:
jstr_filter = '{"rid":' + ds_id + ',"key":"' + key + '"}'
doc = query_mongo.find_one_t(mongo_tuples, jstr_filter, jstr_proj)
dic_list = doc['value']
dic_all_columns = dic_list
feature_count = len(dic_list)
#print "INFO: feature_count=",feature_count
#print "dic_list=",dic_list #{u'123,345':u'136'}
#print "dic_all_columns=",dic_all_columns # {1: u'8215,8216'}
# end
# get {hash : raw string} mapping ==================================
key = "dic_hash_str" #{"123":"openFile"}
jstr_filter = '{"rid":' + row_id_str + ',"key":"' + key + '"}'
jstr_proj = '{"value":1}'
# get parent dataset's data
if ds_id != row_id_str:
jstr_filter = '{"rid":' + ds_id + ',"key":"' + key + '"}'
doc = query_mongo.find_one_t(mongo_tuples, jstr_filter, jstr_proj)
dic_hash_str = doc['value']
'''
# get folder list (labels) from hdfs data_out/<id>/libsvm ==============
libsvm_loc = os.path.join(hdfs_feat_dir , "libsvm_data")
print "INFO: libsvm_loc=", libsvm_loc
samples_rdd = MLUtils.loadLibSVMFile(sc, libsvm_loc)
'''
# filename for featured data
libsvm_data_file = os.path.join(hdfs_feat_dir, "libsvm_data")
print "INFO: libsvm_data_file=", libsvm_data_file
# load feature count file
#feat_count_file=libsvm_data_file+"_feat_count"
#feature_count=zip_feature_util.get_feature_count(sc,feat_count_file)
print "INFO: feature_count=", feature_count
#samples_rdd = MLUtils.loadLibSVMFile(sc, libsvm_data_file)
# load sample RDD from text file
# also exclude selected features in sample ================ =====
# format (LabeledPoint,hash) from str2LabeledPoint_hash()
samples_rdd, feature_count = zip_feature_util.get_sample_rdd(
sc, libsvm_data_file, feature_count, excluded_feat_cslist=None)
labels_and_features_rdd = samples_rdd.map(lambda p:
(p[0].label, p[0].features))
all_data = labels_and_features_rdd.collect()
features_list = [x.toArray() for _, x in all_data]
labels_list_all = [x for x, _ in all_data]
labels_list_all = np.array(labels_list_all)
features_array = np.array(features_list)
### generate sparse matrix (csr) for all samples
features_sparse_mtx = csr_matrix(features_array)
### randomly split the samples into training and testing data
sparse_mtx, sparse_test, labels_training, labels_testing = \
cross_validation.train_test_split(features_sparse_mtx, labels_list_all, test_size=(1-training_fraction))
#print "INFO: sparse_mtx.shape=",sparse_mtx.shape
#print "INFO: sparse_test.shape=",sparse_test.shape
row_num_training = (sparse_mtx.shape)[0]
row_num_testing = (sparse_test.shape)[0]
# why use LinearSVC ?
clf = svm.LinearSVC()
#clf = svm.SVC(C=0.1, kernel='rbf', degree=3, gamma=0.05, coef0=0.0, shrinking=True, probability=False, tol=0.001, cache_size=200, class_weight=None, verbose=False, max_iter=-1, random_state=None)
#clf = svm.NuSVC(nu=0.3, kernel='rbf', degree=3, gamma=0.05, coef0=0.0, shrinking=True, probability=False, tol=0.001, cache_size=200, verbose=False, max_iter=-1, random_state=None)
#print "labels_training=",labels_training
#print "sparse_mtx=",sparse_mtx
clf.fit(sparse_mtx, labels_training)
#print "**model:intercept***"
#print clf.intercept_
#print "**model:coef***"
#print clf.coef_
col_num = len(clf.coef_[0]) # for n_classes==2
print "INFO: col_num=", col_num
labels_pred = clf.predict(sparse_test)
#print "labels_pred:", labels_pred
accuracy = clf.score(sparse_test, labels_testing)
print "INFO: data folder:", hdfs_feat_dir
print "INFO: accuracy: ", accuracy
#####################################################################
##################calculate feature importance with predication labels#######################
#####################################################################
AA = sparse_mtx.todense()
BB = sparse_test.todense()
labels_train_pred = clf.predict(sparse_mtx)
labels_test_pred = labels_pred
print "INFO: ###################################################################################"
print "INFO: ############calculate feature importance with predication labels###################"
print "INFO: ###################################################################################"
dic_importance_label = {}
for j in range(0, col_num): ###for all features in the loop
##############################
#print "====new way with sparse matrix========="
curr_col_train = sparse_mtx.getcol(j)
sum_col = curr_col_train.sum(0)
positive_feature_number = int(sum_col.tolist()[0][0])
labels_value = 3 - labels_train_pred
dot_product = csr_matrix(np.array(labels_value)).dot(curr_col_train)
sum_product = dot_product.sum(1)
labels_positive_sum = int(sum_product.tolist()[0][0])
sum_label_values = sum(labels_value)
labels_negitive_sum = sum_label_values - labels_positive_sum
##############################
#print "====new way with sparse matrix========="
curr_col_test = sparse_test.getcol(j)
sum_col = curr_col_test.sum(0)
positive_feature_number = positive_feature_number + int(
sum_col.tolist()[0][0])
labels_value = 3 - labels_test_pred
dot_product = csr_matrix(np.array(labels_value)).dot(curr_col_test)
sum_product = dot_product.sum(1)
labels_positive_sum = labels_positive_sum + int(
sum_product.tolist()[0][0])
sum_label_values = sum(labels_value)
labels_negitive_sum = labels_negitive_sum + sum_label_values - int(
sum_product.tolist()[0][0])
n_total = row_num_training + row_num_testing
negitive_feature_number = n_total - positive_feature_number
if positive_feature_number == 0:
#print "feature ", j+1, "all 0s!"
dic_importance_label[j + 1] = -100
elif negitive_feature_number == 0:
#print "feature ", j+1, "all 1s!"
dic_importance_label[j + 1] = -200
else:
q_positive = float(labels_positive_sum) / positive_feature_number
q_negitive = float(labels_negitive_sum) / negitive_feature_number
Q = (q_positive - q_negitive) * sqrt(
float(q_positive) * q_negitive / float(n_total) /
float(n_total))
dic_importance_label[j + 1] = Q
sorted_importance = sorted(dic_importance_label.items(),
key=operator.itemgetter(1),
reverse=True)
print "INFO: =======Feature Importance(FIRM score)================"
if os.path.exists(local_score_file):
try:
os.remove(local_score_file)
except OSError, e:
print("ERROR: %s - %s." % (e.local_score_file, e.strerror))
def train(row_id_str, ds_id, hdfs_feat_dir, local_out_dir, ml_opts_jstr,
sp_master, spark_rdd_compress, spark_driver_maxResultSize,
sp_exe_memory, sp_core_max, zipout_dir, zipcode_dir, zip_file_name,
mongo_tuples, labelnameflag, fromweb, src_filename, jobname,
model_data_folder):
# create zip files for Spark workers ================= ================
zip_file_path = ml_build_zip_file(zipout_dir,
zipcode_dir,
zip_file_name,
prefix='zip_feature_util')
print "INFO: zip_file_path=", zip_file_path
#data_folder = hdfs_feat_dir + "/"
#local_out_dir = local_out_dir + "/"
if not os.path.exists(local_out_dir):
os.makedirs(local_out_dir)
# ML model filename ====
model_fname = os.path.join(model_data_folder, row_id_str + '.pkl')
print "INFO: model_data_folder=", model_data_folder
# create out folders and clean up old model files ====
ml_util.ml_prepare_output_dirs(row_id_str, local_out_dir,
model_data_folder, model_fname)
# init Spark context ====
sc = ml_util.ml_get_spark_context(sp_master, spark_rdd_compress,
spark_driver_maxResultSize,
sp_exe_memory, sp_core_max, jobname,
[zip_file_path])
# start here =================================================================== ===============
t0 = time()
### load libsvm file: may or may not be PCA-ed ###
libsvm_data_file = os.path.join(hdfs_feat_dir, src_filename)
print "INFO: libsvm_data_file=", libsvm_data_file
# feature count is a variable if PCA
feature_count = 0
# samples_rdd may be from PCAed data
# load sample RDD from text file
# format (LabeledPoint,hash) from str2LabeledPoint_hash()
samples_rdd, feature_count = zip_feature_util.get_sample_rdd(
sc, libsvm_data_file, feature_count, '')
# collect all data to local for processing ===============
all_data = samples_rdd.collect()
total_sample_count = len(all_data)
# 2-D array, may be PCAed
features_list = [x.features.toArray() for x, _ in all_data]
# label array
labels_list_all = [x.label for x, _ in all_data]
# hash array
hash_list_all = [x for _, x in all_data]
# convert to np array
features_array_reduced = np.array(features_list)
hash_list_all = np.array(hash_list_all)
labels_list_all = np.array(labels_list_all)
true_label_array = np.array(labels_list_all, dtype=np.int8)
print "INFO: total_sample_count=", total_sample_count
print "INFO: features_array_reduced.shape=", features_array_reduced.shape
print "INFO: labels_list_all.shape=", labels_list_all.shape
print "INFO: true_label_array.shape=", true_label_array.shape
t1 = time()
print 'INFO: data generating time: %f' % (t1 - t0)
###############################################
########## build learning model ###############
###############################################
### parse parameters and generate the model ###
(model, alg, n_clusters) = parse_para_and_get_model(ml_opts_jstr)
if model is None:
return
labels_kmeans = None
#### fit the model to training dataset ####
try:
model.fit(features_array_reduced)
labels_kmeans = model.labels_ #'numpy.ndarray'
except:
print "ERROR: Error in model.fit(): ", "model=", model, ", sys.exc_info:", sys.exc_info(
)[0]
return
#### save clf for future use ####
#joblib.dump(model, model_data_folder + row_id_str+'.pkl')
joblib.dump(model, model_fname)
#print "**model:intercept***"
#print clf.intercept_
print "INFO: model type=", type(model), " model=", model
###################################################
### generate label names (family names) ###########
### connect to database to get the column list which contains all column number of the corresponding feature####
###################################################
if labelnameflag == 1:
key = "dic_name_label"
jstr_filter = '{"rid":' + row_id_str + ',"key":"' + key + '"}'
jstr_proj = '{"value":1}'
# get parent dataset's data
if ds_id != row_id_str:
jstr_filter = '{"rid":' + ds_id + ',"key":"' + key + '"}'
doc = query_mongo.find_one_t(mongo_tuples, jstr_filter, jstr_proj)
dic_list = doc['value']
label_dic = {}
for i in range(0, len(dic_list)):
for key in dic_list[i]:
label_dic[dic_list[i][key]] = key.encode('UTF8')
print "INFO: label_dic:", label_dic
else:
label_dic = {}
label_set = set(labels_list_all)
for label_value in label_set:
label_dic[int(label_value)] = str(int(label_value))
print "INFO: generated label_dic:", label_dic
labels_list = []
for key in sorted(label_dic):
labels_list.append(label_dic[key])
print "INFO: labels_list=", labels_list
#Adjusted Mutual Information between two clusterings
amis = adjusted_mutual_info_score(labels_list_all, labels_kmeans)
print "INFO: Adjusted_mutual_info_score=", amis
#Similarity measure between two clusterings
ars = adjusted_rand_score(labels_list_all, labels_kmeans)
print "INFO: Adjusted_rand_score=", ars
###################################################
#######plot histogram ####
###################################################
plot_col_num = int(math.ceil(math.sqrt(n_clusters)))
figsize = (4 * plot_col_num,
3 * int(math.ceil(n_clusters * 1.0 / plot_col_num)))
print "INFO: labels_list_all.shape=", labels_list_all.shape, "labels_kmeans.shape=", labels_kmeans.shape
print "INFO: labels_list_all t=", type(
labels_list_all), "labels_kmeans t=", type(labels_kmeans)
print "INFO: n_clusters=", n_clusters, ",label_dic=", label_dic
print "INFO: plot_col_num=", plot_col_num, ",figsize=", figsize, ",local_out_dir=", local_out_dir
# kmeans histogram
_, p_true = ml_plot_kmeans_histogram_subfigures(labels_list_all,
labels_kmeans,
n_clusters,
names=label_dic,
plot_col_num=plot_col_num,
figsize=figsize,
folder=local_out_dir,
rid=row_id_str)
# normalized kmeans histogram
_, p_true_norm = ml_plot_kmeans_histogram_subfigures(
labels_list_all,
labels_kmeans,
n_clusters,
names=label_dic,
plot_col_num=plot_col_num,
figsize=figsize,
normalize=True,
folder=local_out_dir,
rid=row_id_str)
####plot "reverse" histogram with labels ####
#num_bars = len(np.unique(labels_list_all))
num_bars = max(labels_list_all) + 1
figsize = (4 * plot_col_num,
3 * int(math.ceil(num_bars * 1.0 / plot_col_num)))
_, p_cluster = ml_plot_kmeans_histogram_subfigures(
labels_kmeans,
labels_list_all,
num_bars,
names=label_dic,
plot_col_num=plot_col_num,
figsize=figsize,
reverse=True,
folder=local_out_dir,
rid=row_id_str)
#### plot dot figures ####
#mtx_label = model.labels_
mtx_center = model.cluster_centers_
# dot plot for Kmeans ===========
filename = os.path.join(local_out_dir, row_id_str + '_cluster.png')
filename_3d = os.path.join(local_out_dir, row_id_str + '_cluster_3d.json')
ml_plot_kmeans_dot_graph_save_file(features_array_reduced,
labels_kmeans,
mtx_center,
n_clusters,
figsize=(10, 7),
filename=filename,
title='KMeans',
filename_3d=filename_3d)
#print "features_array_reduced s=",features_array_reduced.shape
# dot plot for True Labels ===========
filename = os.path.join(local_out_dir, row_id_str + '_cluster_tl.png')
filename_3d = os.path.join(local_out_dir,
row_id_str + '_cluster_3d_tl.json')
ml_plot_kmeans_dot_graph_save_file(features_array_reduced,
true_label_array,
mtx_center,
n_clusters,
figsize=(10, 7),
filename=filename,
title='True Labels',
filename_3d=filename_3d)
dataset_info = {
"training_fraction": 1,
"class_count": n_clusters,
"dataset_count": total_sample_count
}
# only update db for web request ===========
if fromweb == "1":
#print "database update"
str_sql="UPDATE atdml_document set accuracy = '" \
+"', status = 'learned', processed_date ='"+str(datetime.datetime.now()) \
+"', total_feature_numb='"+str(feature_count) \
+"', perf_measures='{}" \
+"', dataset_info='"+json.dumps(dataset_info) \
+"' where id="+row_id_str
ret = exec_sqlite.exec_sql(str_sql)
print "INFO: Data update done! ret=", str(ret)
else:
print "INFO: accuracy = '" + str(accuracy * 100) + "%"
t1 = time()
print 'INFO: running time: %f' % (t1 - t0)
#print 'Finished!'
return 0