add new network architecture - server label moves to the middle

This commit is contained in:
René Knaebel 2017-07-29 19:42:36 +02:00
parent 8cd1023165
commit 820a5d1a4d
5 changed files with 151 additions and 40 deletions

View File

@ -120,7 +120,6 @@ def create_dataset_from_flows(user_flow_df, char_dict, max_len, window_size=10):
server_tr = server_tr[idx]
client_tr = np_utils.to_categorical(client_tr, 2)
server_tr = np_utils.to_categorical(server_tr, 2)
return domain_tr, flow_tr, client_tr, server_tr
@ -166,7 +165,7 @@ def create_dataset_from_lists(chunks, vocab, max_len):
logger.info(" select names")
names = np.unique(np.stack(map(lambda f: f.user_hash, chunks)))
logger.info(" select servers")
servers = np.max(np.stack(map(lambda f: f.serverLabel, chunks)), axis=1)
servers = np.stack(map(lambda f: f.serverLabel, chunks))
logger.info(" select trusted hits")
trusted_hits = np.max(np.stack(map(lambda f: f.trustedHits, chunks)), axis=1)

127
main.py
View File

@ -7,7 +7,7 @@ import pandas as pd
import tensorflow as tf
from keras.callbacks import ModelCheckpoint, CSVLogger, EarlyStopping
from keras.models import load_model
from sklearn.decomposition import PCA
from keras.utils import np_utils
from sklearn.utils import class_weight
import arguments
@ -46,8 +46,6 @@ ch.setFormatter(formatter)
# add ch to logger
logger.addHandler(ch)
print = logger.info
args = arguments.parse()
if args.gpu:
@ -104,8 +102,8 @@ def main_hyperband():
def get_custom_class_weights(client_tr, server_tr):
client = client_tr.value.argmax(1)
server = server_tr.value.argmax(1)
client = client_tr.value.argmax(1) if type(client_tr) != np.ndarray else client_tr.argmax(1)
server = server_tr.value.argmax(1) if type(server_tr) != np.ndarray else server_tr.argmax(1)
client_class_weight = class_weight.compute_class_weight('balanced', np.unique(client), client)
server_class_weight = class_weight.compute_class_weight('balanced', np.unique(server), server)
return {
@ -118,7 +116,7 @@ def main_train(param=None):
exists_or_make_path(args.model_path)
char_dict = dataset.get_character_dict()
domain_tr, flow_tr, client_tr, server_tr = load_or_generate_h5data(args.train_h5data, args.train_data,
domain_tr, flow_tr, client_tr, server_windows_tr = load_or_generate_h5data(args.train_h5data, args.train_data,
args.domain_length, args.window)
# parameter
@ -133,10 +131,10 @@ def main_train(param=None):
'dropout': 0.5,
'domain_features': args.domain_embedding,
'embedding_size': args.embedding,
'filter_main': 128,
'filter_main': 64,
'flow_features': 3,
# 'dense_main': 512,
'dense_main': 128,
'dense_main': 64,
'filter_embedding': args.hidden_char_dims,
'hidden_embedding': args.domain_embedding,
'kernel_embedding': 3,
@ -146,7 +144,7 @@ def main_train(param=None):
if not param:
param = p
embedding, model = models.get_models_by_params(param)
embedding, model, _ = models.get_models_by_params(param)
embedding.summary()
model.summary()
logger.info("define callbacks")
@ -166,6 +164,8 @@ def main_train(param=None):
loss='categorical_crossentropy',
metrics=['accuracy'] + custom_metrics)
server_tr = np_utils.to_categorical(np.max(server_windows_tr, axis=1), 2)
if args.class_weights:
logger.info("class weights: compute custom weights")
custom_class_weights = get_custom_class_weights(client_tr, server_tr)
@ -200,7 +200,7 @@ def main_test():
char_dict = dataset.get_character_dict()
user_flow_df = dataset.get_user_flow_data(args.test_data)
domains = user_flow_df.domain.unique()
domains = user_flow_df.domain.unique()[:-1]
def get_domain_features_reduced(d):
return dataset.get_domain_features(d[0], char_dict, args.domain_length)
@ -211,13 +211,93 @@ def main_test():
model = load_model(args.embedding_model)
domain_features = np.stack(domain_features).reshape((-1, 40))
pred = model.predict(domains, batch_size=args.batch_size, verbose=1)
pred = model.predict(domain_features, batch_size=args.batch_size, verbose=1)
np.save("/tmp/rk/domains.npy", domains)
np.save("/tmp/rk/domain_features.npy", domain_features)
np.save("/tmp/rk/domain_embd.npy", pred)
def main_new_model():
exists_or_make_path(args.model_path)
char_dict = dataset.get_character_dict()
domain_tr, flow_tr, client_tr, server_windows_tr = load_or_generate_h5data(args.train_h5data, args.train_data,
args.domain_length, args.window)
# parameter
p = {
"type": args.model_type,
"batch_size": 64,
"window_size": args.window,
"domain_length": args.domain_length,
"flow_features": 3,
"vocab_size": len(char_dict) + 1,
#
'dropout': 0.5,
'domain_features': args.domain_embedding,
'embedding_size': args.embedding,
'filter_main': 64,
'flow_features': 3,
# 'dense_main': 512,
'dense_main': 64,
'filter_embedding': args.hidden_char_dims,
'hidden_embedding': args.domain_embedding,
'kernel_embedding': 3,
'kernels_main': 3,
'input_length': 40
}
embedding, _, model = models.get_models_by_params(p)
embedding.summary()
model.summary()
logger.info("define callbacks")
callbacks = []
callbacks.append(ModelCheckpoint(filepath=args.clf_model,
monitor='val_loss',
verbose=False,
save_best_only=True))
callbacks.append(CSVLogger(args.train_log))
if args.stop_early:
callbacks.append(EarlyStopping(monitor='val_loss',
patience=5,
verbose=False))
logger.info("compile model")
custom_metrics = models.get_metric_functions()
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'] + custom_metrics)
server_tr = np_utils.to_categorical(np.max(server_windows_tr, axis=1), 2)
if args.class_weights:
logger.info("class weights: compute custom weights")
custom_class_weights = get_custom_class_weights(client_tr, server_tr)
logger.info(custom_class_weights)
else:
logger.info("class weights: set default")
custom_class_weights = None
logger.info("start training")
server_tr = np.stack(np_utils.to_categorical(s, 2) for s in server_windows_tr)
model.fit([domain_tr, flow_tr],
[client_tr, server_tr],
batch_size=args.batch_size,
epochs=args.epochs,
callbacks=callbacks,
shuffle=True,
validation_split=0.2,
class_weight=custom_class_weights)
logger.info("save embedding")
embedding.save(args.embedding_model)
def main_embedding():
model = load_model(args.embedding_model)
domain_encs, labels = dataset.load_or_generate_domains(args.train_data, args.domain_length)
domain_embedding = model.predict(domain_encs, batch_size=args.batch_size, verbose=1)
visualize.plot_embedding(domain_embedding, labels, path="results/pp3/embd.png")
def main_visualization():
domain_val, flow_val, client_val, server_val = load_or_generate_h5data(args.test_h5data, args.test_data,
args.domain_length, args.window)
@ -248,27 +328,6 @@ def main_visualization():
"{}/server_cov.png".format(args.model_path),
normalize=False, title="Server Confusion Matrix")
# embedding visi
import matplotlib.pyplot as plt
model = load_model(args.embedding_model)
domains = np.reshape(domain_val, (domain_val.shape[0] * domain_val.shape[1], 40))
domain_embedding = model.predict(domains, batch_size=args.batch_size, verbose=1)
pca = PCA(n_components=2)
domain_reduced = pca.fit_transform(domain_embedding)
print(pca.explained_variance_ratio_)
clients = np.repeat(client_val, 10, axis=0)
clients = clients.argmax(1)
servers = np.repeat(server_val, 10, axis=0)
servers = servers.argmax(1)
plt.scatter(domain_reduced[:, 0], domain_reduced[:, 1], c=clients, cmap=plt.cm.bwr, s=2)
plt.show()
plt.scatter(domain_reduced[:, 0], domain_reduced[:, 1], c=servers, cmap=plt.cm.bwr, s=2)
plt.show()
def main_score():
# mask = dataset.load_mask_eval(args.data, args.test_image)
@ -281,7 +340,7 @@ def main_data():
char_dict = dataset.get_character_dict()
user_flow_df = dataset.get_user_flow_data(args.train_data)
logger.info("create training dataset")
domain_tr, flow_tr, client_tr, server_tr = dataset.create_dataset_from_flows(user_flow_df, char_dict,
domain_tr, flow_tr, client_tr, server_tr, _ = dataset.create_dataset_from_flows(user_flow_df, char_dict,
max_len=args.domain_length,
window_size=args.window)
print(f"domain shape {domain_tr.shape}")
@ -305,6 +364,8 @@ def main():
main_paul_best()
if "data" in args.modes:
main_data()
if "train_new" in args.modes:
main_new_model()
if __name__ == "__main__":

View File

@ -32,7 +32,10 @@ def get_models_by_params(params: dict):
predict_model = networks.get_model(dropout, flow_features, domain_features, window_size, domain_length,
filter_main, kernel_main, dense_dim, embedding_model)
return embedding_model, predict_model
new_model = networks.get_new_model(dropout, flow_features, domain_features, window_size, domain_length,
filter_main, kernel_main, dense_dim, embedding_model)
return embedding_model, predict_model, new_model
def get_metrics():

View File

@ -30,8 +30,8 @@ def get_embedding(vocab_size, embedding_size, input_length,
y = Embedding(input_dim=vocab_size, output_dim=embedding_size)(y)
y = Conv1D(filters, kernel_size, activation='relu')(y)
y = GlobalMaxPooling1D()(y)
y = Dense(hidden_dims)(y)
y = Dropout(drop_out)(y)
y = Dense(hidden_dims)(y)
y = Activation('relu')(y)
return Model(x, y)
@ -56,3 +56,27 @@ def get_model(cnnDropout, flow_features, domain_features, window_size, domain_le
y2 = Dense(2, activation='softmax', name="server")(y)
return Model(inputs=[ipt_domains, ipt_flows], outputs=(y1, y2))
def get_new_model(dropout, flow_features, domain_features, window_size, domain_length, cnn_dims, kernel_size,
dense_dim, cnn):
ipt_domains = Input(shape=(window_size, domain_length), name="ipt_domains")
ipt_flows = Input(shape=(window_size, flow_features), name="ipt_flows")
encoded = TimeDistributed(cnn)(ipt_domains)
y2 = Dense(2, activation="softmax", name="server")(encoded)
merged = keras.layers.concatenate([encoded, ipt_flows, y2], -1)
y = Conv1D(cnn_dims,
kernel_size,
activation='relu',
input_shape=(window_size, domain_features + flow_features))(merged)
# remove temporal dimension by global max pooling
y = GlobalMaxPooling1D()(y)
y = Dropout(dropout)(y)
y = Dense(dense_dim, activation='relu')(y)
y1 = Dense(2, activation='softmax', name="client")(y)
model = Model(inputs=[ipt_domains, ipt_flows], outputs=(y1, y2))
return model

View File

@ -42,3 +42,27 @@ def get_model(cnnDropout, flow_features, domain_features, window_size, domain_le
y2 = Dense(2, activation='softmax', name="server")(y)
return Model(inputs=[ipt_domains, ipt_flows], outputs=(y1, y2))
def get_new_model(dropout, flow_features, domain_features, window_size, domain_length, cnn_dims, kernel_size,
dense_dim, cnn):
ipt_domains = Input(shape=(window_size, domain_length), name="ipt_domains")
ipt_flows = Input(shape=(window_size, flow_features), name="ipt_flows")
encoded = TimeDistributed(cnn)(ipt_domains)
y2 = Dense(2, activation="softmax", name="server")(encoded)
merged = keras.layers.concatenate([encoded, ipt_flows, y2], -1)
y = Conv1D(cnn_dims,
kernel_size,
activation='relu',
input_shape=(window_size, domain_features + flow_features))(merged)
# remove temporal dimension by global max pooling
y = GlobalMaxPooling1D()(y)
y = Dropout(dropout)(y)
y = Dense(dense_dim, activation='relu')(y)
y1 = Dense(2, activation='softmax', name="client")(y)
model = Model(inputs=[ipt_domains, ipt_flows], outputs=(y1, y2))
return model