seyyedsalehi
/
PFP-WGAN
Watch 1
Star 0
Fork 0
Code Issues 0 Pull Requests 0 Releases 0 Wiki Activity
Browse Source

Add existing file

master
seydsalehi 4 years ago
parent
0b25badeab 0226b53ec3
commit
a4e876698c
3 changed files with 0 additions and 2001 deletions
Unified View Show Diff Stats
  1. 0
    902
      conditional_new_nn_hirearchical_seq.py
  2. 0
    917
      conditional_new_nn_hirearchical_seq_mt.py
  3. 0
    182
      conditional_wgan_wrapper_post.py

+ 0
- 902
conditional_new_nn_hirearchical_seq.py View File

#!/usr/bin/env python

"""
parts of codes are based on the work in https://github.com/bio-ontology-research-group/deepgo.

"""
from __future__ import division

import logging
import sys
import time
from collections import deque
from multiprocessing import Pool

import click as ck
import numpy as np
import pandas as pd
import tensorflow as tf
from keras import backend as K
from keras.callbacks import EarlyStopping, ModelCheckpoint
from keras.layers import (
Dense, Input, SpatialDropout1D, Conv1D, MaxPooling1D,
Flatten, Concatenate, Add, Maximum, Embedding, BatchNormalization, Activation, Dropout)
from keras.losses import binary_crossentropy
from keras.models import Sequential, Model, load_model
from keras.preprocessing import sequence
from scipy.spatial import distance
from sklearn.metrics import log_loss
from sklearn.metrics import roc_curve, auc, matthews_corrcoef
from keras.layers import Lambda
from sklearn.metrics import precision_recall_curve

from utils import (
get_gene_ontology,
get_go_set,
get_anchestors,
get_parents,
DataGenerator,
FUNC_DICT,
get_height,
get_ipro)
from conditional_wgan_wrapper_post import WGAN_wrapper, wasserstein_loss, generator_recunstruction_loss_new

config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
K.set_session(sess)

logging.basicConfig(format='%(levelname)s:%(message)s', level=logging.INFO)
sys.setrecursionlimit(100000)

DATA_ROOT = 'data/swiss/'
MAXLEN = 1000
REPLEN = 256
ind = 0


@ck.command()
@ck.option(
'--function',
default='bp',
help='Ontology id (mf, bp, cc)')
@ck.option(
'--device',
default='gpu:0',
help='GPU or CPU device id')
@ck.option(
'--org',
default= None,
help='Organism id for filtering test set')
@ck.option('--train',default = True, is_flag=True)
@ck.option('--param', default=0, help='Param index 0-7')
def main(function, device, org, train, param):
global FUNCTION
FUNCTION = function
global GO_ID
GO_ID = FUNC_DICT[FUNCTION]
global go
go = get_gene_ontology('go.obo')
global ORG
ORG = org
func_df = pd.read_pickle(DATA_ROOT + FUNCTION + '.pkl')
global functions
functions = func_df['functions'].values
global func_set
func_set = set(functions)
global all_functions
all_functions = get_go_set(go, GO_ID)
logging.info('Functions: %s %d' % (FUNCTION, len(functions)))
if ORG is not None:
logging.info('Organism %s' % ORG)
global go_indexes
go_indexes = dict()
for ind, go_id in enumerate(functions):
go_indexes[go_id] = ind
global node_names
node_names = set()
with tf.device('/' + device):
params = {
'fc_output': 1024,
'learning_rate': 0.001,
'embedding_dims': 128,
'embedding_dropout': 0.2,
'nb_conv': 1,
'nb_dense': 1,
'filter_length': 128,
'nb_filter': 32,
'pool_length': 64,
'stride': 32
}
model(params, is_train=train)

def load_data():

df = pd.read_pickle(DATA_ROOT + 'train' + '-' + FUNCTION + '.pkl')
n = len(df)
index = df.index.values
valid_n = int(n * 0.8)
train_df = df.loc[index[:valid_n]]
valid_df = df.loc[index[valid_n:]]
test_df = pd.read_pickle(DATA_ROOT + 'test' + '-' + FUNCTION + '.pkl')
print( test_df['orgs'] )
if ORG is not None:
logging.info('Unfiltered test size: %d' % len(test_df))
test_df = test_df[test_df['orgs'] == ORG]
logging.info('Filtered test size: %d' % len(test_df))

# Filter by type
# org_df = pd.read_pickle('data/prokaryotes.pkl')
# orgs = org_df['orgs']
# test_df = test_df[test_df['orgs'].isin(orgs)]

def reshape(values):
values = np.hstack(values).reshape(
len(values), len(values[0]))
return values

def normalize_minmax(values):
mn = np.min(values)
mx = np.max(values)
if mx - mn != 0.0:
return (values - mn) / (mx - mn)
return values - mn

def get_values(data_frame):
print(data_frame['labels'].values.shape)
labels = reshape(data_frame['labels'].values)
ngrams = sequence.pad_sequences(
data_frame['ngrams'].values, maxlen=MAXLEN)
ngrams = reshape(ngrams)
rep = reshape(data_frame['embeddings'].values)
data = ngrams
return data, labels

train = get_values(train_df)
valid = get_values(valid_df)
test = get_values(test_df)

return train, valid, test, train_df, valid_df, test_df


def get_feature_model(params):
embedding_dims = params['embedding_dims']
max_features = 8001
model = Sequential()
model.add(Embedding(
max_features,
embedding_dims,
input_length=MAXLEN))
model.add(SpatialDropout1D(0.4))
for i in range(params['nb_conv']):
model.add(Conv1D(
activation="relu",
padding="valid",
strides=1,
filters=params['nb_filter'],
kernel_size=params['filter_length']))
model.add(MaxPooling1D(strides=params['stride'], pool_size=params['pool_length']))
model.add(Flatten())
return model

def merge_outputs(outputs, name):
if len(outputs) == 1:
return outputs[0]
## return merge(outputs, mode='concat', name=name, concat_axis=1)
return Concatenate(axis=1, name=name)(outputs)


def merge_nets(nets, name):
if len(nets) == 1:
return nets[0]
## return merge(nets, mode='sum', name=name)
return Add(name=name)(nets)

def get_node_name(go_id, unique=False):
name = go_id.split(':')[1]
if not unique:
return name
if name not in node_names:
node_names.add(name)
return name
i = 1
while (name + '_' + str(i)) in node_names:
i += 1
name = name + '_' + str(i)
node_names.add(name)
return name


def get_layers(inputs):
q = deque()
layers = {}
name = get_node_name(GO_ID)
layers[GO_ID] = {'net': inputs}
for node_id in go[GO_ID]['children']:
if node_id in func_set:
q.append((node_id, inputs))
while len(q) > 0:
node_id, net = q.popleft()
parent_nets = [inputs]
# for p_id in get_parents(go, node_id):
# if p_id in func_set:
# parent_nets.append(layers[p_id]['net'])
# if len(parent_nets) > 1:
# name = get_node_name(node_id) + '_parents'
# net = merge(
# parent_nets, mode='concat', concat_axis=1, name=name)
name = get_node_name(node_id)
net, output = get_function_node(name, inputs)
if node_id not in layers:
layers[node_id] = {'net': net, 'output': output}
for n_id in go[node_id]['children']:
if n_id in func_set and n_id not in layers:
ok = True
for p_id in get_parents(go, n_id):
if p_id in func_set and p_id not in layers:
ok = False
if ok:
q.append((n_id, net))

for node_id in functions:
childs = set(go[node_id]['children']).intersection(func_set)
if len(childs) > 0:
outputs = [layers[node_id]['output']]
for ch_id in childs:
outputs.append(layers[ch_id]['output'])
name = get_node_name(node_id) + '_max'
## layers[node_id]['output'] = merge(
## outputs, mode='max', name=name)
layers[node_id]['output'] = Maximum(name=name)(outputs)
return layers

def get_function_node(name, inputs):
output_name = name + '_out'
# net = Dense(256, name=name, activation='relu')(inputs)
output = Dense(1, name=output_name, activation='sigmoid')(inputs)
return output, output




def get_generator(params, n_classes):
inputs = Input(shape=(MAXLEN,), dtype='int32', name='input1')
feature_model = get_feature_model(params)(inputs)
net = Dense(300, activation='relu')(feature_model)
net = BatchNormalization()(net)
layers = get_layers(net)
output_models = []
for i in range(len(functions)):
output_models.append(layers[functions[i]]['output'])
net = Concatenate(axis=1)(output_models)
output = Dense(n_classes, activation='sigmoid')(net)
model = Model(inputs=inputs, outputs=output)

return model




def get_discriminator(params, n_classes, dropout_rate=0.5):
inputs = Input(shape=(n_classes, ))
inputs2 = Input(shape =(MAXLEN,), dtype ='int32', name='d_input2')
x2 = Embedding(8001,128, input_length=MAXLEN)(inputs2)
x2 = Conv1D(filters =1 , kernel_size= 1, padding = 'valid', activation ='relu', strides=1)(x2)
x2 = Lambda(lambda x: K.squeeze(x, 2))(x2)

#for i in range(params['nb_conv']):
# x2 = Conv1D ( activation="relu", padding="valid", strides=1, filters=params['nb_filter'],kernel_size=params['filter_length'])(x2)

#x2 =MaxPooling1D(strides=params['stride'], pool_size=params['pool_length'])(x2)
#x2 = Flatten()(x2)
size = 40
x = inputs
x = Dropout(dropout_rate)(x)
x = Dense(size)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)


size = 40
x2 = Dropout(dropout_rate)(x2)
x2 = Dense(size)(x2)
x2 = BatchNormalization()(x2)
x2 = Activation('relu')(x2)


x = Concatenate(axis =1 , name = 'merged2')([x, x2])
layer_sizes = [80, 40,30]
for size in layer_sizes:
x = Dropout(dropout_rate)(x)
x = Dense(size)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)


outputs = Dense(1)(x)
model = Model(inputs = [inputs ,inputs2], outputs=outputs, name='Discriminator')

return model



def get_model(params,nb_classes, batch_size, GRADIENT_PENALTY_WEIGHT=10):
generator = get_generator(params, nb_classes)
discriminator = get_discriminator(params, nb_classes)

generator_model, discriminator_model = \
WGAN_wrapper(generator=generator,
discriminator=discriminator,
generator_input_shape=(MAXLEN,),
discriminator_input_shape=(nb_classes,),
discriminator_input_shape2 = (MAXLEN, ),
batch_size=batch_size,
gradient_penalty_weight=GRADIENT_PENALTY_WEIGHT)

logging.info('Compilation finished')
return generator_model, discriminator_model


def train_wgan(generator_model, discriminator_model, batch_size, epochs,
x_train, y_train, x_val, y_val, generator_model_path, discriminator_model_path,
TRAINING_RATIO=10, N_WARM_UP=0):
BATCH_SIZE = batch_size
N_EPOCH = epochs
positive_y = np.ones((batch_size, 1), dtype=np.float32)
zero_y = positive_y * 0
negative_y = -positive_y
positive_full_y = np.ones((BATCH_SIZE * TRAINING_RATIO, 1), dtype=np.float32)
dummy_y = np.zeros((BATCH_SIZE, 1), dtype=np.float32)
positive_full_enable_train = np.ones((len(x_train), 1), dtype = np.float32 )
positive_full_enable_val = np.ones((len(x_val), 1), dtype =np.float32 )
#positive_enable_train = np.ones((1, batch_size),dtype = np.float32 )
#positive_full_train_enable = np.ones((1,BATCH_SIZE * TRAINING_RATIO ), dtype=np.float32 )
best_validation_loss = None

for epoch in range(N_EPOCH):
# np.random.shuffle(X_train)
print("Epoch: ", epoch)
print("Number of batches: ", int(y_train.shape[0] // BATCH_SIZE))
discriminator_loss = []
generator_loss = []
minibatches_size = BATCH_SIZE * TRAINING_RATIO

shuffled_indexes = np.random.permutation(x_train.shape[0])
shuffled_indexes_2 = np.random.permutation(x_train.shape[0])

for i in range(int(y_train.shape[0] // (BATCH_SIZE * TRAINING_RATIO))):
batch_indexes = shuffled_indexes[i * minibatches_size:(i + 1) * minibatches_size]
batch_indexes_2 = shuffled_indexes_2[i * minibatches_size:(i + 1) * minibatches_size]
x = x_train[batch_indexes]
y = y_train[batch_indexes]
y_2 = y_train[batch_indexes_2]
x_2 = x_train[batch_indexes_2]
if epoch < N_WARM_UP:
for j in range(TRAINING_RATIO):
x_batch = x[j * BATCH_SIZE:(j + 1) * BATCH_SIZE]
y_batch = y[j * BATCH_SIZE:(j + 1) * BATCH_SIZE]

generator_loss.append(generator_model.train_on_batch([x_batch, positive_y], [y_batch, zero_y]))
else:
for j in range(TRAINING_RATIO):
x_batch = x[j * BATCH_SIZE:(j + 1) * BATCH_SIZE]
y_batch_2 = y_2[j * BATCH_SIZE:(j + 1) * BATCH_SIZE]
x_batch_2 = x_2[j * BATCH_SIZE:(j + 1) * BATCH_SIZE]
# noise = np.random.rand(BATCH_SIZE, 100).astype(np.float32)
noise = x_batch
#print(sum(y_batch_2))
discriminator_loss.append(discriminator_model.train_on_batch(
[y_batch_2, noise, x_batch_2 ],
[positive_y, negative_y, dummy_y]))

generator_loss.append(generator_model.train_on_batch([x,positive_full_y], [y, positive_full_y]))
# Still needs some code to display losses from the generator and discriminator, progress bars, etc.
predicted_y_train, _ = generator_model.predict([x_train , positive_full_enable_train], batch_size=BATCH_SIZE)
predicted_y_val, _ = generator_model.predict([ x_val , positive_full_enable_val ], batch_size=BATCH_SIZE)
#print(sum(sum(positive_full_enable_train)))
#print(predicted_y_train)
train_loss = log_loss(y_train, predicted_y_train)
val_loss = log_loss(y_val, predicted_y_val)

print("train loss: {:.4f}, validation loss: {:.4f}, discriminator loss: {:.4f}".format(
train_loss, val_loss,
(np.sum(np.asarray(discriminator_loss)) if discriminator_loss else -1) / x_train.shape[0]))

if best_validation_loss is None or best_validation_loss > val_loss:
print('\nEpoch %05d: improved from %0.5f,'
' saving model to %s and %s'
% (epoch + 1, val_loss, generator_model_path, discriminator_model_path))

best_validation_loss = val_loss
generator_model.save(generator_model_path, overwrite=True)
discriminator_model.save(discriminator_model_path, overwrite=True)


def model(params, batch_size=20, nb_epoch=40, is_train=True):
# set parameters:
nb_classes = len(functions)
start_time = time.time()
logging.info("Loading Data")

train, val, test, train_df, valid_df, test_df = load_data()
train_df = pd.concat([train_df, valid_df])
test_gos = test_df['gos'].values
train_data, train_labels = train
val_data, val_labels = val
test_data, test_labels = test


logging.info("Data loaded in %d sec" % (time.time() - start_time))
logging.info("Training data size: %d" % len(train_data))
logging.info("Validation data size: %d" % len(val_data))
logging.info("Test data size: %d" % len(test_data))
generator_model_path = DATA_ROOT + 'models/new_model_seq_' + FUNCTION + '.h5'
discriminator_model_path = DATA_ROOT + 'models/new_model_disc_seq_' + FUNCTION + '.h5'


logging.info('Starting training the model')

train_generator = DataGenerator(batch_size, nb_classes)
train_generator.fit(train_data, train_labels)
valid_generator = DataGenerator(batch_size, nb_classes)
valid_generator.fit(val_data, val_labels)
test_generator = DataGenerator(batch_size, nb_classes)
test_generator.fit(test_data, test_labels)

if is_train:
generator_model, discriminator_model = get_model(params, nb_classes, batch_size)
train_wgan(generator_model, discriminator_model, batch_size=batch_size, epochs=nb_epoch,
x_train=train_data, y_train=train_labels, x_val=val_data, y_val=val_labels,
generator_model_path=generator_model_path,
discriminator_model_path=discriminator_model_path)

logging.info('Loading best model')
model = load_model(generator_model_path,
custom_objects={'generator_recunstruction_loss_new': generator_recunstruction_loss_new,
'wasserstein_loss': wasserstein_loss})

logging.info('Predicting')
preds = model.predict_generator(test_generator, steps=len(test_data) / batch_size)[0]

logging.info('Computing performance')
f, p, r, t, preds_max = compute_performance(preds, test_labels) #, test_gos)
roc_auc = compute_roc(preds, test_labels)
mcc = compute_mcc(preds_max, test_labels)
aupr , _ = compute_aupr(preds, test_labels)
m_pr_max, m_rc_max, m_f1_max, M_pr_max, M_rc_max, M_f1_max = micro_macro_function_centric_f1(preds.T, test_labels.T)

logging.info('Protein centric macro Th, PR, RC, F1: \t %f %f %f %f' % (t, p, r, f))
logging.info('ROC AUC: \t %f ' % (roc_auc, ))
logging.info('MCC: \t %f ' % (mcc, ))
logging.info('AUPR: \t %f ' % (aupr, ))
logging.info('Function centric macro PR, RC, F1: \t %f %f %f' % (M_pr_max, M_rc_max, M_f1_max) )
logging.info('Function centric micro PR, RC, F1: \t %f %f %f' % (m_pr_max, m_rc_max, m_f1_max) )

function_centric_performance(functions, preds.T, test_labels.T, train_labels.T)



def load_prot_ipro():
proteins = list()
ipros = list()
with open(DATA_ROOT + 'swissprot_ipro.tab') as f:
for line in f:
it = line.strip().split('\t')
if len(it) != 3:
continue
prot = it[1]
iprs = it[2].split(';')
proteins.append(prot)
ipros.append(iprs)
return pd.DataFrame({'proteins': proteins, 'ipros': ipros})


def performanc_by_interpro():
pred_df = pd.read_pickle(DATA_ROOT + 'test-' + FUNCTION + '-preds.pkl')
ipro_df = load_prot_ipro()
df = pred_df.merge(ipro_df, on='proteins', how='left')
ipro = get_ipro()

def reshape(values):
values = np.hstack(values).reshape(
len(values), len(values[0]))
return values

for ipro_id in ipro:
if len(ipro[ipro_id]['parents']) > 0:
continue
labels = list()
predictions = list()
gos = list()
for i, row in df.iterrows():
if not isinstance(row['ipros'], list):
continue
if ipro_id in row['ipros']:
labels.append(row['labels'])
predictions.append(row['predictions'])
gos.append(row['gos'])
pr = 0
rc = 0
total = 0
p_total = 0
for i in range(len(labels)):
tp = np.sum(labels[i] * predictions[i])
fp = np.sum(predictions[i]) - tp
fn = np.sum(labels[i]) - tp
all_gos = set()
for go_id in gos[i]:
if go_id in all_functions:
all_gos |= get_anchestors(go, go_id)
all_gos.discard(GO_ID)
all_gos -= func_set
fn += len(all_gos)
if tp == 0 and fp == 0 and fn == 0:
continue
total += 1
if tp != 0:
p_total += 1
precision = tp / (1.0 * (tp + fp))
recall = tp / (1.0 * (tp + fn))
pr += precision
rc += recall
if total > 0 and p_total > 0:
rc /= total
pr /= p_total
if pr + rc > 0:
f = 2 * pr * rc / (pr + rc)
logging.info('%s\t%d\t%f\t%f\t%f' % (
ipro_id, len(labels), f, pr, rc))


def function_centric_performance(functions, preds, labels, labels_train):
results = []
preds = np.round(preds, 2)
for i in range(preds.shape[0]):
f_max = 0
p_max = 0
r_max = 0
for t in range(1, 100):
threshold = t / 100.0
predictions = (preds[i, :] > threshold).astype(np.int32)
tp = np.sum(predictions * labels[i, :])
fp = np.sum(predictions) - tp
fn = np.sum(labels[i, :]) - tp
if tp > 0:
precision = tp / (1.0 * (tp + fp))
recall = tp / (1.0 * (tp + fn))
f = 2 * precision * recall / (precision + recall)
else:
if fp == 0 and fn == 0:
precision = 1
recall = 1
f = 1
else:
precision = 0
recall = 0
f = 0

if f_max < f:
f_max = f
p_max = precision
r_max = recall
num_prots_train = np.sum(labels_train[i, :])
height = get_height(go, functions[i])
results.append([functions[i], num_prots_train, height, f_max, p_max, r_max])
results = pd.DataFrame(results)
results.to_csv('Con_GodGanSeq_results_' + FUNCTION + '.txt', sep='\t', index=False)



def function_centric_performance_backup(functions, preds, labels, labels_train):
results = []
preds = np.round(preds, 2)
for i in range(len(functions)):
f_max = 0
p_max = 0
r_max = 0
x = list()
y = list()
total = 0
for t in range(1, 100):
threshold = t / 100.0
predictions = (preds[i, :] > threshold).astype(np.int32)
tp = np.sum(predictions * labels[i, :])
fp = np.sum(predictions) - tp
fn = np.sum(labels[i, :]) - tp
if tp >0:
sn = tp / (1.0 * np.sum(labels[i, :]))
sp = np.sum((predictions ^ 1) * (labels[i, :] ^ 1))
sp /= 1.0 * np.sum(labels[i, :] ^ 1)
fpr = 1 - sp
x.append(fpr)
y.append(sn)
precision = tp / (1.0 * (tp + fp))
recall = tp / (1.0 * (tp + fn))
f = 2 * precision * recall / (precision + recall)
total +=1
if f_max < f:
f_max = f
p_max = precision
r_max = recall
num_prots = np.sum(labels[i, :])
num_prots_train = np.sum(labels_train[i,:])
if total >1 :
roc_auc = auc(x, y)
else:
roc_auc =0
height = get_height(go , functions[i])
results.append([functions[i], f_max, p_max, r_max, num_prots, num_prots_train, height,roc_auc])
results = pd.DataFrame(results)
#results.to_csv('new_results.txt' , sep='\t' , index = False)
results.to_csv('Con_GodGanSeq_results_'+FUNCTION +'.txt', sep='\t', index=False)
#results = np.array(results)
#p_mean = (np.sum(results[:,2])) / len(functions)
#r_mean = (np.sum(results[:,3])) / len(functions)
#f_mean = (2*p_mean*r_mean)/(p_mean+r_mean)
#roc_auc_mean = (np.sum(results[:,7])) / len(functions)
#print('Function centric performance (macro) ' '%f %f %f %f' % (f_mean, p_mean, r_mean, roc_auc_mean))



def micro_macro_function_centric_f1_backup(preds, labels):
preds = np.round(preds, 2)
m_f1_max = 0
M_f1_max = 0
for t in range(1, 100):
threshold = t / 100.0
predictions = (preds > threshold).astype(np.int32)
m_tp = 0
m_fp = 0
m_fn = 0
M_pr = 0
M_rc = 0
total = 0
p_total = 0
for i in range(len(preds)):
tp = np.sum(predictions[i, :] * labels[i, :])
fp = np.sum(predictions[i, :]) - tp
fn = np.sum(labels[i, :]) - tp
if tp == 0 and fp == 0 and fn == 0:
continue
total += 1
if tp > 0:
pr = tp / (1.0 * (tp + fp))
rc = tp / (1.0 * (tp + fn))
m_tp += tp
m_fp += fp
m_fn += fn
M_pr += pr
M_rc += rc
p_total += 1

if p_total == 0:
continue
if total > 0:
m_tp /= total
m_fn /= total
m_fp /= total
m_pr = m_tp / (1.0 * (m_tp + m_fp))
m_rc = m_tp / (1.0 * (m_tp + m_fn))
M_pr /= p_total
M_rc /= total
m_f1 = 2 * m_pr * m_rc / (m_pr + m_rc)
M_f1 = 2 * M_pr * M_rc / (M_pr + M_rc)

if m_f1 > m_f1_max:
m_f1_max = m_f1
m_pr_max = m_pr
m_rc_max = m_rc

if M_f1 > M_f1_max:
M_f1_max = M_f1
M_pr_max = M_pr
M_rc_max = M_rc

return m_pr_max, m_rc_max, m_f1_max, M_pr_max, M_rc_max, M_f1_max





def micro_macro_function_centric_f1(preds, labels):
preds = np.round(preds, 2)
m_f1_max = 0
M_f1_max = 0
for t in range(1, 200):
threshold = t / 200.0
predictions = (preds > threshold).astype(np.int32)
m_tp = 0
m_fp = 0
m_fn = 0
M_pr = 0
M_rc = 0
for i in range(preds.shape[0]):
tp = np.sum(predictions[i, :] * labels[i, :])
fp = np.sum(predictions[i, :]) - tp
fn = np.sum(labels[i, :]) - tp
m_tp += tp
m_fp += fp
m_fn += fn
if tp > 0:
pr = 1.0 * tp / (1.0 * (tp + fp))
rc = 1.0 * tp / (1.0 * (tp + fn))
else:
if fp == 0 and fn == 0:
pr = 1
rc = 1
else:
pr = 0
rc = 0
M_pr += pr
M_rc += rc

if m_tp > 0:
m_pr = 1.0 * m_tp / (1.0 * (m_tp + m_fp))
m_rc = 1.0 * m_tp / (1.0 * (m_tp + m_fn))
m_f1 = 2.0 * m_pr * m_rc / (m_pr + m_rc)
else:
if m_fp == 0 and m_fn == 0:
m_pr = 1
m_rc = 1
m_f1 = 1
else:
m_pr = 0
m_rc = 0
m_f1 = 0

M_pr /= preds.shape[0]
M_rc /= preds.shape[0]
if M_pr == 0 and M_rc == 0:
M_f1 = 0
else:
M_f1 = 2.0 * M_pr * M_rc / (M_pr + M_rc)
if m_f1 > m_f1_max:
m_f1_max = m_f1
m_pr_max = m_pr
m_rc_max = m_rc

if M_f1 > M_f1_max:
M_f1_max = M_f1
M_pr_max = M_pr
M_rc_max = M_rc

return m_pr_max, m_rc_max, m_f1_max, M_pr_max, M_rc_max, M_f1_max


def compute_roc(preds, labels):
# Compute ROC curve and ROC area for each class
fpr, tpr, _ = roc_curve(labels.flatten(), preds.flatten())
roc_auc = auc(fpr, tpr)
return roc_auc


def compute_aupr(preds, labels):
# Compute ROC curve and ROC area for each class
pr, rc, threshold =precision_recall_curve(labels.flatten(), preds.flatten())
pr_auc = auc(rc, pr)
#pr, rc, threshold =precision_recall_curve(labels.flatten(), preds.flatten(),average ='macro' )
M_pr_auc = 0
return pr_auc, M_pr_auc



def compute_mcc(preds, labels):
# Compute ROC curve and ROC area for each class
mcc = matthews_corrcoef(labels.flatten(), preds.flatten())
return mcc


def compute_performance(preds, labels): #, gos):
preds = np.round(preds, 2)
f_max = 0
p_max = 0
r_max = 0
t_max = 0
for t in range(1, 100):
threshold = t / 100.0
predictions = (preds > threshold).astype(np.int32)
total = 0
f = 0.0
p = 0.0
r = 0.0
p_total = 0
for i in range(labels.shape[0]):
tp = np.sum(predictions[i, :] * labels[i, :])
fp = np.sum(predictions[i, :]) - tp
fn = np.sum(labels[i, :]) - tp
all_gos = set()
#for go_id in gos[i]:
# if go_id in all_functions:
# all_gos |= get_anchestors(go, go_id)
#all_gos.discard(GO_ID)
#all_gos -= func_set
#fn += len(all_gos)
if tp == 0 and fp == 0 and fn == 0:
continue
total += 1
if tp != 0:
p_total += 1
precision = tp / (1.0 * (tp + fp))
recall = tp / (1.0 * (tp + fn))
p += precision
r += recall
if p_total == 0:
continue
r /= total
p /= p_total
if p + r > 0:
f = 2 * p * r / (p + r)
if f_max < f:
f_max = f
p_max = p
r_max = r
t_max = threshold
predictions_max = predictions
return f_max, p_max, r_max, t_max, predictions_max


def get_gos(pred):
mdist = 1.0
mgos = None
for i in range(len(labels_gos)):
labels, gos = labels_gos[i]
dist = distance.cosine(pred, labels)
if mdist > dist:
mdist = dist
mgos = gos
return mgos


def compute_similarity_performance(train_df, test_df, preds):
logging.info("Computing similarity performance")
logging.info("Training data size %d" % len(train_df))
train_labels = train_df['labels'].values
train_gos = train_df['gos'].values
global labels_gos
labels_gos = zip(train_labels, train_gos)
p = Pool(64)
pred_gos = p.map(get_gos, preds)
total = 0
p = 0.0
r = 0.0
f = 0.0
test_gos = test_df['gos'].values
for gos, tgos in zip(pred_gos, test_gos):
preds = set()
test = set()
for go_id in gos:
if go_id in all_functions:
preds |= get_anchestors(go, go_id)
for go_id in tgos:
if go_id in all_functions:
test |= get_anchestors(go, go_id)
tp = len(preds.intersection(test))
fp = len(preds - test)
fn = len(test - preds)
if tp == 0 and fp == 0 and fn == 0:
continue
total += 1
if tp != 0:
precision = tp / (1.0 * (tp + fp))
recall = tp / (1.0 * (tp + fn))
p += precision
r += recall
f += 2 * precision * recall / (precision + recall)
return f / total, p / total, r / total


def print_report(report, go_id):
with open(DATA_ROOT + 'reports.txt', 'a') as f:
f.write('Classification report for ' + go_id + '\n')
f.write(report + '\n')


if __name__ == '__main__':
main()

+ 0
- 917
conditional_new_nn_hirearchical_seq_mt.py View File

#!/usr/bin/env python

"""

"""
from __future__ import division

import logging
import sys
import time
from collections import deque
from multiprocessing import Pool

import click as ck
import numpy as np
import pandas as pd
import tensorflow as tf
from keras import backend as K
from keras.callbacks import EarlyStopping, ModelCheckpoint
from keras.layers import (
Dense, Input, SpatialDropout1D, Conv1D, MaxPooling1D,
Flatten, Concatenate, Add, Maximum, Embedding, BatchNormalization, Activation, Dropout)
from keras.losses import binary_crossentropy
from keras.models import Sequential, Model, load_model
from keras.preprocessing import sequence
from scipy.spatial import distance
from sklearn.metrics import log_loss
from sklearn.metrics import roc_curve, auc, matthews_corrcoef
from keras.layers import Lambda
from sklearn.metrics import precision_recall_curve

from utils import (
get_gene_ontology,
get_go_set,
get_anchestors,
get_parents,
DataGenerator,
FUNC_DICT,
get_height,
get_ipro)
from conditional_wgan_wrapper_post import WGAN_wrapper, wasserstein_loss, generator_recunstruction_loss_new

config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
K.set_session(sess)

logging.basicConfig(format='%(levelname)s:%(message)s', level=logging.INFO)
sys.setrecursionlimit(100000)

DATA_ROOT = 'data/swiss/'
MAXLEN = 258 #1000
REPLEN = 256
ind = 0


@ck.command()
@ck.option(
'--function',
default='bp',
help='Ontology id (mf, bp, cc)')
@ck.option(
'--device',
default='gpu:0',
help='GPU or CPU device id')
@ck.option(
'--org',
default= None,
help='Organism id for filtering test set')
@ck.option('--train',default = True, is_flag=True)
@ck.option('--param', default=0, help='Param index 0-7')
def main(function, device, org, train, param):
global FUNCTION
FUNCTION = function
global GO_ID
GO_ID = FUNC_DICT[FUNCTION]
global go
go = get_gene_ontology('go.obo')
global ORG
ORG = org
func_df = pd.read_pickle(DATA_ROOT + FUNCTION + '.pkl')
global functions
functions = func_df['functions'].values
global func_set
func_set = set(functions)
global all_functions
all_functions = get_go_set(go, GO_ID)
logging.info('Functions: %s %d' % (FUNCTION, len(functions)))
if ORG is not None:
logging.info('Organism %s' % ORG)
global go_indexes
go_indexes = dict()
for ind, go_id in enumerate(functions):
go_indexes[go_id] = ind
global node_names
node_names = set()
with tf.device('/' + device):
params = {
'fc_output': 1024,
'learning_rate': 0.001,
'embedding_dims': 128,
'embedding_dropout': 0.2,
'nb_conv': 1,
'nb_dense': 1,
'filter_length': 128,
'nb_filter': 32,
'pool_length': 64,
'stride': 32
}
model(params, is_train=train)
#dims = [64, 128, 256, 512]
#nb_filters = [16, 32, 64, 128]
#nb_convs = [1, 2, 3, 4]
#nb_dense = [1, 2, 3, 4]
#for i in range(param * 32, param * 32 + 32):
# dim = i % 4
# i = i / 4
# nb_fil = i % 4
# i /= 4
# conv = i % 4
# i /= 4
# den = i
# params['embedding_dims'] = dims[dim]
# params['nb_filter'] = nb_filters[nb_fil]
# params['nb_conv'] = nb_convs[conv]
# params['nb_dense'] = nb_dense[den]

# performanc_by_interpro()


def load_data2():
all_data_x_fn = 'data2/all_data_X.csv'
all_data_x = pd.read_csv(all_data_x_fn, sep='\t', header=0, index_col=0)
all_proteins_train = [p.replace('"', '') for p in all_data_x.index]
all_data_x.index = all_proteins_train
all_data_y_fn = 'data2/all_data_Y.csv'
all_data_y = pd.read_csv(all_data_y_fn, sep='\t', header=0, index_col=0)
branch = pd.read_csv('data2/'+FUNCTION +'_branches.txt', sep='\t', header=0, index_col=0)
all_x = all_data_x.values
branches = [p for p in branch.index.tolist() if p in all_data_y.columns.tolist()]
t= pd.DataFrame(all_data_y, columns=branches)
all_y = t.values
number_of_test = int(np.ceil(0.2 * len(all_x)))
index = np.random.rand(1,number_of_test)
index_test = [int(p) for p in np.ceil(index*len(all_x))[0] ]
index_train = [p for p in range(len(all_x)) if p not in index_test]
train_data = all_x[index_train, : ] #[ :20000, : ]
test_data = all_x[index_test, : ] #[20000: , : ]
train_labels = all_y[index_train, : ] #[ :20000, : ]
test_labels = all_y[index_test, :] #[20000: , : ]
val_data = test_data
val_labels = test_labels
#print(sum(sum(train_labels)))
#print(train_data.shape)
print(train_labels.shape)
print(test_labels.shape)
return train_data, train_labels, test_data, test_labels, val_data, val_labels




def load_data():

df = pd.read_pickle(DATA_ROOT + 'train' + '-' + FUNCTION + '.pkl')
n = len(df)
index = df.index.values
valid_n = int(n * 0.8)
train_df = df.loc[index[:valid_n]]
valid_df = df.loc[index[valid_n:]]
test_df = pd.read_pickle(DATA_ROOT + 'test' + '-' + FUNCTION + '.pkl')
print( test_df['orgs'] )
if ORG is not None:
logging.info('Unfiltered test size: %d' % len(test_df))
test_df = test_df[test_df['orgs'] == ORG]
logging.info('Filtered test size: %d' % len(test_df))

# Filter by type
# org_df = pd.read_pickle('data/prokaryotes.pkl')
# orgs = org_df['orgs']
# test_df = test_df[test_df['orgs'].isin(orgs)]

def reshape(values):
values = np.hstack(values).reshape(
len(values), len(values[0]))
return values

def normalize_minmax(values):
mn = np.min(values)
mx = np.max(values)
if mx - mn != 0.0:
return (values - mn) / (mx - mn)
return values - mn

def get_values(data_frame):
print(data_frame['labels'].values.shape)
labels = reshape(data_frame['labels'].values)
ngrams = sequence.pad_sequences(
data_frame['ngrams'].values, maxlen=MAXLEN)
ngrams = reshape(ngrams)
rep = reshape(data_frame['embeddings'].values)
data = ngrams
return data, labels

train = get_values(train_df)
valid = get_values(valid_df)
test = get_values(test_df)

return train, valid, test, train_df, valid_df, test_df


def get_feature_model(params):
embedding_dims = params['embedding_dims']
max_features = 8001
model = Sequential()
model.add(Embedding(
max_features,
embedding_dims,
input_length=MAXLEN))
model.add(SpatialDropout1D(0.4))
for i in range(params['nb_conv']):
model.add(Conv1D(
activation="relu",
padding="valid",
strides=1,
filters=params['nb_filter'],
kernel_size=params['filter_length']))
model.add(MaxPooling1D(strides=params['stride'], pool_size=params['pool_length']))
model.add(Flatten())
return model


def merge_outputs(outputs, name):
if len(outputs) == 1:
return outputs[0]
## return merge(outputs, mode='concat', name=name, concat_axis=1)
return Concatenate(axis=1, name=name)(outputs)


def merge_nets(nets, name):
if len(nets) == 1:
return nets[0]
## return merge(nets, mode='sum', name=name)
return Add(name=name)(nets)

def get_node_name(go_id, unique=False):
name = go_id.split(':')[1]
if not unique:
return name
if name not in node_names:
node_names.add(name)
return name
i = 1
while (name + '_' + str(i)) in node_names:
i += 1
name = name + '_' + str(i)
node_names.add(name)
return name


def get_function_node(name, inputs):
output_name = name + '_out'
# net = Dense(256, name=name, activation='relu')(inputs)
output = Dense(1, name=output_name, activation='sigmoid')(inputs)
return output, output


def get_generator(params, n_classes):
inputs = Input(shape=(MAXLEN,), dtype='float32', name='input1')
#feature_model = get_feature_model(params)(inputs)
net0 = Dense(150, activation='relu')(inputs)
net0 = Dense(150, activation='relu')(net0)
#net0 = Dense(50, activation='relu')(net0)
net = Dense(70, activation = 'relu')(net0)
output = Dense(n_classes, activation='sigmoid')(net)
model = Model(inputs=inputs, outputs=output)

return model


def get_discriminator(params, n_classes, dropout_rate=0.5):
inputs = Input(shape=(n_classes, ))
inputs2 = Input(shape =(MAXLEN,), dtype ='int32', name='d_input2')
x2 = Embedding(8001,128, input_length=MAXLEN)(inputs2)
x2 = Conv1D(filters =1 , kernel_size= 1, padding = 'valid', activation ='relu', strides=1)(x2)
x2 = Lambda(lambda x: K.squeeze(x, 2))(x2)

#for i in range(params['nb_conv']):
# x2 = Conv1D ( activation="relu", padding="valid", strides=1, filters=params['nb_filter'],kernel_size=params['filter_length'])(x2)

#x2 =MaxPooling1D(strides=params['stride'], pool_size=params['pool_length'])(x2)
#x2 = Flatten()(x2)
size = 40
x = inputs
x = Dropout(dropout_rate)(x)
x = Dense(size)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)


size = 40
x2 = Dropout(dropout_rate)(x2)
x2 = Dense(size)(x2)
x2 = BatchNormalization()(x2)
x2 = Activation('relu')(x2)


x = Concatenate(axis =1 , name = 'merged2')([x, x2])
layer_sizes = [80, 40,30]
for size in layer_sizes:
x = Dropout(dropout_rate)(x)
x = Dense(size)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)


outputs = Dense(1)(x)
model = Model(inputs = [inputs ,inputs2], outputs=outputs, name='Discriminator')

return model



def get_model(params,nb_classes, batch_size, GRADIENT_PENALTY_WEIGHT=10):
generator = get_generator(params, nb_classes)
discriminator = get_discriminator(params, nb_classes)

generator_model, discriminator_model = \
WGAN_wrapper(generator=generator,
discriminator=discriminator,
generator_input_shape=(MAXLEN,),
discriminator_input_shape=(nb_classes,),
discriminator_input_shape2 = (MAXLEN, ),
batch_size=batch_size,
gradient_penalty_weight=GRADIENT_PENALTY_WEIGHT)

logging.info('Compilation finished')
return generator_model, discriminator_model


def train_wgan(generator_model, discriminator_model, batch_size, epochs,
x_train, y_train, x_val, y_val, generator_model_path, discriminator_model_path,
TRAINING_RATIO=10, N_WARM_UP=0):
BATCH_SIZE = batch_size
N_EPOCH = epochs
positive_y = np.ones((batch_size, 1), dtype=np.float32)
zero_y = positive_y * 0
negative_y = -positive_y
positive_full_y = np.ones((BATCH_SIZE * TRAINING_RATIO, 1), dtype=np.float32)
dummy_y = np.zeros((BATCH_SIZE, 1), dtype=np.float32)
positive_full_enable_train = np.ones((len(x_train), 1), dtype = np.float32 )
positive_full_enable_val = np.ones((len(x_val), 1), dtype =np.float32 )
#positive_enable_train = np.ones((1, batch_size),dtype = np.float32 )
#positive_full_train_enable = np.ones((1,BATCH_SIZE * TRAINING_RATIO ), dtype=np.float32 )
best_validation_loss = None

for epoch in range(N_EPOCH):
# np.random.shuffle(X_train)
print("Epoch: ", epoch)
print("Number of batches: ", int(y_train.shape[0] // BATCH_SIZE))
discriminator_loss = []
generator_loss = []
minibatches_size = BATCH_SIZE * TRAINING_RATIO

shuffled_indexes = np.random.permutation(x_train.shape[0])
shuffled_indexes_2 = np.random.permutation(x_train.shape[0])

for i in range(int(y_train.shape[0] // (BATCH_SIZE * TRAINING_RATIO))):
batch_indexes = shuffled_indexes[i * minibatches_size:(i + 1) * minibatches_size]
batch_indexes_2 = shuffled_indexes_2[i * minibatches_size:(i + 1) * minibatches_size]
x = x_train[batch_indexes]
y = y_train[batch_indexes]
y_2 = y_train[batch_indexes_2]
x_2 = x_train[batch_indexes_2]
if epoch < N_WARM_UP:
for j in range(TRAINING_RATIO):
x_batch = x[j * BATCH_SIZE:(j + 1) * BATCH_SIZE]
y_batch = y[j * BATCH_SIZE:(j + 1) * BATCH_SIZE]

generator_loss.append(generator_model.train_on_batch([x_batch, positive_y], [y_batch, zero_y]))
else:
for j in range(TRAINING_RATIO):
x_batch = x[j * BATCH_SIZE:(j + 1) * BATCH_SIZE]
y_batch_2 = y_2[j * BATCH_SIZE:(j + 1) * BATCH_SIZE]
x_batch_2 = x_2[j * BATCH_SIZE:(j + 1) * BATCH_SIZE]
# noise = np.random.rand(BATCH_SIZE, 100).astype(np.float32)
noise = x_batch
#print(sum(y_batch_2))
discriminator_loss.append(discriminator_model.train_on_batch(
[y_batch_2, noise, x_batch_2 ],
[positive_y, negative_y, dummy_y]))

generator_loss.append(generator_model.train_on_batch([x,positive_full_y], [y, positive_full_y]))
# Still needs some code to display losses from the generator and discriminator, progress bars, etc.
predicted_y_train, _ = generator_model.predict([x_train , positive_full_enable_train], batch_size=BATCH_SIZE)
predicted_y_val, _ = generator_model.predict([ x_val , positive_full_enable_val ], batch_size=BATCH_SIZE)
#print(sum(sum(positive_full_enable_train)))
#print(predicted_y_train)
train_loss = log_loss(y_train, predicted_y_train)
val_loss = log_loss(y_val, predicted_y_val)

print("train loss: {:.4f}, validation loss: {:.4f}, discriminator loss: {:.4f}".format(
train_loss, val_loss,
(np.sum(np.asarray(discriminator_loss)) if discriminator_loss else -1) / x_train.shape[0]))

if best_validation_loss is None or best_validation_loss > val_loss:
print('\nEpoch %05d: improved from %0.5f,'
' saving model to %s and %s'
% (epoch + 1, val_loss, generator_model_path, discriminator_model_path))

best_validation_loss = val_loss
generator_model.save(generator_model_path, overwrite=True)
discriminator_model.save(discriminator_model_path, overwrite=True)


def model(params, batch_size=20, nb_epoch=40, is_train=True):
# set parameters:
#nb_classes = len(functions)
start_time = time.time()
logging.info("Loading Data")
##
#train, val, test, train_df, valid_df, test_df = load_data()
#train_df = pd.concat([train_df, valid_df])
#test_gos = test_df['gos'].values
#train_data, train_labels = train
#val_data, val_labels = val
#test_data, test_labels = test
##

train_data, train_labels, test_data, test_labels, val_data, val_labels = load_data2()
nb_classes = train_labels.shape[1]

logging.info("Data loaded in %d sec" % (time.time() - start_time))
logging.info("Training data size: %d" % len(train_data))
logging.info("Validation data size: %d" % len(val_data))
logging.info("Test data size: %d" % len(test_data))
generator_model_path = DATA_ROOT + 'models/new_model_seq_' + FUNCTION + '.h5'
discriminator_model_path = DATA_ROOT + 'models/new_model_disc_seq_' + FUNCTION + '.h5'


logging.info('Starting training the model')

train_generator = DataGenerator(batch_size, nb_classes)
train_generator.fit(train_data, train_labels)
valid_generator = DataGenerator(batch_size, nb_classes)
valid_generator.fit(val_data, val_labels)
test_generator = DataGenerator(batch_size, nb_classes)
test_generator.fit(test_data, test_labels)

if is_train:
generator_model, discriminator_model = get_model(params, nb_classes, batch_size)
train_wgan(generator_model, discriminator_model, batch_size=batch_size, epochs=nb_epoch,
x_train=train_data, y_train=train_labels, x_val=val_data, y_val=val_labels,
generator_model_path=generator_model_path,
discriminator_model_path=discriminator_model_path)

logging.info('Loading best model')
model = load_model(generator_model_path,
custom_objects={'generator_recunstruction_loss_new': generator_recunstruction_loss_new,
'wasserstein_loss': wasserstein_loss})

logging.info('Predicting')
preds = model.predict_generator(test_generator, steps=len(test_data) / batch_size)[0]
# incon = 0
# for i in xrange(len(test_data)):
# for j in xrange(len(functions)):
# childs = set(go[functions[j]]['children']).intersection(func_set)
# ok = True`
# for n_id in childs:
# if preds[i, j] < preds[i, go_indexes[n_id]]:
# preds[i, j] = preds[i, go_indexes[n_id]]
# ok = False
# if not ok:
# incon += 1
logging.info('Computing performance')
f, p, r, t, preds_max = compute_performance(preds, test_labels) #, test_gos)
roc_auc = compute_roc(preds, test_labels)
mcc = compute_mcc(preds_max, test_labels)
aupr , _ = compute_aupr(preds, test_labels)
m_pr_max, m_rc_max, m_f1_max, M_pr_max, M_rc_max, M_f1_max = micro_macro_function_centric_f1(preds.T, test_labels.T)

logging.info('Protein centric macro Th, PR, RC, F1: \t %f %f %f %f' % (t, p, r, f))
logging.info('ROC AUC: \t %f ' % (roc_auc, ))
logging.info('MCC: \t %f ' % (mcc, ))
logging.info('AUPR: \t %f ' % (aupr, ))
logging.info('Function centric macro PR, RC, F1: \t %f %f %f' % (M_pr_max, M_rc_max, M_f1_max) )
logging.info('Function centric micro PR, RC, F1: \t %f %f %f' % (m_pr_max, m_rc_max, m_f1_max) )

function_centric_performance(functions, preds.T, test_labels.T, train_labels.T)



def load_prot_ipro():
proteins = list()
ipros = list()
with open(DATA_ROOT + 'swissprot_ipro.tab') as f:
for line in f:
it = line.strip().split('\t')
if len(it) != 3:
continue
prot = it[1]
iprs = it[2].split(';')
proteins.append(prot)
ipros.append(iprs)
return pd.DataFrame({'proteins': proteins, 'ipros': ipros})


def performanc_by_interpro():
pred_df = pd.read_pickle(DATA_ROOT + 'test-' + FUNCTION + '-preds.pkl')
ipro_df = load_prot_ipro()
df = pred_df.merge(ipro_df, on='proteins', how='left')
ipro = get_ipro()

def reshape(values):
values = np.hstack(values).reshape(
len(values), len(values[0]))
return values

for ipro_id in ipro:
if len(ipro[ipro_id]['parents']) > 0:
continue
labels = list()
predictions = list()
gos = list()
for i, row in df.iterrows():
if not isinstance(row['ipros'], list):
continue
if ipro_id in row['ipros']:
labels.append(row['labels'])
predictions.append(row['predictions'])
gos.append(row['gos'])
pr = 0
rc = 0
total = 0
p_total = 0
for i in range(len(labels)):
tp = np.sum(labels[i] * predictions[i])
fp = np.sum(predictions[i]) - tp
fn = np.sum(labels[i]) - tp
all_gos = set()
for go_id in gos[i]:
if go_id in all_functions:
all_gos |= get_anchestors(go, go_id)
all_gos.discard(GO_ID)
all_gos -= func_set
fn += len(all_gos)
if tp == 0 and fp == 0 and fn == 0:
continue
total += 1
if tp != 0:
p_total += 1
precision = tp / (1.0 * (tp + fp))
recall = tp / (1.0 * (tp + fn))
pr += precision
rc += recall
if total > 0 and p_total > 0:
rc /= total
pr /= p_total
if pr + rc > 0:
f = 2 * pr * rc / (pr + rc)
logging.info('%s\t%d\t%f\t%f\t%f' % (
ipro_id, len(labels), f, pr, rc))


def function_centric_performance(functions, preds, labels, labels_train):
results = []
preds = np.round(preds, 2)
for i in range(preds.shape[0]):
f_max = 0
p_max = 0
r_max = 0
for t in range(1, 100):
threshold = t / 100.0
predictions = (preds[i, :] > threshold).astype(np.int32)
tp = np.sum(predictions * labels[i, :])
fp = np.sum(predictions) - tp
fn = np.sum(labels[i, :]) - tp
if tp > 0:
precision = tp / (1.0 * (tp + fp))
recall = tp / (1.0 * (tp + fn))
f = 2 * precision * recall / (precision + recall)
else:
if fp == 0 and fn == 0:
precision = 1
recall = 1
f = 1
else:
precision = 0
recall = 0
f = 0

if f_max < f:
f_max = f
p_max = precision
r_max = recall
num_prots_train = np.sum(labels_train[i, :])
height = get_height(go, functions[i])
results.append([functions[i], num_prots_train, height, f_max, p_max, r_max])
results = pd.DataFrame(results)
results.to_csv('Con_GodGanSeq_results_' + FUNCTION + '.txt', sep='\t', index=False)



def function_centric_performance_backup(functions, preds, labels, labels_train):
results = []
preds = np.round(preds, 2)
for i in range(len(functions)):
f_max = 0
p_max = 0
r_max = 0
x = list()
y = list()
total = 0
for t in range(1, 100):
threshold = t / 100.0
predictions = (preds[i, :] > threshold).astype(np.int32)
tp = np.sum(predictions * labels[i, :])
fp = np.sum(predictions) - tp
fn = np.sum(labels[i, :]) - tp
if tp >0:
sn = tp / (1.0 * np.sum(labels[i, :]))
sp = np.sum((predictions ^ 1) * (labels[i, :] ^ 1))
sp /= 1.0 * np.sum(labels[i, :] ^ 1)
fpr = 1 - sp
x.append(fpr)
y.append(sn)
precision = tp / (1.0 * (tp + fp))
recall = tp / (1.0 * (tp + fn))
f = 2 * precision * recall / (precision + recall)
total +=1
if f_max < f:
f_max = f
p_max = precision
r_max = recall
num_prots = np.sum(labels[i, :])
num_prots_train = np.sum(labels_train[i,:])
if total >1 :
roc_auc = auc(x, y)
else:
roc_auc =0
height = get_height(go , functions[i])
results.append([functions[i], f_max, p_max, r_max, num_prots, num_prots_train, height,roc_auc])
results = pd.DataFrame(results)
#results.to_csv('new_results.txt' , sep='\t' , index = False)
results.to_csv('Con_GodGanSeq_results_'+FUNCTION +'.txt', sep='\t', index=False)
#results = np.array(results)
#p_mean = (np.sum(results[:,2])) / len(functions)
#r_mean = (np.sum(results[:,3])) / len(functions)
#f_mean = (2*p_mean*r_mean)/(p_mean+r_mean)
#roc_auc_mean = (np.sum(results[:,7])) / len(functions)
#print('Function centric performance (macro) ' '%f %f %f %f' % (f_mean, p_mean, r_mean, roc_auc_mean))



def micro_macro_function_centric_f1_backup(preds, labels):
preds = np.round(preds, 2)
m_f1_max = 0
M_f1_max = 0
for t in range(1, 100):
threshold = t / 100.0
predictions = (preds > threshold).astype(np.int32)
m_tp = 0
m_fp = 0
m_fn = 0
M_pr = 0
M_rc = 0
total = 0
p_total = 0
for i in range(len(preds)):
tp = np.sum(predictions[i, :] * labels[i, :])
fp = np.sum(predictions[i, :]) - tp
fn = np.sum(labels[i, :]) - tp
if tp == 0 and fp == 0 and fn == 0:
continue
total += 1
if tp > 0:
pr = tp / (1.0 * (tp + fp))
rc = tp / (1.0 * (tp + fn))
m_tp += tp
m_fp += fp
m_fn += fn
M_pr += pr
M_rc += rc
p_total += 1

if p_total == 0:
continue
if total > 0:
m_tp /= total
m_fn /= total
m_fp /= total
m_pr = m_tp / (1.0 * (m_tp + m_fp))
m_rc = m_tp / (1.0 * (m_tp + m_fn))
M_pr /= p_total
M_rc /= total
m_f1 = 2 * m_pr * m_rc / (m_pr + m_rc)
M_f1 = 2 * M_pr * M_rc / (M_pr + M_rc)

if m_f1 > m_f1_max:
m_f1_max = m_f1
m_pr_max = m_pr
m_rc_max = m_rc

if M_f1 > M_f1_max:
M_f1_max = M_f1
M_pr_max = M_pr
M_rc_max = M_rc

return m_pr_max, m_rc_max, m_f1_max, M_pr_max, M_rc_max, M_f1_max





def micro_macro_function_centric_f1(preds, labels):
preds = np.round(preds, 2)
m_f1_max = 0
M_f1_max = 0
for t in range(1, 200):
threshold = t / 200.0
predictions = (preds > threshold).astype(np.int32)
m_tp = 0
m_fp = 0
m_fn = 0
M_pr = 0
M_rc = 0
for i in range(preds.shape[0]):
tp = np.sum(predictions[i, :] * labels[i, :])
fp = np.sum(predictions[i, :]) - tp
fn = np.sum(labels[i, :]) - tp
m_tp += tp
m_fp += fp
m_fn += fn
if tp > 0:
pr = 1.0 * tp / (1.0 * (tp + fp))
rc = 1.0 * tp / (1.0 * (tp + fn))
else:
if fp == 0 and fn == 0:
pr = 1
rc = 1
else:
pr = 0
rc = 0
M_pr += pr
M_rc += rc

if m_tp > 0:
m_pr = 1.0 * m_tp / (1.0 * (m_tp + m_fp))
m_rc = 1.0 * m_tp / (1.0 * (m_tp + m_fn))
m_f1 = 2.0 * m_pr * m_rc / (m_pr + m_rc)
else:
if m_fp == 0 and m_fn == 0:
m_pr = 1
m_rc = 1
m_f1 = 1
else:
m_pr = 0
m_rc = 0
m_f1 = 0

M_pr /= preds.shape[0]
M_rc /= preds.shape[0]
if M_pr == 0 and M_rc == 0:
M_f1 = 0
else:
M_f1 = 2.0 * M_pr * M_rc / (M_pr + M_rc)
if m_f1 > m_f1_max:
m_f1_max = m_f1
m_pr_max = m_pr
m_rc_max = m_rc

if M_f1 > M_f1_max:
M_f1_max = M_f1
M_pr_max = M_pr
M_rc_max = M_rc

return m_pr_max, m_rc_max, m_f1_max, M_pr_max, M_rc_max, M_f1_max


def compute_roc(preds, labels):
# Compute ROC curve and ROC area for each class
fpr, tpr, _ = roc_curve(labels.flatten(), preds.flatten())
roc_auc = auc(fpr, tpr)
return roc_auc


def compute_aupr(preds, labels):
# Compute ROC curve and ROC area for each class
pr, rc, threshold =precision_recall_curve(labels.flatten(), preds.flatten())
pr_auc = auc(rc, pr)
#pr, rc, threshold =precision_recall_curve(labels.flatten(), preds.flatten(),average ='macro' )
M_pr_auc = 0
return pr_auc, M_pr_auc



def compute_mcc(preds, labels):
# Compute ROC curve and ROC area for each class
mcc = matthews_corrcoef(labels.flatten(), preds.flatten())
return mcc


def compute_performance(preds, labels): #, gos):
preds = np.round(preds, 2)
f_max = 0
p_max = 0
r_max = 0
t_max = 0
for t in range(1, 100):
threshold = t / 100.0
predictions = (preds > threshold).astype(np.int32)
total = 0
f = 0.0
p = 0.0
r = 0.0
p_total = 0
for i in range(labels.shape[0]):
tp = np.sum(predictions[i, :] * labels[i, :])
fp = np.sum(predictions[i, :]) - tp
fn = np.sum(labels[i, :]) - tp
all_gos = set()
#for go_id in gos[i]:
# if go_id in all_functions:
# all_gos |= get_anchestors(go, go_id)
#all_gos.discard(GO_ID)
#all_gos -= func_set
#fn += len(all_gos)
if tp == 0 and fp == 0 and fn == 0:
continue
total += 1
if tp != 0:
p_total += 1
precision = tp / (1.0 * (tp + fp))
recall = tp / (1.0 * (tp + fn))
p += precision
r += recall
if p_total == 0:
continue
r /= total
p /= p_total
if p + r > 0:
f = 2 * p * r / (p + r)
if f_max < f:
f_max = f
p_max = p
r_max = r
t_max = threshold
predictions_max = predictions
return f_max, p_max, r_max, t_max, predictions_max


def get_gos(pred):
mdist = 1.0
mgos = None
for i in range(len(labels_gos)):
labels, gos = labels_gos[i]
dist = distance.cosine(pred, labels)
if mdist > dist:
mdist = dist
mgos = gos
return mgos


def compute_similarity_performance(train_df, test_df, preds):
logging.info("Computing similarity performance")
logging.info("Training data size %d" % len(train_df))
train_labels = train_df['labels'].values
train_gos = train_df['gos'].values
global labels_gos
labels_gos = zip(train_labels, train_gos)
p = Pool(64)
pred_gos = p.map(get_gos, preds)
total = 0
p = 0.0
r = 0.0
f = 0.0
test_gos = test_df['gos'].values
for gos, tgos in zip(pred_gos, test_gos):
preds = set()
test = set()
for go_id in gos:
if go_id in all_functions:
preds |= get_anchestors(go, go_id)
for go_id in tgos:
if go_id in all_functions:
test |= get_anchestors(go, go_id)
tp = len(preds.intersection(test))
fp = len(preds - test)
fn = len(test - preds)
if tp == 0 and fp == 0 and fn == 0:
continue
total += 1
if tp != 0:
precision = tp / (1.0 * (tp + fp))
recall = tp / (1.0 * (tp + fn))
p += precision
r += recall
f += 2 * precision * recall / (precision + recall)
return f / total, p / total, r / total


def print_report(report, go_id):
with open(DATA_ROOT + 'reports.txt', 'a') as f:
f.write('Classification report for ' + go_id + '\n')
f.write(report + '\n')


if __name__ == '__main__':
main()

+ 0
- 182
conditional_wgan_wrapper_post.py View File

from functools import partial

import numpy as np
from keras import backend as K
from keras.layers import Input
from keras.layers.merge import _Merge
from keras.models import Model
from keras.optimizers import Adam
from keras.losses import binary_crossentropy

def wasserstein_loss(y_true, y_pred):
"""Calculates the Wasserstein loss for a sample batch.

The Wasserstein loss function is very simple to calculate. In a standard GAN, the discriminator
has a sigmoid output, representing the probability that samples are real or generated. In Wasserstein
GANs, however, the output is linear with no activation function! Instead of being constrained to [0, 1],
the discriminator wants to make the distance between its output for real and generated samples as large as possible.

The most natural way to achieve this is to label generated samples -1 and real samples 1, instead of the
0 and 1 used in normal GANs, so that multiplying the outputs by the labels will give you the loss immediately.

Note that the nature of this loss means that it can be (and frequently will be) less than 0."""
return K.mean(y_true * y_pred)







def generator_recunstruction_loss(y_true, y_pred, enableTrain):
return binary_crossentropy(y_true, y_pred) * enableTrain


global enable_train
enable_train = Input(shape = (1,))
global generator_recunstruction_loss_new
generator_recunstruction_loss_new = partial(generator_recunstruction_loss, enableTrain = enable_train)
generator_recunstruction_loss_new.__name__ = 'generator_recunstruction_loss_new'


def WGAN_wrapper(generator, discriminator, generator_input_shape, discriminator_input_shape, discriminator_input_shape2,
batch_size, gradient_penalty_weight):

BATCH_SIZE = batch_size
GRADIENT_PENALTY_WEIGHT = gradient_penalty_weight


def set_trainable_state(model, state):
for layer in model.layers:
layer.trainable = state
model.trainable = state

def gradient_penalty_loss(y_true, y_pred, averaged_samples, gradient_penalty_weight):
"""Calculates the gradient penalty loss for a batch of "averaged" samples.

In Improved WGANs, the 1-Lipschitz constraint is enforced by adding a term to the loss function
that penalizes the network if the gradient norm moves away from 1. However, it is impossible to evaluate
this function at all points in the input space. The compromise used in the paper is to choose random points
on the lines between real and generated samples, and check the gradients at these points. Note that it is the
gradient w.r.t. the input averaged samples, not the weights of the discriminator, that we're penalizing!

In order to evaluate the gradients, we must first run samples through the generator and evaluate the loss.
Then we get the gradients of the discriminator w.r.t. the input averaged samples.
The l2 norm and penalty can then be calculated for this gradient.

Note that this loss function requires the original averaged samples as input, but Keras only supports passing
y_true and y_pred to loss functions. To get around this, we make a partial() of the function with the
averaged_samples argument, and use that for model training."""
# first get the gradients:
# assuming: - that y_pred has dimensions (batch_size, 1)
# - averaged_samples has dimensions (batch_size, nbr_features)
# gradients afterwards has dimension (batch_size, nbr_features), basically
# a list of nbr_features-dimensional gradient vectors
gradients = K.gradients(y_pred, averaged_samples)[0]
# compute the euclidean norm by squaring ...
gradients_sqr = K.square(gradients)
# ... summing over the rows ...
gradients_sqr_sum = K.sum(gradients_sqr,
axis=np.arange(1, len(gradients_sqr.shape)))
# ... and sqrt
gradient_l2_norm = K.sqrt(gradients_sqr_sum)
# compute lambda * (1 - ||grad||)^2 still for each single sample
gradient_penalty = gradient_penalty_weight * K.square(1 - gradient_l2_norm)
# return the mean as loss over all the batch samples
return K.mean(gradient_penalty)

class RandomWeightedAverage(_Merge):
"""Takes a randomly-weighted average of two tensors. In geometric terms, this outputs a random point on the line
between each pair of input points.

Inheriting from _Merge is a little messy but it was the quickest solution I could think of.
Improvements appreciated."""

def _merge_function(self, inputs):
weights = K.random_uniform((BATCH_SIZE, 1))
print(inputs[0])
return (weights * inputs[0]) + ((1 - weights) * inputs[1])

# The generator_model is used when we want to train the generator layers.
# As such, we ensure that the discriminator layers are not trainable.
# Note that once we compile this model, updating .trainable will have no effect within it. As such, it
# won't cause problems if we later set discriminator.trainable = True for the discriminator_model, as long
# as we compile the generator_model first.

set_trainable_state(discriminator, False)
set_trainable_state(generator, True)
#enable_train = Input(shape = (1,))
generator_input = Input(shape=generator_input_shape)
generator_layers = generator(generator_input)
#discriminator_noise_in = Input(shape=(1,))
#input_seq_g = Input(shape = discriminator_input_shape2)
discriminator_layers_for_generator = discriminator([generator_layers, generator_input])
generator_model = Model(inputs=[generator_input, enable_train],
outputs=[generator_layers, discriminator_layers_for_generator])
# We use the Adam paramaters from Gulrajani et al.
#global generator_recunstruction_loss_new
#generator_recunstruction_loss_new = partial(generator_recunstruction_loss, enableTrain = enable_train)
#generator_recunstruction_loss_new.__name__ = 'generator_RLN'

loss = [generator_recunstruction_loss_new, wasserstein_loss]
loss_weights = [30, 1]
generator_model.compile(optimizer=Adam(lr=1E-3, beta_1=0.9, beta_2=0.999, epsilon=1e-08),
loss=loss, loss_weights=loss_weights)

# Now that the generator_model is compiled, we can make the discriminator layers trainable.

set_trainable_state(discriminator, True)
set_trainable_state(generator, False)

# The discriminator_model is more complex. It takes both real image samples and random noise seeds as input.
# The noise seed is run through the generator model to get generated images. Both real and generated images
# are then run through the discriminator. Although we could concatenate the real and generated images into a
# single tensor, we don't (see model compilation for why).
real_samples = Input(shape=discriminator_input_shape)
input_seq = Input(shape = discriminator_input_shape2)
generator_input_for_discriminator = Input(shape=generator_input_shape)
generated_samples_for_discriminator = generator(generator_input_for_discriminator)
discriminator_output_from_generator = discriminator([generated_samples_for_discriminator, generator_input_for_discriminator] )
discriminator_output_from_real_samples = discriminator([real_samples, input_seq])

# We also need to generate weighted-averages of real and generated samples, to use for the gradient norm penalty.
averaged_samples = RandomWeightedAverage()([real_samples, generated_samples_for_discriminator])
average_seq = RandomWeightedAverage()([input_seq, generator_input_for_discriminator])
# We then run these samples through the discriminator as well. Note that we never really use the discriminator
# output for these samples - we're only running them to get the gradient norm for the gradient penalty loss.
#print('hehehe')
#print(averaged_samples)
averaged_samples_out = discriminator([averaged_samples, average_seq] )

# The gradient penalty loss function requires the input averaged samples to get gradients. However,
# Keras loss functions can only have two arguments, y_true and y_pred. We get around this by making a partial()
# of the function with the averaged samples here.
partial_gp_loss = partial(gradient_penalty_loss,
averaged_samples=averaged_samples,
gradient_penalty_weight=GRADIENT_PENALTY_WEIGHT)
partial_gp_loss.__name__ = 'gradient_penalty' # Functions need names or Keras will throw an error

# Keras requires that inputs and outputs have the same number of samples. This is why we didn't concatenate the
# real samples and generated samples before passing them to the discriminator: If we had, it would create an
# output with 2 * BATCH_SIZE samples, while the output of the "averaged" samples for gradient penalty
# would have only BATCH_SIZE samples.

# If we don't concatenate the real and generated samples, however, we get three outputs: One of the generated
# samples, one of the real samples, and one of the averaged samples, all of size BATCH_SIZE. This works neatly!
discriminator_model = Model(inputs=[real_samples, generator_input_for_discriminator,input_seq],
outputs=[discriminator_output_from_real_samples,
discriminator_output_from_generator,
averaged_samples_out])
# We use the Adam paramaters from Gulrajani et al. We use the Wasserstein loss for both the real and generated
# samples, and the gradient penalty loss for the averaged samples.
discriminator_model.compile(optimizer=Adam(lr=1E-3, beta_1=0.9, beta_2=0.999, epsilon=1e-08),
loss=[wasserstein_loss,
wasserstein_loss,
partial_gp_loss])

# set_trainable_state(discriminator, True)
# set_trainable_state(generator, True)

return generator_model, discriminator_model

Write Preview
Loading…
Cancel
Save