def main(args=None):
"""Generate random periodic time series and train an autoencoder model.
args: dict
Dictionary of values to override default values in `keras_util.parse_model_args`;
can also be passed via command line. See `parse_model_args` for full list of
possible arguments.
"""
args = ku.parse_model_args(args)
train = np.arange(args.N_train); test = np.arange(args.N_test) + args.N_train
X, Y, X_raw = sample_data.periodic(args.N_train + args.N_test, args.n_min, args.n_max,
even=args.even, noise_sigma=args.sigma,
kind=args.data_type)
if args.even:
X = X[:, :, 1:2]
X_raw = X_raw[:, :, 1:2]
else:
X[:, :, 0] = ku.times_to_lags(X_raw[:, :, 0])
X[np.isnan(X)] = -1.
X_raw[np.isnan(X_raw)] = -1.
model_type_dict = {'gru': GRU, 'lstm': LSTM, 'vanilla': SimpleRNN}
main_input = Input(shape=(X.shape[1], X.shape[-1]), name='main_input')
if args.even:
model_input = main_input
aux_input = None
else:
aux_input = Input(shape=(X.shape[1], X.shape[-1] - 1), name='aux_input')
model_input = [main_input, aux_input]
encode = encoder(main_input, layer=model_type_dict[args.model_type],
output_size=args.embedding, **vars(args))
decode = decoder(encode, num_layers=args.decode_layers if args.decode_layers
else args.num_layers,
layer=model_type_dict[args.decode_type if args.decode_type
else args.model_type],
n_step=X.shape[1], aux_input=aux_input,
**{k: v for k, v in vars(args).items() if k != 'num_layers'})
model = Model(model_input, decode)
run = ku.get_run_id(**vars(args))
if args.even:
history = ku.train_and_log(X[train], X_raw[train], run, model, **vars(args))
else:
sample_weight = (X[train, :, -1] != -1)
history = ku.train_and_log({'main_input': X[train], 'aux_input': X[train, :, 0:1]},
X_raw[train, :, 1:2], run, model,
sample_weight=sample_weight, **vars(args))
return X, Y, X_raw, model, args
def main(args=None):
args = ku.parse_model_args(args)
N = args.N_train + args.N_test
train = np.arange(args.N_train)
test = np.arange(args.N_test) + args.N_train
X, Y, X_raw = sample_data.periodic(N,
args.n_min,
args.n_max,
even=args.even,
noise_sigma=args.sigma,
kind=args.data_type)
if args.even:
X = X[:, :, 1:2]
else:
X[:, :, 0] = ku.times_to_lags(X_raw[:, :, 0])
X[np.isnan(X)] = -1.
X_raw[np.isnan(X_raw)] = -1.
Y = sample_data.phase_to_sin_cos(Y)
scaler = StandardScaler(copy=False, with_mean=True, with_std=True)
scaler.fit_transform(Y)
if args.loss_weights: # so far, only used to zero out some columns
Y *= args.loss_weights
model_type_dict = {'gru': GRU, 'lstm': LSTM, 'vanilla': SimpleRNN}
model_input = Input(shape=(X.shape[1], X.shape[-1]), name='main_input')
encode = encoder(model_input,
layer=model_type_dict[args.model_type],
output_size=Y.shape[-1],
**vars(args))
model = Model(model_input, encode)
run = ku.get_run_id(**vars(args))
history = ku.train_and_log(X[train], Y[train], run, model, **vars(args))
return X, Y, X_raw, scaler, model, args
def main(args=None):
args = parse_model_args(args)
K.set_floatx('float64')
run = get_run_id(**vars(args))
log_dir = os.path.join(os.getcwd(), 'keras_logs', args.sim_type,
run) #Loading the model architecture
weights_path = os.path.join(log_dir,
'weights.h5') #Loading the model weights
print("log_dir", log_dir)
print("Weight matrix read...")
#Load the model
#How do I access dict using index? I just want to load the model
#Why not use the main args?
model = list(survey_autoencoder(vars(args)))[2]
#LOADING GMM PARAMTERS
# Where is gmm.mu updated?
gmm_para = np.load(log_dir + '/gmm_parameters')
gmm_mu = gmm_para[gmm_mu] #Size = embedding size * #classes
#LOADING THE MODEL
decode_model = Model(inputs=model.input,
outputs=model.get_layer('time_dist').output)
#What is X[valid], X[new]
#TRAINING SAMPLES
gmm_mu = np.float64(gmm_mu)
#NO AUX INPUT FOR DECODER! DOES IT ASSUME EVEN INPUTS?
decoding_train = decode_model.predict(gmm_mu)
print(decoding_train.shape)
phase = np.linspace(0, 1, len(decoding_train[0]))
for i in range(len(decoding_train)):
plt.plot(phase, decoding_train[i])
def main(args=None):
"""Train an autoencoder model from `LightCurve` objects saved in
`args.survey_files`.
args: dict
Dictionary of values to override default values in `keras_util.parse_model_args`;
can also be passed via command line. See `parse_model_args` for full list of
possible arguments.
"""
args = ku.parse_model_args(args)
np.random.seed(0)
K.set_floatx('float64')
run = ku.get_run_id(**vars(args))
if not args.survey_files:
raise ValueError("No survey files given")
lc_lists = [joblib.load(f) for f in args.survey_files]
n_reps = [max(len(y) for y in lc_lists) // len(x) for x in lc_lists]
combined = sum([x * i for x, i in zip(lc_lists, n_reps)], [])
# Preparation for classification probability
classprob_pkl = joblib.load("./data/asassn/class_probs.pkl")
class_probability = dict(classprob_pkl)
# Combine subclasses into eight superclasses:
# CEPH, DSCT, ECL, RRAB, RRCD, M, ROT, SR
for lc in combined:
if ((lc.label == 'EW') or (lc.label == 'EA') or (lc.label == 'EB')):
lc.label = 'ECL'
if ((lc.label == 'CWA') or (lc.label == 'CWB') or (lc.label == 'DCEP') or
(lc.label == 'DCEPS') or (lc.label == 'RVA')):
lc.label = "CEPH"
if ((lc.label == 'DSCT') or (lc.label == 'HADS')):
lc.label = "DSCT"
if ((lc.label == 'RRD') or (lc.label == 'RRC')):
lc.label = "RRCD"
top_classes = ['SR', 'RRAB', 'RRCD', 'M', 'ROT', 'ECL', 'CEPH', 'DSCT']
print ("Number of raw LCs:", len(combined))
if args.lomb_score:
combined = [lc for lc in combined if lc.best_score >= args.lomb_score]
if args.ss_resid:
combined = [lc for lc in combined if lc.ss_resid <= args.ss_resid]
if args.class_prob:
combined = [lc for lc in combined if float(class_probability[lc.name.split("/")[-1][2:-4]]) >= args.class_prob]
#combined = [lc for lc in combined if lc.class_prob >= args.class_prob]
# Select only superclasses for training
combined = [lc for lc in combined if lc.label in top_classes]
split = [el for lc in combined for el in lc.split(args.n_min, args.n_max)]
if args.period_fold:
for lc in split:
lc.period_fold()
X_list = [np.c_[lc.times, lc.measurements, lc.errors] for lc in split]
classnames, indices = np.unique([lc.label for lc in split], return_inverse=True)
y = classnames[indices]
periods = np.array([np.log10(lc.p) for lc in split])
periods = periods.reshape(len(split), 1)
X_raw = pad_sequences(X_list, value=np.nan, dtype='float64', padding='post')
model_type_dict = {'gru': GRU, 'lstm': LSTM, 'vanilla': SimpleRNN}
X, means, scales, wrong_units, X_err = preprocess(X_raw, args.m_max)
y = y[~wrong_units]
periods = periods[~wrong_units]
# Prepare the indices for training and validation
train, valid = list(StratifiedKFold(n_splits=5, shuffle=True, random_state=SEED).split(X, y))[0]
X_valid = X[valid]
y_valid = y[valid]
means_valid = means[valid]
scales_valid = scales[valid]
periods_valid = periods[valid]
energy_dummy_valid = np.zeros((X_valid.shape[0], 1))
X_err_valid = X_err[valid]
sample_weight_valid = 1. / X_err_valid
X = X[train]
y = y[train]
means = means[train]
scales = scales[train]
periods = periods[train]
energy_dummy = np.zeros((X.shape[0], 1))
X_err = X_err[train]
sample_weight = 1. / X_err
supports_valid = np.concatenate((means_valid, scales_valid, periods_valid), axis=1)
supports = np.concatenate((means, scales, periods), axis=1)
num_supports_train = supports.shape[-1]
num_supports_valid = supports_valid.shape[-1]
assert(num_supports_train == num_supports_valid)
num_additional = num_supports_train + 2 # 2 for reconstruction error
if (args.gmm_on):
gmm = GMM(args.num_classes, args.embedding+num_additional)
### Covert labels into one-hot vectors for training
label_encoder = LabelEncoder()
label_encoder.fit(y)
### Transform the integers into one-hot vector for softmax
train_y_encoded = label_encoder.transform(y)
train_y = np_utils.to_categorical(train_y_encoded)
### Repeat for validation dataset
label_encoder = LabelEncoder()
label_encoder.fit(y_valid)
### Transform the integers into one-hot vector for softmax
valid_y_encoded = label_encoder.transform(y_valid)
valid_y = np_utils.to_categorical(valid_y_encoded)
main_input = Input(shape=(X.shape[1], 2), name='main_input') # dim: (200, 2) = dt, mag
aux_input = Input(shape=(X.shape[1], 1), name='aux_input') # dim: (200, 1) = dt's
model_input = [main_input, aux_input]
if (args.gmm_on):
support_input = Input(shape=(num_supports_train,), name='support_input')
model_input = [main_input, aux_input, support_input]
encode = encoder(main_input, layer=model_type_dict[args.model_type],
output_size=args.embedding, **vars(args))
decode = decoder(encode, num_layers=args.decode_layers if args.decode_layers
else args.num_layers,
layer=model_type_dict[args.decode_type if args.decode_type
else args.model_type],
n_step=X.shape[1], aux_input=aux_input,
**{k: v for k, v in vars(args).items() if k != 'num_layers'})
optimizer = Adam(lr=args.lr if not args.finetune_rate else args.finetune_rate)
if (not args.gmm_on):
model = Model(model_input, decode)
model.compile(optimizer=optimizer, loss='mse',
sample_weight_mode='temporal')
else:
est_net = EstimationNet(args.estnet_size, args.num_classes, K.tanh)
# extract x_i and hat{x}_{i} for reconstruction error feature calculations
main_input_slice = Lambda(lambda x: x[:,:,1], name="main_slice")(main_input)
decode_slice = Lambda(lambda x: x[:,:,0], name="decode_slice")(decode)
z_both = Lambda(extract_features, name="concat_zs")([main_input_slice, decode_slice, encode, support_input])
gamma = est_net.inference(z_both, args.estnet_drop_frac)
print ("z_both shape", z_both.shape)
print ("gamma shape", gamma.shape)
sigma_i = Lambda(lambda x: gmm.fit(x[0], x[1]),
output_shape=(args.num_classes, args.embedding+num_additional, args.embedding+num_additional),
name="sigma_i")([z_both, gamma])
energy = Lambda(lambda x: gmm.energy(x[0], x[1]), name="energy")([z_both, sigma_i])
model_output = [decode, gamma, energy]
model = Model(model_input, model_output)
optimizer = Adam(lr=args.lr if not args.finetune_rate else args.finetune_rate)
model.compile(optimizer=optimizer,
loss=['mse', 'categorical_crossentropy', energy_loss],
loss_weights=[1.0, 1.0, args.lambda1],
metrics={'gamma':'accuracy'},
sample_weight_mode=['temporal', None, None])
# summary for checking dimensions
print(model.summary())
if (args.gmm_on):
history = ku.train_and_log( {'main_input':X, 'aux_input':np.delete(X, 1, axis=2), 'support_input':supports},
{'time_dist':X[:, :, [1]], 'gamma':train_y, 'energy':energy_dummy}, run, model,
sample_weight={'time_dist':sample_weight, 'gamma':None, 'energy':None},
validation_data=(
{'main_input':X_valid, 'aux_input':np.delete(X_valid, 1, axis=2), 'support_input':supports_valid},
{'time_dist':X_valid[:, :, [1]], 'gamma':valid_y, 'energy':energy_dummy_valid},
{'time_dist':sample_weight_valid, 'gamma':None, 'energy':None}),
**vars(args))
else: # Just train RNN autoencoder
history = ku.train_and_log({'main_input':X, 'aux_input':np.delete(X, 1, axis=2)},
X[:, :, [1]], run, model,
sample_weight=sample_weight,
validation_data=(
{'main_input':X_valid, 'aux_input':np.delete(X_valid, 1, axis=2)},
X_valid[:, :, [1]], sample_weight_valid), **vars(args))
return X, X_raw, model, means, scales, wrong_units, args
def main(args=None):
args = ku.parse_model_args(args)
args.loss = 'categorical_crossentropy'
np.random.seed(0)
if not args.survey_files:
raise ValueError("No survey files given")
classes = [
'RR_Lyrae_FM', 'W_Ursae_Maj', 'Classical_Cepheid', 'Beta_Persei',
'Semireg_PV'
]
lc_lists = [joblib.load(f) for f in args.survey_files]
combined = [lc for lc_list in lc_lists for lc in lc_list]
combined = [lc for lc in combined if lc.label in classes]
if args.lomb_score:
combined = [lc for lc in combined if lc.best_score >= args.lomb_score]
split = [el for lc in combined for el in lc.split(args.n_min, args.n_max)]
if args.period_fold:
for lc in split:
lc.period_fold()
X_list = [np.c_[lc.times, lc.measurements, lc.errors] for lc in split]
classnames, y_inds = np.unique([lc.label for lc in split],
return_inverse=True)
Y = to_categorical(y_inds, len(classnames))
X_raw = pad_sequences(X_list, value=np.nan, dtype='float', padding='post')
X, means, scales, wrong_units = preprocess(X_raw, args.m_max)
Y = Y[~wrong_units]
# Remove errors
X = X[:, :, :2]
if args.even:
X = X[:, :, 1:]
# shuffled_inds = np.random.permutation(np.arange(len(X)))
# train = np.sort(shuffled_inds[:args.N_train])
# valid = np.sort(shuffled_inds[args.N_train:])
train, valid = list(
StratifiedKFold(n_splits=5, shuffle=True,
random_state=0).split(X_list, y_inds))[0]
model_type_dict = {
'gru': GRU,
'lstm': LSTM,
'vanilla': SimpleRNN,
'conv': Conv1D
} #, 'atrous': AtrousConv1D, 'phased': PhasedLSTM}
# if args.pretrain:
# auto_args = {k: v for k, v in args.__dict__.items() if k != 'pretrain'}
# auto_args['sim_type'] = args.pretrain
## auto_args['no_train'] = True
# auto_args['epochs'] = 1; auto_args['loss'] = 'mse'; auto_args['batch_size'] = 32; auto_args['sim_type'] = 'test'
# _, _, auto_model, _ = survey_autoencoder(auto_args)
# for layer in auto_model.layers:
# layer.trainable = False
# model_input = auto_model.input[0]
# encode = auto_model.get_layer('encoding').output
# else:
# model_input = Input(shape=(X.shape[1], X.shape[-1]), name='main_input')
# encode = encoder(model_input, layer=model_type_dict[args.model_type],
# output_size=args.embedding, **vars(args))
model_input = Input(shape=(X.shape[1], X.shape[-1]), name='main_input')
encode = encoder(model_input,
layer=model_type_dict[args.model_type],
output_size=args.embedding,
**vars(args))
scale_param_input = Input(shape=(2, ), name='scale_params')
merged = merge([encode, scale_param_input], mode='concat')
out = Dense(args.size + 2, activation='relu')(merged)
out = Dense(Y.shape[-1], activation='softmax')(out)
model = Model([model_input, scale_param_input], out)
run = ku.get_run_id(**vars(args))
if args.pretrain:
for layer in model.layers:
layer.trainable = False
pretrain_weights = os.path.join('keras_logs', args.pretrain, run,
'weights.h5')
else:
pretrain_weights = None
history = ku.train_and_log(
[X[train], np.c_[means, scales][train]],
Y[train],
run,
model,
metrics=['accuracy'],
validation_data=([X[valid], np.c_[means, scales][valid]], Y[valid]),
pretrain_weights=pretrain_weights,
**vars(args))
return X, X_raw, Y, model, args
def main(args=None):
"""Train an autoencoder model from `LightCurve` objects saved in
`args.survey_files`.
args: dict
Dictionary of values to override default values in `keras_util.parse_model_args`;
can also be passed via command line. See `parse_model_args` for full list of
possible arguments.
"""
args = ku.parse_model_args(args)
np.random.seed(0)
if not args.survey_files:
raise ValueError("No survey files given")
lc_lists = [joblib.load(f) for f in args.survey_files]
n_reps = [max(len(y) for y in lc_lists) // len(x) for x in lc_lists]
combined = sum([x * i for x, i in zip(lc_lists, n_reps)], [])
if args.lomb_score:
combined = [lc for lc in combined if lc.best_score >= args.lomb_score]
if args.ss_resid:
combined = [lc for lc in combined if lc.ss_resid <= args.ss_resid]
split = [el for lc in combined for el in lc.split(args.n_min, args.n_max)]
if args.period_fold:
for lc in split:
lc.period_fold()
X_list = [np.c_[lc.times, lc.measurements, lc.errors] for lc in split]
X_raw = pad_sequences(X_list, value=np.nan, dtype='float', padding='post')
model_type_dict = {'gru': GRU, 'lstm': LSTM, 'vanilla': SimpleRNN}
X, means, scales, wrong_units = preprocess(X_raw, args.m_max)
main_input = Input(shape=(X.shape[1], X.shape[-1]), name='main_input')
aux_input = Input(shape=(X.shape[1], X.shape[-1] - 1), name='aux_input')
model_input = [main_input, aux_input]
encode = encoder(main_input,
layer=model_type_dict[args.model_type],
output_size=args.embedding,
**vars(args))
decode = decoder(
encode,
num_layers=args.decode_layers
if args.decode_layers else args.num_layers,
layer=model_type_dict[args.decode_type if args.decode_type else args.
model_type],
n_step=X.shape[1],
aux_input=aux_input,
**{k: v
for k, v in vars(args).items() if k != 'num_layers'})
model = Model(model_input, decode)
run = ku.get_run_id(**vars(args))
errors = X_raw[:, :, 2] / scales
sample_weight = 1. / errors
sample_weight[np.isnan(sample_weight)] = 0.0
X[np.isnan(X)] = 0.
history = ku.train_and_log(
{
'main_input': X,
'aux_input': np.delete(X, 1, axis=2)
},
X[:, :, [1]],
run,
model,
sample_weight=sample_weight,
errors=errors,
validation_split=0.0,
**vars(args))
return X, X_raw, model, means, scales, wrong_units, args
def main(args=None):
args = parse_model_args(args)
K.set_floatx('float64')
run = get_run_id(**vars(args))
log_dir = os.path.join(os.getcwd(), 'keras_logs', args.sim_type, run)
weights_path = os.path.join(log_dir, 'weights.h5')
if not os.path.exists(weights_path):
raise FileNotFoundError(weights_path)
X_fold, X_raw_fold, model, means, scales, wrong_units, args = survey_autoencoder(
vars(args))
print("log_dir", log_dir)
print("Weight matrix read...")
full = joblib.load(args.survey_files[0])
# Combine subclasses
# Resulting in five classes: RR, EC, SR, M, ROT
for lc in full:
if ((lc.label == 'EW') or (lc.label == 'EA') or (lc.label == 'EB')):
lc.label = 'ECL'
if ((lc.label == 'CWA') or (lc.label == 'CWB') or (lc.label == 'DCEP')
or (lc.label == 'DCEPS') or (lc.label == 'RVA')):
lc.label = "CEPH"
if ((lc.label == 'DSCT') or (lc.label == 'HADS')):
lc.label = "DSCT"
if ((lc.label == 'RRD') or (lc.label == 'RRC')):
lc.label = "RRCD"
top_classes = ['SR', 'RRAB', 'RRCD', 'M', 'ROT', 'ECL', 'CEPH', 'DSCT']
new_classes = ['VAR']
top = [lc for lc in full if lc.label in top_classes]
new = [lc for lc in full if lc.label in new_classes]
# Preparation for classification probability
classprob_pkl = joblib.load("./data/asassn/class_probs.pkl")
class_probability = dict(classprob_pkl)
if args.ss_resid:
top = [lc for lc in top if lc.ss_resid <= args.ss_resid]
if args.class_prob:
top = [
lc for lc in top
if float(class_probability[lc.name.split("/")[-1][2:-4]]) >= 0.9
]
#top = [lc for lc in top if lc.class_prob >= args.class_prob]
split = [el for lc in top for el in lc.split(args.n_min, args.n_max)]
split_new = [el for lc in new for el in lc.split(args.n_min, args.n_max)]
if args.period_fold:
for lc in split:
lc.period_fold()
for lc in split_new:
if lc.p is not None:
lc.period_fold()
X_list = [np.c_[lc.times, lc.measurements, lc.errors] for lc in split]
classnames, indices = np.unique([lc.label for lc in split],
return_inverse=True)
y = classnames[indices]
periods = np.array([np.log10(lc.p) for lc in split])
periods = periods.reshape(len(split), 1)
X_raw = pad_sequences(X_list, value=0., dtype='float64', padding='post')
X, means, scales, wrong_units, X_err = preprocess(X_raw, args.m_max)
y = y[~wrong_units]
periods = periods[~wrong_units]
train, valid = list(
StratifiedKFold(n_splits=5, shuffle=True,
random_state=SEED).split(X, y))[0]
X_train = X[train]
y_train = y[train]
means_train = means[train]
scales_train = scales[train]
periods_train = periods[train]
energy_dummy = np.zeros((X_train.shape[0], 1))
X_valid = X[valid]
y_valid = y[valid]
means_valid = means[valid]
scales_valid = scales[valid]
periods_valid = periods[valid]
energy_dummy_valid = np.zeros((X_valid.shape[0], 1))
supports_train = np.concatenate((means_train, scales_train, periods_train),
axis=1)
supports_valid = np.concatenate((means_valid, scales_valid, periods_valid),
axis=1)
# New class data (VAR type)
X_list_new = [
np.c_[lc.times, lc.measurements, lc.errors] for lc in split_new
]
classnames_new, indices_new = np.unique([lc.label for lc in split_new],
return_inverse=True)
y_new = classnames_new[indices_new]
periods_new = np.array(
[np.log10(lc.p) if lc.p is not None else 99.0 for lc in split_new])
periods_new = periods_new.reshape(len(split_new), 1)
X_raw_new = pad_sequences(X_list_new,
value=0.,
dtype='float64',
padding='post')
X_new, means_new, scales_new, wrong_units_new, X_err_new = preprocess(
X_raw_new, args.m_max)
y_new = y_new[~wrong_units_new]
periods_new = periods_new[~wrong_units_new]
supports_new = np.concatenate((means_new, scales_new, periods_new), axis=1)
### Concatenating validation data and data from new classes for testing novelty detection
y_new = np.concatenate((y_valid, y_new), axis=0)
X_new = np.concatenate((X_valid, X_new), axis=0)
supports_new = np.concatenate((supports_valid, supports_new), axis=0)
num_supports_train = supports_train.shape[-1]
num_supports_valid = supports_valid.shape[-1]
assert (num_supports_train == num_supports_valid)
num_additional = num_supports_train + 2 # 2 for reconstruction error
### flagging novel samples as 1, samples in superclasses 0.
true_flags = np.array([1 if (l not in top_classes) else 0 for l in y_new])
### Making one-hot labels
label_encoder1 = LabelEncoder()
label_encoder1.fit(y_train)
train_y_encoded = label_encoder1.transform(y_train)
train_y = np_utils.to_categorical(train_y_encoded)
label_encoder2 = LabelEncoder()
label_encoder2.fit(y_valid)
valid_y_encoded = label_encoder2.transform(y_valid)
valid_y = np_utils.to_categorical(valid_y_encoded)
label_encoder3 = LabelEncoder()
label_encoder3.fit(y_new)
new_y_encoded = label_encoder3.transform(y_new)
new_y = np_utils.to_categorical(new_y_encoded)
### Loading trained autoencoder network
encode_model = Model(inputs=model.input,
outputs=model.get_layer('encoding').output)
decode_model = Model(inputs=model.input,
outputs=model.get_layer('time_dist').output)
X_train = np.float64(X_train)
X_valid = np.float64(X_valid)
X_new = np.float64(X_new)
# Passing samples through trained layers
encoding_train = encode_model.predict({
'main_input':
X_train,
'aux_input':
np.delete(X_train, 1, axis=2),
'support_input':
supports_train
})
encoding_valid = encode_model.predict({
'main_input':
X_valid,
'aux_input':
np.delete(X_valid, 1, axis=2),
'support_input':
supports_valid
})
encoding_new = encode_model.predict({
'main_input':
X_new,
'aux_input':
np.delete(X_new, 1, axis=2),
'support_input':
supports_new
})
decoding_train = decode_model.predict({
'main_input':
X_train,
'aux_input':
np.delete(X_train, 1, axis=2),
'support_input':
supports_train
})
decoding_valid = decode_model.predict({
'main_input':
X_valid,
'aux_input':
np.delete(X_valid, 1, axis=2),
'support_input':
supports_valid
})
decoding_new = decode_model.predict({
'main_input':
X_new,
'aux_input':
np.delete(X_new, 1, axis=2),
'support_input':
supports_new
})
z_both_train = extract_features([
X_train[:, :, 1], decoding_train[:, :, 0], encoding_train,
supports_train
])
z_both_valid = extract_features([
X_valid[:, :, 1], decoding_valid[:, :, 0], encoding_valid,
supports_valid
])
z_both_new = extract_features(
[X_new[:, :, 1], decoding_new[:, :, 0], encoding_new, supports_new])
z_both_train = K.eval(z_both_train)
z_both_valid = K.eval(z_both_valid)
z_both_new = K.eval(z_both_new)
# Retrieve estimation network if gmm is on
if (args.gmm_on):
estnet_model = Model(inputs=model.input,
outputs=model.get_layer('gamma').output)
gamma_train = estnet_model.predict({
'main_input':
X_train,
'aux_input':
np.delete(X_train, 1, axis=2),
'support_input':
supports_train
})
gamma_valid = estnet_model.predict({
'main_input':
X_valid,
'aux_input':
np.delete(X_valid, 1, axis=2),
'support_input':
supports_valid
})
else:
est_net = EstimationNet(args.estnet_size, args.num_classes, K.tanh)
# Fit data to gmm if joint, train gmm if sequential
gmm = GMM(args.num_classes, args.embedding + num_additional)
gmm_init = gmm.init_gmm_variables()
# If sequential training, create and train the estimation net
if (not args.gmm_on):
z_input = Input(shape=(args.embedding + num_additional, ),
name='z_input')
gamma = est_net.inference(z_input, args.estnet_drop_frac)
sigma_i = Lambda(lambda x: gmm.fit(x[0], x[1]),
output_shape=(args.num_classes,
args.embedding + num_additional,
args.embedding + num_additional),
name="sigma_i")([z_input, gamma])
energy = Lambda(lambda x: gmm.energy(x[0], x[1]),
name="energy")([z_input, sigma_i])
# Setting up the GMM model
model_output = [gamma, energy]
model = Model(z_input, model_output)
optimizer = Adam(
lr=args.lr if not args.finetune_rate else args.finetune_rate)
model.compile(optimizer=optimizer,
loss=['categorical_crossentropy', energy_loss],
metrics={'gamma': 'accuracy'},
loss_weights=[1.0, args.lambda1])
# Controlling outputs
estnet_size = args.estnet_size
estsize = "_".join(str(s) for s in estnet_size)
gmm_run = 'estnet{}_estdrop{}_l1{}'.format(
estsize, int(100 * args.estnet_drop_frac), args.lambda1)
gmm_dir = os.path.join(os.getcwd(), 'keras_logs', args.sim_type, run,
gmm_run)
if (not os.path.isdir(gmm_dir)):
os.makedirs(gmm_dir)
# For classifier training only
args.nb_epoch = args.cls_epoch
history = ku.train_and_log(z_both_train, {
'gamma': train_y,
'energy': energy_dummy
},
gmm_dir,
model,
validation_data=(z_both_valid, {
'gamma':
valid_y,
'energy':
energy_dummy_valid
}, {
'gamma': None,
'energy': None
}),
**vars(args))
gamma_train, energy_train = model.predict(z_both_train)
gamma_valid, energy_valid = model.predict(z_both_valid)
if (not args.gmm_on):
log_dir = gmm_dir
plot_dir = os.path.join(log_dir, 'figures/')
if (not os.path.isdir(plot_dir)):
os.makedirs(plot_dir)
# Converting to Keras variables to use Keras.backend functions
z_both_train_K = K.variable(z_both_train)
gamma_train_K = K.variable(gamma_train)
z_both_valid_K = K.variable(z_both_valid)
gamma_valid_K = K.variable(gamma_valid)
z_both_new_K = K.variable(z_both_new)
# Fitting GMM parameters only with training set
sigma_i_train = gmm.fit(z_both_train_K, gamma_train_K)
sigma_i_dummy = 1.0
assert (gmm.fitted == True)
# Energy calculation
energy_train = K.eval(gmm.energy(z_both_train_K, sigma_i_train))[:, 0]
energy_valid = K.eval(gmm.energy(z_both_valid_K, sigma_i_dummy))[:, 0]
energy_new = K.eval(gmm.energy(z_both_new_K, sigma_i_dummy))[:, 0]
energy_known = [
e for i, e in enumerate(energy_new) if (true_flags[i] == 0)
]
energy_unknown = [
e for i, e in enumerate(energy_new) if (true_flags[i] == 1)
]
print("known/unknown", len(energy_known), len(energy_unknown))
gmm_phi = K.eval(gmm.phi)
gmm_mu = K.eval(gmm.mu)
gmm_sigma = K.eval(gmm.sigma)
np.savez(log_dir + '/gmm_parameters.npz',
gmm_phi=gmm_phi,
gmm_mu=gmm_mu,
gmm_sigma=gmm_sigma)
percentile_list = [80.0, 95.0]
txtfilename = log_dir + "/novel_detection_scores.txt"
txtfile = open(txtfilename, 'w')
for per in percentile_list:
new_class_energy_threshold = np.percentile(energy_train, per)
print(
f"Energy threshold to detect new class: {new_class_energy_threshold:.2f}"
)
new_pred_flag = np.where(energy_new >= new_class_energy_threshold, 1,
0)
prec, recall, fscore, _ = precision_recall_fscore_support(
true_flags, new_pred_flag, average="binary")
print(f"Detecting new using {per:.1f}% percentile")
print(f" Precision = {prec:.3f}")
print(f" Recall = {recall:.3f}")
print(f" F1-Score = {fscore:.3f}")
txtfile.write(f"Detecting new using {per:.1f}% percentile \n")
txtfile.write(f" Precision = {prec:.3f}\n")
txtfile.write(f" Recall = {recall:.3f}\n")
txtfile.write(f" F1-Score = {fscore:.3f}\n")
txtfile.close()
### Make plots of energy
nbin = 100
fig = plt.figure()
ax = fig.add_subplot(111)
plt.hist(energy_train,
nbin,
normed=True,
color='black',
histtype='step',
label='Training Set')
plt.hist(np.isfinite(energy_unknown),
nbin,
normed=True,
color='blue',
histtype='step',
label='Unknown Classes')
plt.hist(energy_known,
nbin,
normed=True,
color='green',
histtype='step',
label='Known Classes')
plt.legend()
plt.xlabel(r"Energy E(z)")
plt.ylabel("Probability")
plt.savefig(plot_dir + 'energy_histogram.pdf',
dpi=300,
bbox_inches='tight')
plt.clf()
plt.cla()
plt.close()
### Generate confusion matrix
le_list = list(label_encoder2.classes_)
predicted_onehot = gamma_valid
predicted_labels = [
le_list[np.argmax(onehot, axis=None, out=None)]
for onehot in predicted_onehot
]
corr_num, tot_num = plot_confusion_matrix(
y_valid, predicted_labels, classnames,
plot_dir + 'asassn_nn_confusion.pdf')
nn_acc = corr_num / tot_num
### Generate confusion matrix for RF
RF_PARAM_GRID = {
'n_estimators': [50, 100, 250],
'criterion': ['gini', 'entropy'],
'max_features': [0.05, 0.1, 0.2, 0.3],
'min_samples_leaf': [1, 2, 3]
}
rf_model = GridSearchCV(RandomForestClassifier(random_state=0),
RF_PARAM_GRID)
rf_model.fit(encoding_train, y[train])
rf_train_acc = 100 * rf_model.score(encoding_train, y[train])
rf_valid_acc = 100 * rf_model.score(encoding_valid, y[valid])
plot_confusion_matrix(y[valid], rf_model.predict(encoding_valid),
classnames, plot_dir + 'asassn_rf_confusion.pdf')
### Text output
txtfilename = log_dir + "/classification_accuracy.txt"
txtfile = open(txtfilename, 'w')
### Writing results
txtfile.write("===== Classification Accuracy =====\n")
txtfile.write(f"Neural Network Classifier: {nn_acc:.2f}\n")
txtfile.write("==========================\n")
txtfile.write("Random Forest Classifier\n")
txtfile.write(f"Training accuracy: {rf_train_acc:2.2f}%\n")
txtfile.write(f"Validation accuracy: {rf_valid_acc:2.2f}%\n")
txtfile.write(f"Best RF {rf_model.best_params_}\n")
txtfile.write("==========================\n")
txtfile.close()
