def main(args):
if args.model is not None:
model_class = Model_lstm
model = model_class()
elif args.file is not None:
model = tf.keras.models.load_model(f"models/{args.file}.h5")
else:
raise Exception("No valid model")
model.summary()
utils.plot_model(model, args.file)
def main(args):
model = None
use_mfcc = False
use_gabor = False
if args.model == "model_1":
model_class = Model_1
model = model_class()
if args.model == "model_mfcc":
use_mfcc = True
model_class = Model_mfcc
model = model_class()
if args.model == "model_lstm":
model_class = Model_lstm
model = create_model()
checkpoint_prefix = os.path.join(Config.CHECKPOINTS_DIR, "ckpt_{epoch}")
log_dir = Config.TENSORBOARD_LOGDIR + "\\fit\\" + datetime.datetime.now(
).strftime("%Y%m%d-%H%M%S")
callbacks = [
# Interrupt training if `val_loss` stops improving for over 2 epochs
# tf.keras.callbacks.EarlyStopping(patience=10, monitor='val_loss'),
# tf.keras.callbacks.ModelCheckpoint(checkpoint_prefix, save_best_only=True, period=2),
# Write TensorBoard logs to `./logs` directory
tf.keras.callbacks.TensorBoard(log_dir=log_dir, histogram_freq=1)
]
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
trained_model = train(model=model,
epochs=args.epochs,
callbacks=callbacks,
use_mfcc=use_mfcc,
use_gabor=use_gabor)
if args.save_model == True:
# The model description can also be saved as a JSON
# model.save_weights(f"{args.model}.h5", overwrite=False, save_format="HDF5")
# new_model = model_class()
# dataset, _ = create_training_dataset(batch_size=5, use_mfcc=use_mfcc, use_gabor=use_gabor)
# new_model = init_model(new_model, dataset)
# new_model.load_weights(f"{args.model}.h5")
trained_model.save("latest.h5", overwrite=False)
if args.print_summary == True:
loaded_model = tf.keras.models.load_model("latest.h5")
loaded_model.summary()
utils.plot_model(loaded_model)
DURATION = 250 * pq.ms
# # The JIT model is compiled.
# The first model evaluation compiles the model to C code implicitly, from that point onwards evaluation speeds are fractions of ms. This is fast for python as you can see below.
# In[2]:
IinRange = [60, 70, 85, 100]
params = {}
params['amplitude'] = 500 * pq.pA
params['delay'] = DELAY
params['duration'] = 600 * pq.ms
fig_title = 'Layer 5 regular spiking (RS) pyramidal cell (fig 8.12)'
plot_model(IinRange,
reduced_cells,
params,
cell_key='RS',
title=fig_title,
timed=True)
# In[3]:
IinRange = [290, 370, 500, 550]
params = {}
params['delay'] = DELAY
params['duration'] = 600 * pq.ms
fig_title = 'Layer 5 intrinsic bursting (IB) pyramidal cell (fig 8.19)'
plot_model(IinRange, reduced_cells, params, cell_key='IB', title=fig_title)
# In[4]:
IinRange = [200, 300, 400, 600]
import glob
from pathlib import Path
def pull_data(output_dir: str) -> pd.DataFrame:
""" Pulls and concatenates separate csvs from model run"""
data_files = glob.glob(f"{output_dir}/*.csv")
df = pd.concat((pd.read_csv(f) for f in data_files), sort=True)
return df
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("--output_path")
parser.add_argument("--title")
args = parser.parse_args()
output_path = args.output_path # Root of directory to read in inputs from
title = args.title
df = pull_data(output_path)
plot_model(predicted=df.predicted,
observed=df.observed,
title=title,
plot_output_path=Path(output_path) / f"diagnostics/{title}.png")
# Save compiled data frame
df.to_csv(Path(output_path) / "diagnostics/summarized.csv")
print('\nValidate model on {} unknown validation samples:'.format(
X_val.shape[0]))
val_score = model.evaluate(X_val, y_val, verbose=0)
print('Val loss:', val_score[0])
print('Val accuracy:', val_score[1])
return model, model_results, history
print("Load data")
data = load_data()
X_train, y_train, X_test, y_test, X_val, y_val = get_shuffled_splitted_data()
print("Start training")
model, model_results, history = train()
utils.plot_history(model_results)
plt.savefig('alexnet_model_history.png')
# Always same results
np.random.seed(1)
y_val_pred = model.predict(X_val, batch_size=32, verbose=0)
y_val_pred = np.round(y_val_pred).astype(int)
is_lgg = y_val.argmax(axis=1) == 0
utils.plot_predicted_samples(4, X_val[is_lgg], y_val[is_lgg],
y_val_pred[is_lgg], 'Validation set - LGG',
(WIDTH, HEIGHT))
plt.savefig('alexnet_samples_lgg.png')
utils.plot_predicted_samples(4, X_val[is_lgg == False], y_val[is_lgg == False],
y_val_pred[is_lgg == False],
'Validation set - HGG', (WIDTH, HEIGHT))
plt.savefig('alexnet_samples_hgg.png')
utils.plot_model(model, 'skull_classification_model_alexnet.png', show=False)
lr_scheduler = LearningRateScheduler(lr_sch)
lr_reducer = ReduceLROnPlateau(monitor='val_my_metric',
factor=0.2,
patience=5,
mode='max',
min_lr=1e-3)
checkpoint = ModelCheckpoint('model.h5',
monitor='val_loss',
verbose=0,
save_best_only=True,
mode='auto')
model_details = model.fit(data_train,
box_train,
batch_size=128,
epochs=700,
shuffle=True,
validation_split=1,
callbacks=[lr_scheduler, lr_reducer, checkpoint],
verbose=1)
model.save('model.h5')
scores = model.evaluate(data_test, box_test, verbose=1)
print('Test loss : ', scores[0])
print('Test accuracy : ', scores[1])
plot_model(model_details)
def fit_model(sess,model,t,Y,Nw=10,num_iter=500,print_every=10,eta=5e-3,dec_step=20,dec_rate=0.99,plot_=True):
""" Fits the NPDE model to a dataset and returns the fitted object
Args:
sess: TensowFlow session needed for initialization and optimization
t: Python array of numpy vectors storing observation times
Y: Python array of numpy matrices storing observations. Observations
are stored in rows.
Nw: Integer number of samples used for optimization in SDE model
num_iter: Integer number of optimization steps
num_iter: Integer interval of optimization logs to be printed
eta: Float step size used in optimization, must be carefully tuned
dec_step: Float decay interval of the step size
dec_rate: Float decay rate of the step size
plot_: Boolean for plotting the model fit. Valid only for demo
Returns:
npde: Fitted model
"""
print('Building loss function...')
x0 = np.vstack([Y_[0] for Y_ in Y])
if model.name is 'npode':
with tf.name_scope("cost"):
X = model.forward(x0,t)
ll = []
for i in range(len(X)):
mvn = tf.contrib.distributions.MultivariateNormalFullCovariance(loc=X[i],covariance_matrix=tf.diag(model.sn))
ll.append(tf.reduce_sum(mvn.log_prob(Y[i])))
ll = tf.reduce_logsumexp(ll)
ode_prior = model.build_prior()
cost = -(ll + ode_prior)
elif model.name is 'npsde':
with tf.name_scope("cost"):
Xs = model.forward(x0,t,Nw=Nw)
ll = 0
for i in range(len(Y)):
mvn = tf.contrib.distributions.MultivariateNormalFullCovariance(loc=Y[i],covariance_matrix=tf.diag(model.sn))
ll_i = tf.stack([mvn.log_prob(Xs[i][j,:,:]) for j in range(Xs[i].shape[0])]) # Nw x D
ll_i = tf.reduce_sum(tf.log(tf.reduce_mean(tf.exp(ll_i),axis=0)))
ll += ll_i
sde_prior = model.build_prior()
cost = -(ll + sde_prior)
else:
raise NotImplementedError("model parameter should be either 'ode' or 'sde', not {:s}\n".format(model))
print('Adam optimizer being initialized...')
with tf.name_scope("adam"):
global_step = tf.Variable(0, dtype=tf.int32, trainable=False, name='global_step')
expdec = tf.train.exponential_decay(eta,global_step,dec_step,dec_rate,staircase=True)
optimizer = tf.train.AdamOptimizer(expdec).minimize(cost,global_step)
sess.run(tf.global_variables_initializer())
# global_vars = tf.global_variables()
# sess.run(tf.variables_initializer(var_list=[v for v in global_vars if 'adam' in v.name]))
print('Optimization starts.')
print('{:>16s}'.format("iteration")+'{:>16s}'.format("objective"))
for i in range(1,num_iter+1):
_cost,_ = sess.run([cost,optimizer])
if i==1 or i%print_every==0 or i==num_iter:
print('{:>16d}'.format(i)+'{:>16.3f}'.format(_cost))
print('Optimization ends.')
if plot_:
print('Plotting...')
plot_model(model,t,Y,Nw=50)
return model
def npde_fit(sess,
t,
Y,
model='sde',
sf0=1.0,
ell0=[2, 2],
sfg0=1.0,
ellg0=[1e5],
W=6,
ktype="id",
whiten=True,
Nw=50,
fix_ell=False,
fix_sf=False,
fix_Z=False,
fix_U=False,
fix_sn=False,
fix_ellg=False,
fix_sfg=False,
fix_Zg=True,
fix_Ug=False,
num_iter=500,
print_int=10,
eta=5e-3,
dec_step=20,
dec_rate=0.99,
plot_=True):
""" Fits the NPDE model to a dataset and returns the fitted object
Args:
sess: TensowFlow session needed for initialization and optimization
t: Python array of numpy vectors storing observation times
Y: Python array of numpy matrices storing observations. Observations
are stored in rows.
model: 'sde' or 'ode'
sf0: Integer initial value of the signal variance of drift GP
ell0: Python/numpy array of floats for the initial value of the
lengthscale of drift GP
sfg0: Integer initial value of the signal variance of diffusion GP
ellg0: Python/numpy array of a single float for the initial value of the
lengthscale of diffusion GP
W: Integer denoting the width of the inducing point grid. If the problem
dimension is D, total number of inducing points is W**D
ktype: Kernel type. We have made experiments only with Kronecker kernel,
denoted by 'id'. The other kernels are not supported.
whiten: Boolean. Currently we perform the optimization only in the
white domain
Nw: Integer number of samples used for optimization in SDE model
fix_ell: Boolean - whether drift GP lengthscale is fixed or optimized
fix_sf: Boolean - whether drift GP signal variance is fixed or optimized
fix_Z: Boolean - whether drift GP inducing locations are fixed or optimized
fix_U: Boolean - whether drift GP inducing vectors are fixed or optimized
fix_sn: Boolean - whether noise variance is fixed or optimized
fix_ellg: Boolean - whether diffusion GP lengthscale is fixed or optimized
fix_sfg: Boolean - whether diffusion GP signal variance is fixed or optimized
fix_Zg: Boolean - whether diffusion GP inducing locations are fixed or optimized
fix_Ug: Boolean - whether diffusion GP inducing vectors are fixed or optimized
num_iter: Integer number of optimization steps
num_iter: Integer interval of optimization logs to be printed
eta: Float step size used in optimization, must be carefully tuned
dec_step: Float decay interval of the step size
dec_rate: Float decay rate of the step size
plot_: Boolean for plotting the model fit. Valid only for demo
Returns:
npde: Fitted model
"""
print('Model being initialized...')
def init_U0(Y=None, t=None, kern=None, Z0=None, whiten=None):
Ug = (Y[1:, :] - Y[:-1, :]) / np.reshape(t[1:] - t[:-1], (-1, 1))
tmp = NPODE(Y[0, :].reshape((1, -1)),
t,
Y,
Z0=Y[:-1, :],
U0=Ug,
sn0=0,
kern=kern,
jitter=0.2,
whiten=False,
fix_Z=True,
fix_U=True,
fix_sn=True)
U0 = tmp.f(X=Z0)
if whiten:
Lz = tf.cholesky(kern.K(Z0))
U0 = tf.matrix_triangular_solve(Lz, U0, lower=True)
U0 = sess.run(U0)
return U0
D = len(ell0)
Nt = len(Y)
x0 = np.zeros((Nt, D))
Ys = np.zeros((0, D))
for i in range(Nt):
x0[i, :] = Y[i][0, :]
Ys = np.vstack((Ys, Y[i]))
maxs = np.max(Ys, 0)
mins = np.min(Ys, 0)
grids = []
for i in range(D):
grids.append(np.linspace(mins[i], maxs[i], W))
vecs = np.meshgrid(*grids)
Z0 = np.zeros((0, W**D))
for i in range(D):
Z0 = np.vstack((Z0, vecs[i].T.flatten()))
Z0 = Z0.T
tmp_kern = OperatorKernel(sf0, ell0, ktype="id", fix_ell=True, fix_sf=True)
U0 = np.zeros(Z0.shape, dtype=np.float64)
for i in range(len(Y)):
U0 += init_U0(Y[i], t[i], tmp_kern, Z0, whiten)
U0 /= len(Y)
sn0 = 0.5 * np.ones(D)
Ug0 = np.ones([Z0.shape[0], 1]) * 0.01
ell0 = np.asarray(ell0, dtype=np.float64)
ellg0 = np.asarray(ellg0, dtype=np.float64)
kern = OperatorKernel(sf0=sf0,
ell0=ell0,
ktype=ktype,
fix_ell=fix_ell,
fix_sf=fix_sf)
if model is 'ode':
npde = NPODE(x0=x0,
t=t,
Y=Y,
Z0=Z0,
U0=U0,
sn0=sn0,
kern=kern,
whiten=whiten,
fix_Z=fix_Z,
fix_U=fix_U,
fix_sn=fix_sn)
with tf.name_scope("cost"):
X = npde.forward()
ll = []
for i in range(len(X)):
mvn = tf.contrib.distributions.MultivariateNormalFullCovariance(
loc=X[i], covariance_matrix=tf.diag(npde.sn))
ll.append(tf.reduce_sum(mvn.log_prob(Y[i])))
ll = tf.reduce_logsumexp(ll)
ode_prior = npde.build_prior()
cost = -(ll + ode_prior)
elif model is 'sde':
diffus = BrownianMotion(sf0=sfg0,
ell0=ellg0,
U0=Ug0,
Z0=Z0,
whiten=whiten,
fix_sf=fix_sfg,
fix_ell=fix_ellg,
fix_Z=fix_Zg,
fix_U=fix_Ug)
npde = NPSDE(x0=x0,
t=t,
Y=Y,
Z0=Z0,
U0=U0,
sn0=sn0,
kern=kern,
diffus=diffus,
whiten=whiten,
fix_Z=fix_Z,
fix_U=fix_U,
fix_sn=fix_sn)
with tf.name_scope("cost"):
Xs = npde.forward(Nw=Nw)
ll = 0
for i in range(len(Y)):
mvn = tf.contrib.distributions.MultivariateNormalFullCovariance(
loc=Y[i], covariance_matrix=tf.diag(npde.sn))
ll_i = tf.stack([
mvn.log_prob(Xs[i][j, :, :]) for j in range(Xs[i].shape[0])
]) # Nw x D
ll_i = tf.reduce_sum(
tf.log(tf.reduce_mean(tf.exp(ll_i), axis=0)))
ll += ll_i
sde_prior = npde.build_prior()
cost = -(ll + sde_prior)
else:
raise NotImplementedError(
"model parameter should be either 'ode' or 'sde', not {:s}\n".
format(model))
print('Adam optimizer being initialized...')
global_step = tf.Variable(0,
dtype=tf.int32,
trainable=False,
name='global_step')
expdec = tf.train.exponential_decay(eta,
global_step,
dec_step,
dec_rate,
staircase=True)
optimizer = tf.train.AdamOptimizer(expdec).minimize(cost, global_step)
sess.run(tf.global_variables_initializer())
print('Optimization starts.')
print('{:>16s}'.format("iteration") + '{:>16s}'.format("objective"))
for i in range(1, num_iter + 1):
_cost, _ = sess.run([cost, optimizer])
if i == 1 or i % print_int == 0 or i == num_iter:
print('{:>16d}'.format(i) + '{:>16.3f}'.format(_cost))
print('Optimization ends.')
if plot_:
print('Plotting...')
plot_model(npde, Nw=50)
return npde
