def fit_ln(cells, train_stimuli, exptdate, stim_shape, l2=1e-3, readme=None):
"""Fits an LN model using keras"""
ncells = len(cells)
batchsize = 5000
# get the layers
layers = ln(stim_shape, ncells, weight_init='normal', l2_reg=l2)
# compile it
model = sequential(layers, RMSprop(lr=1e-4), loss='poisson')
# load experiment data
test_stimuli = ['whitenoise', 'naturalscene']
data = Experiment(exptdate,
cells,
train_stimuli,
test_stimuli,
stim_shape[0],
batchsize,
nskip=6000)
# create a monitor that keeps track of progress
monitor = KerasMonitor('ln', model, data, readme, save_every=20)
# train
train(model, data, monitor, num_epochs=30)
return model
def fit_convnet(cells, train_stimuli, exptdate, nclip=0, readme=None):
"""Main script for fitting a convnet
author: Niru Maheswaranathan
"""
stim_shape = (40, 50, 50)
ncells = len(cells)
batchsize = 5000
# get the convnet layers
layers = convnet(stim_shape, ncells, num_filters=(8, 16),
filter_size=(15, 7), weight_init='normal',
l2_reg_weights=(0.01, 0.01, 0.01),
l1_reg_activity=(0.0, 0.0, 0.001),
dropout=(0.1, 0.0))
# compile the keras model
model = sequential(layers, 'adam', loss='poisson')
# load experiment data
test_stimuli = ['whitenoise', 'naturalscene']
data = Experiment(exptdate, cells, train_stimuli, test_stimuli, stim_shape[0], batchsize, nskip=nclip)
# create a monitor to track progress
monitor = KerasMonitor('convnet', model, data, readme, save_every=20)
# train
train(model, data, monitor, num_epochs=50)
return model
def fit_fc_rnn(expt, stim):
train(fc_rnn,
expt,
stim,
model_args=("flatten", "mse"),
lr=1e-3,
nb_epochs=250,
val_split=0.05)
def fit_bn_spat_cnn(expt, stim):
train(bn_spat_cnn,
expt,
stim,
model_args=("spatial"),
lr=1e-2,
nb_epochs=250,
val_split=0.05)
def fit_copy_cnn(expt, stim):
train(copy_cnn,
expt,
stim,
model_args=(),
lr=1e-3,
nb_epochs=250,
val_split=0.05)
def fit_bn_rnn(expt, stim):
train(bn_rnn,
expt,
stim,
lr=5e-3,
model_args=("add_dim", "mse"),
nb_epochs=250,
val_split=0.05,
bz=1024)
def fit_cn_tcn(expt, stim):
train(cn_tcn,
expt,
stim,
model_args=("add_dim", "mse"),
lr=1e-3,
nb_epochs=250,
val_split=0.05,
bz=2048)
def fit_conv_to_rnn(expt, stim):
train(conv_to_rnn,
expt,
stim,
model_args=("add_dim", "mse"),
lr=1e-4,
nb_epochs=250,
val_split=0.05,
bz=1024)
def fit_conv_to_lstm(expt, stim):
train(conv_to_lstm,
expt,
stim,
model_args=("add_dim"),
lr=0.1,
nb_epochs=250,
val_split=0.05,
bz=2048)
def fit_conv_lstm(expt, stim):
train(conv_lstm,
expt,
stim,
model_args=("add_dim"),
lr=1e-3,
bz=128,
nb_epochs=250,
val_split=0.05)
def fit_ln(expt, ci, stim, activation, l2_reg=0.1):
if activation.lower() == 'rbf':
model_args = (30, 6)
else:
model_args = ()
model = functools.partial(linear_nonlinear,
activation=activation,
l2_reg=l2_reg)
#tp.banner(f'Training LN-{activation}, expt {args.expt}, {args.stim}, cell {ci+1:02d}')
train(model,
expt,
stim,
model_args=model_args,
lr=1e-2,
nb_epochs=500,
val_split=0.05,
cells=[ci])
def fit_bn_cnn(expt, stim):
train(bn_cnn, expt, stim, lr=1e-2, nb_epochs=250, val_split=0.05)
