def create_all_amplitude_perturbation_corrs(folder_name,
params,
start,
stop,
with_square,
with_square_cov,
after_softmax,
n_samples,
perturbations='default'):
if perturbations == 'default':
perturbations = (
(
'no_dev',
FuncAndArgs(
rand_diff,
with_blocks=False, #just return 1
deviation_func=lambda arr, axis, keepdims: 1)),
('rand_mad',
FuncAndArgs(rand_diff,
with_blocks=False,
deviation_func=median_absolute_deviation)),
('rand_std',
FuncAndArgs(rand_diff, with_blocks=False, deviation_func=np.std)),
)
elif perturbations == 'simple_no_scale':
perturbations = (
(
'no_dev',
FuncAndArgs(
rand_diff,
with_blocks=False, #just return 1
deviation_func=lambda arr, axis, keepdims: 1)), )
assert not (with_square and with_square_cov)
res_pool = ResultPool()
res_pool.load_results(folder_name, params=params)
res_file_names = res_pool.result_file_names()
all_base_names = [
name.replace('.result.pkl', '') for name in res_file_names
]
start = start or 0
stop = stop or len(all_base_names)
for i_file, base_name in enumerate(all_base_names[start:stop]):
log.info("Running {:s} ({:d} of {:d})".format(base_name,
i_file + start + 1,
stop))
create_amplitude_perturbation_corrs(base_name,
with_square=with_square,
with_square_cov=with_square_cov,
after_softmax=after_softmax,
n_samples=n_samples,
perturbations=perturbations)
def get_templates():
return {'categorical_crossentropy': lambda : categorical_crossentropy,
'tied_loss': lambda : FuncAndArgs(sum_of_losses,
loss_expressions=[categorical_crossentropy,
tied_neighbours_cnt_model ,
]
)
}
def _create_train_function(self):
# Maybe replace self.exp.final_layer by self.cnt_model?
# not clear to me why I am using self.exp.final_layer here
targets = T.ivector()
input_var = get_input_var(self.exp.final_layer)
updates_expression = FuncAndArgs(adam,
learning_rate=self.learning_rate)
prediction = lasagne.layers.get_output(
self.exp.final_layer,
deterministic=self.deterministic_training,
input_var=input_var,
inputs=input_var)
# Loss function might need layers or not...
loss_fn = masked_loss_func(categorical_crossentropy)
loss = loss_fn(prediction, targets).mean()
# create parameter update expressions
params = lasagne.layers.get_all_params(self.exp.final_layer,
trainable=True)
updates = updates_expression(loss, params)
if self.exp.updates_modifier is not None:
# put norm constraints on all layer, for now fixed to max kernel norm
# 2 and max col norm 0.5
updates = self.exp.updates_modifier.modify(updates,
self.exp.final_layer)
# store only the parameters for training,
# assumes parameters for layers already set
self.train_params = []
all_update_params = updates.keys()
for update_param in all_update_params:
if update_param not in params:
self.train_params.append(update_param)
self.train_func = theano.function([input_var, targets],
updates=updates)
# Set optimizer/train parameter values if not done
if self.train_param_values is not None:
log.info("Setting train parameter values")
for param, val in zip(self.train_params, self.train_param_values):
param.set_value(val)
log.info("...Done setting parameter train values")
else:
log.info(
"Not setting train parameter values, optimization values "
"start from scratch (model params may be loaded anyways.)")
if len(sys.argv) > 2:
stop = int(sys.argv[2])
folder = 'data/models/paper/ours/cnt/deep4/car/'
params = dict(cnt_preprocessors="$cz_zero_resample_car_demean",
trial_start=1500,
trial_stop=4000) #3500 4000 # 1000 1500
with_square = False
with_square_cov = False
with_blocks = False
after_softmax = False
n_samples = 300
perturbations = (
(
'no_dev',
FuncAndArgs(
rand_diff,
with_blocks=False, #just return 1
deviation_func=lambda arr, axis, keepdims: 1)), ) #,
# ('rand_mad',
# FuncAndArgs(rand_diff,
# with_blocks=False,
# deviation_func=median_absolute_deviation)),
# ('rand_std', FuncAndArgs(rand_diff,
# with_blocks=False,
# deviation_func=np.std)),
# )
create_all_amplitude_perturbation_corrs(folder,
params=params,
start=start,
stop=stop,
with_square=with_square,