def check_shallow_env_corrs():
i_layer = 5
for i_exp in xrange(1, 21):
file_base_name = 'data/models-backup/paper/ours/cnt/shallow//car/{:d}'.format(
i_exp)
result = np.load(file_base_name + '.result.pkl')
log.info("Running {:d} of {:d}:\n{:s}".format(i_exp, 20,
file_base_name))
exp_shallow, model_shallow = load_exp_and_model(file_base_name)
exp_shallow.dataset.load()
env_filename = dataset_to_env_file(
result.parameters['dataset_filename'])
trial_env = load_trial_env(env_filename,
model_shallow,
i_layer,
exp_shallow.dataset.sets[0],
n_inputs_per_trial=2)
trial_acts = compute_trial_acts(model_shallow, i_layer,
exp_shallow.iterator,
exp_shallow.dataset.train_set)
log.info("Compute topo corrs")
topo_corrs = compute_topo_corrs(trial_env, trial_acts)
topo_corrs_old = np.load(file_base_name + '5.env_corrs.npy')
diff = np.mean(np.abs(topo_corrs - topo_corrs_old))
log.info("Diff ok {:f}".format(diff))
assert diff < 1e-3, ("Diff too big for {:s}: {:f}".format(
file_base_name, diff))
def compute_trial_topo_corrs(model, i_layer, train_set, iterator, trial_env):
"""Now split per class."""
trial_acts = compute_trial_acts(model, i_layer, iterator, train_set)
trial_labels = determine_trial_labels(train_set, iterator)
topo_corrs_per_class = []
for i_label in xrange(trial_labels.shape[1]):
this_trial_mask = trial_labels[:,i_label] == 1
this_trial_env = trial_env[:, this_trial_mask]
this_trial_acts = trial_acts[this_trial_mask]
topo_corrs = compute_topo_corrs(this_trial_env, this_trial_acts)
topo_corrs_per_class.append(topo_corrs)
topo_corrs_per_class = np.array(topo_corrs_per_class)
# TODELETE (old without split input per label):
# topo_corrs = compute_topo_corrs(trial_env, trial_acts)
# TODO compute within trial corrs -> remove mean for each trial
# compute between trial corrs -> compute means for trial env and trial acts
# with keepdims=True
return topo_corrs_per_class
def unit_output_class_corrs(model, iterator, train_set, i_layer):
# only need targets, ignore trials
# always get targets as from final layer
this_final_layer = lasagne.layers.get_all_layers(model)[-1]
_, targets = get_trials_targets(train_set,
get_n_sample_preds(this_final_layer), iterator)
trial_acts = compute_trial_acts(this_final_layer, i_layer, iterator,
train_set)
# only take those targets where we have predictions for
# a bit hacky: we know targets are same for each trial, so we just
# take last ones, eventhough they come form overlapping batches
# targets are #trials x #samples x #classes
unmeaned_targets = targets - np.mean(targets, axis=(1), keepdims=True)
assert np.all(unmeaned_targets == 0), ("Each trial should only have one "
"unique label")
relevant_targets = targets[:,:trial_acts.shape[2]]
unit_class_corrs = wrap_reshape_topo(corr, trial_acts, relevant_targets,
axis_a=(0,2), axis_b=(0,1))
return unit_class_corrs
def compute_trial_topo_corrs(model, i_layer, train_set, iterator, trial_env,
split_per_class):
"""Now with split per class arg."""
trial_acts = compute_trial_acts(model, i_layer, iterator, train_set)
trial_labels = determine_trial_labels(train_set, iterator)
if split_per_class:
topo_corrs_per_class = []
for i_label in xrange(trial_labels.shape[1]):
this_trial_mask = trial_labels[:,i_label] == 1
this_trial_env = trial_env[:, this_trial_mask]
this_trial_acts = trial_acts[this_trial_mask]
topo_corrs = compute_topo_corrs(this_trial_env, this_trial_acts)
topo_corrs_per_class.append(topo_corrs)
topo_corrs_per_class = np.array(topo_corrs_per_class)
# TODO compute within trial corrs -> remove mean for each trial
# compute between trial corrs -> compute means for trial env and trial acts
# with keepdims=True
return topo_corrs_per_class
else:
topo_corrs = compute_topo_corrs(trial_env, trial_acts)
return topo_corrs
def check_shallow_env_corrs():
i_layer = 5
for i_exp in xrange(1,21):
file_base_name = 'data/models-backup/paper/ours/cnt/shallow//car/{:d}'.format(
i_exp)
result = np.load(file_base_name + '.result.pkl')
log.info("Running {:d} of {:d}:\n{:s}".format(i_exp, 20,
file_base_name))
exp_shallow, model_shallow = load_exp_and_model(file_base_name)
exp_shallow.dataset.load()
env_filename = dataset_to_env_file(result.parameters['dataset_filename'])
trial_env = load_trial_env(env_filename, model_shallow,
i_layer, exp_shallow.dataset.sets[0],
n_inputs_per_trial=2)
trial_acts = compute_trial_acts(model_shallow, i_layer,
exp_shallow.iterator, exp_shallow.dataset.train_set)
log.info("Compute topo corrs")
topo_corrs = compute_topo_corrs(trial_env, trial_acts)
topo_corrs_old = np.load(file_base_name + '5.env_corrs.npy')
diff = np.mean(np.abs(topo_corrs - topo_corrs_old))
log.info("Diff ok {:f}".format(diff))
assert diff < 1e-3, (
"Diff too big for {:s}: {:f}".format(file_base_name, diff))
def unit_output_class_corrs(model, iterator, train_set, i_layer):
# only need targets, ignore trials
# always get targets as from final layer
this_final_layer = lasagne.layers.get_all_layers(model)[-1]
_, targets = get_trials_targets(train_set,
get_n_sample_preds(this_final_layer),
iterator)
trial_acts = compute_trial_acts(this_final_layer, i_layer, iterator,
train_set)
# only take those targets where we have predictions for
# a bit hacky: we know targets are same for each trial, so we just
# take last ones, eventhough they come form overlapping batches
# targets are #trials x #samples x #classes
unmeaned_targets = targets - np.mean(targets, axis=(1), keepdims=True)
assert np.all(unmeaned_targets == 0), ("Each trial should only have one "
"unique label")
relevant_targets = targets[:, :trial_acts.shape[2]]
unit_class_corrs = wrap_reshape_topo(corr,
trial_acts,
relevant_targets,
axis_a=(0, 2),
axis_b=(0, 1))
return unit_class_corrs
