def run_BD(i):
tf.reset_default_graph()
data = DataReader.read_BD()
ncells = configs[i]['cells']
learning_rate = configs[i]['lr']
group = configs[i]['g']
iters = configs[i]['iters']
model_path = configs[i]['model_path']
output_path = Paths.local_path + 'BD/rnn-opt-rand-init/' + 'run_' + str(
configs[i]['s']) + '/' + str(ncells) + 'cells/' + group + '/'
with LogFile(output_path, 'run.log'):
DLogger.logger().debug("group: " + str(group))
gdata = data.loc[data.diag == group]
ids = gdata['id'].unique().tolist()
dftr = pd.DataFrame({'id': ids, 'train': 'train'})
tdftr = pd.DataFrame({'id': ids, 'train': 'test'})
train, test = DataProcess.train_test_between_subject(
gdata, pd.concat((dftr, tdftr)),
[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
train = DataProcess.merge_data(train)
DLogger.logger().debug("total points: " + str(get_total_pionts(train)))
worker = LSTMBeh(2, 0, n_cells=ncells)
OptBEH.optimise(worker,
output_path,
train,
None,
learning_rate=learning_rate,
global_iters=iters,
load_model_path=model_path)
def run_BD_Q(i):
tf.reset_default_graph()
learning_rate = configs[i]['lr']
group = configs[i]['g']
cv_index = configs[i]['cv_index']
indx_data = cv_lists_group[group][cv_index]
gdata = data.loc[data.diag == group]
train, test = DataProcess.train_test_between_subject(
gdata, indx_data, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
output_path = Paths.local_path + 'BD/gql10d-ml-cv/' + group + '/' + 'fold' + str(
cv_index) + '/'
with LogFile(output_path, 'run.log'):
indx_data.to_csv(output_path + 'train_test.csv')
worker = GQL.get_instance(2, 10, {})
train = DataProcess.merge_data(train)
OptML.optimise(worker,
output_path,
train,
test,
learning_rate=learning_rate,
global_iters=1000)
def run_BD_RNN(i):
tf.reset_default_graph()
ncells = configs[i]['cells']
learning_rate = configs[i]['lr']
group = configs[i]['g']
cv_index = configs[i]['cv_index']
output_path = Paths.local_path + 'BD/rnn-cv/' + str(
ncells) + 'cells/' + group + '/' + 'fold' + str(cv_index) + '/'
with LogFile(output_path, 'run.log'):
indx_data = cv_lists_group[group][cv_index]
gdata = data.loc[data.diag == group]
train, test = DataProcess.train_test_between_subject(
gdata, indx_data, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
indx_data.to_csv(output_path + 'train_test.csv')
train_merged = DataProcess.merge_data(train)
DLogger.logger().debug("total points: " +
str(get_total_pionts(train_merged)))
del train
worker = LSTMBeh(2, 0, n_cells=ncells)
OptBEH.optimise(worker,
output_path,
train_merged,
None,
learning_rate=learning_rate,
global_iters=3000,
load_model_path='../inits/rnn-init/' + str(ncells) +
'cells/model-final/')
def run_BD(i):
tf.reset_default_graph()
data = DataReader.read_BD()
learning_rate = configs[i]['lr']
group = configs[i]['g']
output_path = Paths.local_path + 'BD/gql-ml-opt/' + group + '/'
with LogFile(output_path, 'run.log'):
DLogger.logger().debug("group: " + str(group))
gdata = data.loc[data.diag == group]
ids = gdata['id'].unique().tolist()
dftr = pd.DataFrame({'id': ids, 'train': 'train'})
tdftr = pd.DataFrame({'id': ids, 'train': 'test'})
train, test = DataProcess.train_test_between_subject(
gdata, pd.concat((dftr, tdftr)),
[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
DLogger.logger().debug("total points: " + str(get_total_pionts(train)))
worker = GQL.get_instance(2, 2, {})
train = DataProcess.merge_data(train)
OptML.optimise(worker,
output_path,
train,
test,
global_iters=1000,
learning_rate=learning_rate)
def run_BD_GQL(i):
tf.reset_default_graph()
learning_rate = configs[i]['lr']
group = configs[i]['g']
cv_index = configs[i]['cv_index']
iters = configs[i]['iters']
output_path = Paths.local_path + 'BD/gql-ml-rand-opt/' + group + '/' + 'fold' + str(
cv_index) + '/'
with LogFile(output_path, 'run.log'):
indx_data = cv_lists_group[group][cv_index]
gdata = data.loc[data.diag == group]
train, test = DataProcess.train_test_between_subject(
gdata, indx_data, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
indx_data.to_csv(output_path + 'train_test.csv')
DLogger.logger().debug("total points: " + str(get_total_pionts(train)))
worker = GQL.get_instance(2, 2, {})
train = DataProcess.merge_data(train)
OptML.optimise(worker,
output_path,
train,
None,
learning_rate=learning_rate,
global_iters=iters)
def GQL_classify_subjects():
tf.reset_default_graph()
data = DataReader.read_BD()
ids = data['id'].unique().tolist()
dftr = pd.DataFrame({'id': ids, 'train': 'train'})
train, test = DataProcess.train_test_between_subject(
data, dftr, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
model_iter = 'model-final'
df = pd.DataFrame(columns=('model', 'id', 'loss'))
config = tf.ConfigProto(device_count={'GPU': 0})
subj_paths = finding_CV(Paths.rest_path + 'archive/beh/gql-ml-pred-diag/')
worker = GQL.get_instance(2, 2, {})
worker.set_params(OptML.get_variables(worker.get_params()))
for k, tr in train.iteritems():
for g, p in subj_paths[k].iteritems():
DLogger.logger().debug('subject ' + k + ' group ' + g + ' path ' +
p)
model_path = p + model_iter + '/'
with tf.Session(config=config) as sess:
load_model(sess, model_path)
total_loss = 0
for v in tr:
ell, _, _ = worker.simulate(sess, v['reward'], v['action'])
total_loss += -ell
df.loc[len(df)] = [g, k, total_loss]
df.to_csv(Paths.local_path + 'BD/gql_diag.csv')
def run_BD(i):
data = DataReader.read_BD()
ncells = configs[i]['cells']
group = configs[i]['g']
input_path = Paths.rest_path + 'archive/beh/rnn-opt-rand-init/' + 'run_' + \
str(configs[i]['s']) + '/' + str(ncells) + 'cells/' + group + '/model-final/'
output_path = Paths.local_path + 'BD/rnn-opt-rand-init-evals/' + 'run_' + \
str(configs[i]['s']) + '/' + str(ncells) + 'cells/' + group + '/'
gdata = data.loc[data.diag == group]
ids = gdata['id'].unique().tolist()
dftr = pd.DataFrame({'id': ids, 'train': 'train'})
tdftr = pd.DataFrame({'id': ids, 'train': 'test'})
train, test = DataProcess.train_test_between_subject(
gdata, pd.concat((dftr, tdftr)),
[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
test = DataProcess.merge_data(test)
tf.reset_default_graph()
worker = LSTMBeh(2, 0, ncells)
saver = tf.train.Saver(max_to_keep=None)
with tf.Session() as sess:
DLogger.logger().debug("loading model from: " + str(input_path))
ckpt = tf.train.get_checkpoint_state(input_path)
tf.train.import_meta_graph(input_path + 'model.cptk.meta')
saver.restore(sess, ckpt.model_checkpoint_path)
for k, tr in test.iteritems():
for v in tr:
_, _, _, ell = worker.simulate(sess, v['reward'], v['action'],
v['state'])
DLogger.logger().debug("input path: " + input_path)
DLogger.logger().debug("output path: " + output_path)
DLogger.logger().debug("total nlp: {} ".format(str(ell)))
return pd.DataFrame({
'total nlp': [ell],
'group': group,
'cell': ncells,
'fold': None,
'model_iter': 'model-final',
's': configs[i]['s']
})
def simulate_model(input_folder, output_folder, data, n_cells):
dftr = pd.DataFrame({'id': data['id'].unique().tolist(), 'train': 'train'})
train, _ = DataProcess.train_test_between_subject(
data, dftr, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
tf.reset_default_graph()
worker = LSTMBeh(2, 0, n_cells)
DLogger.logger().debug('started simulations')
Simulator.simulate_worker(worker, input_folder, train, output_folder)
def run_BD_RNN(i):
tf.reset_default_graph()
ncells = configs[i]['cells']
lr = configs[i]['lr']
output_path = Paths.local_path + 'BD/rnn-init/' + str(ncells) + 'cells/'
with LogFile(output_path, 'run.log'):
ids = data['id'].unique().tolist()
dftr = pd.DataFrame({'id': ids, 'train': 'train'})
train, test = DataProcess.train_test_between_subject(
data, dftr, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
train = DataProcess.merge_data(train)
DLogger.logger().debug("total points: " + str(get_total_pionts(train)))
worker = LSTMBeh(2, 0, n_cells=ncells)
lrh.OptBEH.optimise(worker,
output_path,
train,
None,
learning_rate=lr,
global_iters=0)
def test_and_save(label, output_path, saver, sess, test, worker):
policies_list = {}
if test is not None and len(test) > 0:
DLogger.logger().debug("started testing...")
output = [['sid', 'pre accu', 'post accu']]
for k, tr in test.iteritems():
merged_data = DataProcess.merge_data({'tmp': tr})['merged'][0]
ell, policies, _ = worker.simulate(sess, merged_data['reward'],
merged_data['action'])
policies_list[k] = {}
post_accu = []
for i in range(len(tr)):
v = tr[i]
pol = policies[i, :v['action'].shape[1]]
post_accu.append(Assessor.evaluate_fit(pol, v['action']))
policies_list[k][v['block']] = pol
output.append([v['id'], None, post_accu])
DLogger.logger().debug("finished testing.")
if output_path is not None:
Export.write_to_csv(output, output_path, 'accuracy.csv')
Export.policies(policies_list, output_path, 'policies.csv')
if output_path is not None:
vparams = sess.run(worker.get_params())
params = [worker.get_param_names(), vparams]
Export.write_to_csv(params, output_path, 'params.csv')
trans_params = [
f(v)
for f, v in zip(worker.get_trans_func(), worker.get_params())
]
output = sess.run(trans_params)
params = [worker.get_param_names(), output]
Export.write_to_csv(params, output_path, 'params-trans.csv')
if output_path is not None and saver is not None:
model_path = output_path + 'model-' + str(label) + '/'
ensure_dir(model_path)
saver.save(sess, model_path + 'model.cptk')
train_writer = tf.summary.FileWriter(model_path)
train_writer.add_graph(sess.graph)
train_writer.close()
return policies_list
def RNN_classify_subjects():
data = DataReader.read_BD()
ids = data['id'].unique().tolist()
dftr = pd.DataFrame({'id': ids, 'train': 'train'})
train, test = DataProcess.train_test_between_subject(
data, dftr, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
n_cells = {'Healthy': 10, 'Depression': 10, 'Bipolar': 20}
model_iter = 'model-final'
df = pd.DataFrame(columns=('model', 'id', 'loss'))
config = tf.ConfigProto(device_count={'GPU': 0})
subj_paths = finding_CV(Paths.rest_path + 'archive/beh/rnn-pred-diag/')
for k, tr in train.iteritems():
for g, p in subj_paths[k].iteritems():
tf.reset_default_graph()
worker = LSTMBeh(2, 0, n_cells[g])
saver = tf.train.Saver(max_to_keep=5)
DLogger.logger().debug('subject ' + k + ' group ' + g + ' path ' +
p)
model_path = p + model_iter + '/'
ckpt = tf.train.get_checkpoint_state(model_path)
tf.train.import_meta_graph(model_path + 'model.cptk.meta')
with tf.Session(config=config) as sess:
saver.restore(sess, ckpt.model_checkpoint_path)
total_loss = 0
for v in tr:
policies, c_track, h_track, loss = worker.simulate(
sess, v['reward'], v['action'], v['state'])
total_loss += loss
df.loc[len(df)] = [g, k, total_loss]
df.to_csv(Paths.local_path + 'BD/rnn_diag.csv')
def evaluate_CV(base_input_folder,
base_output_folder,
test_and_save,
data,
folds,
model_iters,
trials,
random_tie=True):
df = pd.DataFrame()
saver = tf.train.Saver(max_to_keep=5)
config = tf.ConfigProto(device_count={'GPU': 0})
with tf.Session(config=config) as sess:
for group in sorted(folds.keys()):
for fold in folds[group]:
for model_iter in model_iters:
input_folder = base_input_folder + '/' + group + '/' + fold + '/'
if base_output_folder is not None:
output_folder = base_output_folder + group + '/' + fold + '/' + model_iter + '/'
else:
output_folder = None
DLogger.logger().debug(
"input folder: {}".format(input_folder))
DLogger.logger().debug(
"output folder: {}".format(output_folder))
tr_tst = pd.read_csv(input_folder + 'train_test.csv')
if 'ID' in tr_tst:
DLogger.logger().debug(
"id column was not found. Replaced id column with ID."
)
tr_tst['id'] = tr_tst['ID']
tst_ids = tr_tst.loc[tr_tst.train == 'test']['id']
dftr = pd.DataFrame({'id': tst_ids, 'train': 'train'})
train, _ = DataProcess.train_test_between_subject(
data, dftr, trials)
model_path = input_folder + model_iter + '/'
ckpt = tf.train.get_checkpoint_state(model_path)
# tf.train.import_meta_graph(model_path + 'model.cptk.meta')
saver.restore(sess, ckpt.model_checkpoint_path)
policies = test_and_save(sess, train, output_folder)
if output_folder is not None:
Export.export_train(train, output_folder,
'train.csv')
train_merged = Export.merge_train(train)
#add a dummy column at the beginning
train_merged.insert(loc=0, column='tmp', value='')
policies_merged = Export.merge_policies(policies)
#add a dummy column at the beginning
policies_merged.insert(loc=0, column='tmp', value='')
# train_merged = pd.read_csv(output_folder + 'train.csv')
# policies_merged = pd.read_csv(output_folder + 'policies.csv')
acc, nlp, total_nlp = Assessor.evaluate_fit_multi(
policies_merged,
train_merged,
pol_in_log=True,
random_tie=random_tie)
df = df.append(
pd.DataFrame({
'acc': [acc],
'nlp': [nlp],
'total nlp': [total_nlp],
'group': group,
'option': Helper.dicstr({}),
'fold': fold,
'model_iter': model_iter
}))
if base_output_folder is not None:
df.to_csv(base_output_folder + 'accu.csv')
return df
