def sample_replay_for_update(self):
for i in range(self.state_space_parameters.replay_number):
net = np.random.choice(self.replay_dictionary['net'])
''' Uncomment for training '''
# accuracy_best_val = self.replay_dictionary[self.replay_dictionary['net'] == net]['accuracy_best_val'].values[0]
# accuracy_last_val = self.replay_dictionary[self.replay_dictionary['net'] == net]['accuracy_last_val'].values[0]
# train_flag = self.replay_dictionary[self.replay_dictionary['net'] == net]['train_flag'].values[0]
''' Uncommenting ends '''
loss_inverse = self.replay_dictionary[
self.replay_dictionary['net'] == net]['loss_inverse'].values[0]
loss = self.replay_dictionary[self.replay_dictionary['net'] ==
net]['loss'].values[0]
computeLoss_flag = self.replay_dictionary[
self.replay_dictionary['net'] ==
net]['computeLoss_flag'].values[0]
state_list = self.stringutils.convert_model_string_to_states(
cnn.parse('net', net))
# print('During update:')
# for state in state_list:
# print('{} {} {} {} {} {} {} {}'.format(state.layer_type,state.layer_depth, state.filter_depth,\
# state.filter_size, state.stride, state.image_size, state.fc_size, state.terminate))
# print(state_list)
''' Uncomment while training '''
# state_list = self.stringutils.remove_drop_out_states(state_list)
''' Uncommenting ends '''
# state_list = [self.enum.bucket_state(state) for state in state_list]
''' Uncomment while training '''
# if train_flag == True:
# self.update_q_value_sequence(state_list, self.accuracy_to_reward(accuracy_best_val))
''' Uncommenting ends '''
if computeLoss_flag == True:
self.update_q_value_sequence(
state_list, self.accuracy_to_reward(loss_inverse))
def main(_):
_model = __import__(
'models.' + 'cifar10', #args.model,
globals(),
locals(),
['state_space_parameters', 'hyper_parameters'],
-1)
########### Benchmark end-to-end network ##########
net_list = cnn.parse('net', args.net_string)
StateStringUtils(
_model.state_space_parameters).convert_model_string_to_states(net_list)
def sample_replay_for_update_in_latency(self):
# Experience replay to update Q-Values
for i in range(self.state_space_parameters.replay_number):
net = np.random.choice(self.replay_dictionary['net'])
latency = self.replay_dictionary[self.replay_dictionary['net'] ==
net]['latency'].values[0]
state_list = self.stringutils.convert_model_string_to_states(
cnn.parse('net', net))
state_list = self.stringutils.remove_drop_out_states(state_list)
# Convert States so they are bucketed
state_list = [
self.enum.bucket_state(state) for state in state_list
]
self.update_q_value_sequence(
state_list, self.accuracy_to_reward_in_latency(latency))
to_state = State(
state_list=gen.qstore.q[start_state.as_tuple()]['to_states'][0])
gen.qstore.q = enum.enumerate_state(to_state, gen.qstore.q)
gen.qstore.q[to_state.as_tuple()]['utilities'][0] = 0.6
gen._update_q_value(start_state, to_state, 0.4)
assert (gen.qstore.q[start_state.as_tuple()]['utilities'][0] == 0.5 +
gen.learning_rate * (0.4 + 0.6 - 0.5))
for i in range(100):
previous_action_values = gen.qstore.q.copy()
net, acc_best_val, acc_last_val, acc_best_test, acc_last_test = gen.generate_net(
)
stringutils = StateStringUtils(state_space_parameters)
print net
net_lists = cnn.parse('net', net)
print '---------------------'
print 'Net String'
print net
print '----------------------'
assert (net == stringutils.state_list_to_string(
stringutils.convert_model_string_to_states(net_lists)))
for key in previous_action_values.keys():
old_to_state = [
state for state in previous_action_values[key]['to_states']
]
new_to_state = [state for state in gen.qstore.q[key]['to_states']]
for i in range(len(old_to_state)):
assert (old_to_state[i] in new_to_state)
from yapps import cli_tool
cli_tool.generate('cnn.g')
import cnn
print cnn.parse('net', '[C(1,2,3), C(1,2,3), SM(10)]')
print cnn.parse('net', '[C(1,2,3), C(4, 5, 6), FC(40), FC(40), SM(10)]')
print cnn.parse('net', '[C(1,1,1), S{[C(2,2,2), C(22,22,22)], [C(3,3,3)]}, C(4,4,4), P(2), SM(10)]')
print cnn.parse('net', '[C(1,2,3), NIN(100), BN, NIN(100), GAP(100)]')