def test_parameters(_gamma):
start_time = time_module.clock()
environment = GridEnvironment(n_rows, n_cols, circuit)
agent = SimpleDQNAgent(environment)
train_model(environment,
agent,
training_episodes=training_episodes,
should_print=False)
average_test_time = 0.0
average_circuit_depth_overhead = 0.0
for e in range(test_episodes):
actions, circuit_depth = schedule_swaps(environment,
agent,
experience_db=None)
average_test_time += (1.0 / test_episodes) * len(actions)
average_circuit_depth_overhead += (1.0 /
test_episodes) * (circuit_depth - 1)
end_time = time_module.clock()
total_time = end_time - start_time
return average_circuit_depth_overhead, total_time
def train_model_on_random_circuits(model_number):
model_name = "random_circuits_" + str(model_number)
training_circuit_generation_function = lambda: generate_completely_random_circuit(
20, 50).to_dqn_rep()
environment = IBMQ20Tokyo(training_circuit_generation_function())
agent = DoubleDQNAgent(environment)
train_model(
environment,
agent,
training_episodes=training_episodes,
circuit_generation_function=training_circuit_generation_function,
should_print=False)
agent.save_model(model_name)
def train_model_on_full_layers(model_number):
model_name = "full_layers_" + str(model_number)
def training_circuit_generation_function():
return generate_multi_layer_circuit(20, 2).to_dqn_rep()
environment = IBMQ20Tokyo(training_circuit_generation_function())
agent = DoubleDQNAgent(environment)
train_model(
environment,
agent,
training_episodes=training_episodes,
circuit_generation_function=training_circuit_generation_function,
should_print=False)
agent.save_model(model_name)
def train_model_on_random_circuits(f_model_number):
model_name = "random_circuits_" + str(f_model_number)
def training_circuit_generation_function():
return generate_completely_random_circuit(16, 50).to_dqn_rep()
environment = GridEnvironment(4, 4, training_circuit_generation_function())
agent = DoubleDQNAgent(environment)
train_model(
environment,
agent,
training_episodes=training_episodes,
circuit_generation_function=training_circuit_generation_function,
should_print=False)
agent.save_model(model_name)
def perform_run(num_rows, num_cols, training_episodes, test_episodes=100):
"""
:param num_rows: dummy arg, number of rows of TODO???
:param num_cols: dummy arg, number of rows of TODO???
:param training_episodes: number of episodes to train for
:param test_episodes: number of episodes to test for
:return: datapoint: TODO what does that mean?
"""
def circuit_generation_function():
return generate_full_layer_circuit(num_rows * num_cols).to_dqn_rep()
environment = GridEnvironment(num_rows, num_cols,
circuit_generation_function())
agent = DoubleDQNAgent(environment)
start_time = time_module.perf_counter()
# Pass the number of training steps and episodes to control how much we train for
train_model(environment,
agent,
training_episodes=training_episodes,
circuit_generation_function=circuit_generation_function,
should_print=True)
average_circuit_depth_overhead = 0.0
for e in range(test_episodes):
actions, circuit_depth = schedule_swaps(
environment,
agent,
circuit=circuit_generation_function(),
experience_db=None)
average_circuit_depth_overhead += (1.0 /
test_episodes) * (circuit_depth - 1)
end_time = time_module.perf_counter()
time_taken = end_time - start_time
datapoint = (num_rows, num_cols, average_circuit_depth_overhead,
time_taken)
print('Completed run:', datapoint)
return datapoint
def perform_run(nrows, ncols, training_episodes):
test_episodes = 100
circuit_generation_function = lambda: generate_full_layer_circuit(
nrows * ncols).to_dqn_rep()
environment = GridEnvironment(nrows, ncols, circuit_generation_function())
agent = DQNAgent(environment)
start_time = time_module.clock()
train_model(environment,
agent,
training_episodes=training_episodes,
circuit_generation_function=circuit_generation_function,
should_print=False)
average_circuit_depth_overhead = 0.0
for e in range(test_episodes):
actions, circuit_depth = schedule_swaps(
environment,
agent,
circuit=circuit_generation_function(),
experience_db=None)
average_circuit_depth_overhead += (1.0 /
test_episodes) * (circuit_depth - 1)
end_time = time_module.clock()
total_time = end_time - start_time
datapoint = (nrows, ncols, average_circuit_depth_overhead, total_time)
print('Completed run:', datapoint)
return datapoint
print(circuit)
environment = GridEnvironment(nrows,ncols,circuit)
agent = DQNAgent(environment)
db = ExperienceDB()
training_episodes = 200 if nrows*ncols >= 36 else 100
test_episodes = 100
if __name__ == "__main__":
if use_saved_model:
agent.load_model(experiment_name)
else:
train_model(environment, agent, training_episodes=training_episodes, circuit_generation_function=circuit_generation_function, should_print=True)
agent.save_model(experiment_name)
average_test_time = 0.0
average_circuit_depth_overhead = 0.0
for e in range(test_episodes):
if circuit_generation_function is not None:
circuit = circuit_generation_function()
else:
circuit = None
actions, circuit_depth = schedule_swaps(environment, agent, circuit=circuit, experience_db=db)
average_test_time += (1.0/test_episodes) * len(actions)
average_circuit_depth_overhead += (1.0/test_episodes) * (circuit_depth - 1)
