keras.backend.clear_session()
agent = DQNAgent(INPUT_SHAPE, training=True,
replay_memory=minibatch, copy_target_model=False
)
agent.target_model = load_model(target_model_path)
#NOTE: train newly loaded model on new data
if len(minibatch) > 0:
minibatch_history = agent.train_only(len(minibatch), len(minibatch))
if minibatch_history == None:
debug_log('ERROR: Unable to train on iteration\'s data')
replay_memory.extend(minibatch)
else:
debug_log('WARNING: Skipping minibatch training since no new data was found')
#NOTE: train newly loaded model on random selection of old data
agent.replay_memory = replay_memory
sum_loss = 0
if len(replay_memory) > MIN_REPLAY_MEMORY_SIZE:
train_loops = 50
for train_iteration in range(train_loops):
history = agent.train_only(MINIBATCH_SIZE,
MIN_REPLAY_MEMORY_SIZE
)
sum_loss += history.history.get('loss', [0])[0]
average_loss = sum_loss / float(train_loops)
else:
history = None
debug_log(f'on iteration {iteration}, replay_memory has size {len(replay_memory)}')
#NOTE: decay epsilon
