replay_initial = 10000
learning_rate = 1e-5
train_replay_buffer = ReplayBuffer(100000)
analysis_replay_buffer = ReplayBuffer(100000)
policy_model = QLearner(env, train_num_frames, batch_size, gamma,
train_replay_buffer)
target_model = QLearner(env, train_num_frames, batch_size, gamma,
train_replay_buffer)
target_model.load_state_dict(policy_model.state_dict())
target_model.eval()
optimizer = optim.Adam(policy_model.parameters(), lr=learning_rate)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if USE_CUDA:
policy_model = policy_model.to(device)
target_model = target_model.to(device)
epsilon_by_frame = lambda frame_idx: epsilon_final + (
epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
def play_to_train(num_frames, policy_model, target_model, buffer):
losses = []
all_rewards = []
mean_losses = []
mean_rewards = []
episode_reward = 0
state = env.reset()
start_training = time.time()
batch_size = 32
gamma = 0.99
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
"""loading the saved model"""
device = torch.device("cuda")
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
filename = 'newdqnModel.pt'
model.load_state_dict(torch.load(filename))
model.to(device)
model.eval()
"""choosing 1000 frames randomly!"""
frame_range = 50000
frame_list = set(random.sample(range(1, frame_range), 1000))
vis_feature_matrix = []
vis_rewards = []
vis_actions = []
state = env.reset()
indx = 0
episode_reward = 0
for frame_idx in range(0,frame_range):
#print(frame_idx)
action= model.act(state, 0)