def test_train(tmp_path):
"""Tests imbDRL.agents.ddqn.TrainDDQN.train."""
tmp_models = str(
tmp_path / "test_model"
) # No support for pathLib https://github.com/tensorflow/tensorflow/issues/37357
tmp_logs = str(tmp_path / "test_log")
model = TrainDDQN(10,
10,
0.001,
0.0,
0.1,
5,
model_path=tmp_models,
log_dir=tmp_logs,
val_every=2,
memory_length=30)
with pytest.raises(Exception) as exc:
model.train()
assert "must be compiled" in str(exc.value)
model.compile_model(
np.random.rand(4, 10).astype(np.float32),
np.random.choice(2, size=4).astype(np.int32),
[Dense(4), Dense(2)])
model.train(
np.random.rand(4, 10).astype(np.float32),
np.random.choice(2, size=4).astype(np.int32))
assert model.replay_buffer.num_frames().numpy(
) >= 10 + 10 # 10 for warmup + 1 for each episode
assert model.global_episode == 10
model = TrainDDQN(10,
10,
0.001,
0.0,
0.1,
5,
model_path=tmp_models,
log_dir=tmp_logs,
val_every=2)
with pytest.raises(Exception) as exc:
model.train()
assert "must be compiled" in str(exc.value)
model.compile_model(
np.random.rand(4, 10).astype(np.float32),
np.random.choice(2, size=4).astype(np.int32),
[Dense(4), Dense(2)])
model.train(
np.random.rand(4, 10).astype(np.float32),
np.random.choice(2, size=4).astype(np.int32))
assert model.replay_buffer.num_frames(
) >= 10 # 10 in total since no memory length is defined
assert model.global_episode == 10
) # Normalization should happen after splitting train and test sets
X_train, X_test, y_train, y_test = train_test_split(df.to_numpy(),
y,
stratify=y,
test_size=0.2)
X_train, y_train, X_test, y_test, X_val, y_val = get_train_test_val(
X_train, y_train, X_test, y_test, min_class, maj_class, val_frac=0.2)
model = TrainDDQN(episodes,
warmup_steps,
learning_rate,
gamma,
min_epsilon,
decay_episodes,
target_update_period=target_update_period,
target_update_tau=target_update_tau,
batch_size=batch_size,
collect_steps_per_episode=collect_steps_per_episode,
memory_length=memory_length,
collect_every=collect_every,
n_step_update=n_step_update)
model.compile_model(X_train, y_train, layers)
model.q_net.summary()
model.train(X_val, y_val, "F1")
stats = model.evaluate(X_test, y_test, X_train, y_train)
print(rounded_dict(stats))
# {'Gmean': 0.824172, 'F1': 0.781395, 'Precision': 0.730435, 'Recall': 0.84, 'TP': 84, 'TN': 131, 'FP': 31, 'FN': 16}
X_train,
y_train,
X_test,
y_test,
min_class,
maj_class,
val_frac=0.1,
imb_ratio=imb_ratio,
imb_test=False)
model = TrainDDQN(episodes,
warmup_steps,
learning_rate,
gamma,
min_epsilon,
decay_episodes,
target_update_period=target_update_period,
target_update_tau=target_update_tau,
batch_size=batch_size,
collect_steps_per_episode=collect_steps_per_episode,
memory_length=memory_length,
collect_every=collect_every,
n_step_update=n_step_update)
model.compile_model(X_train, y_train, layers, imb_ratio=imb_ratio)
model.train(X_val, y_val, "Gmean")
stats = model.evaluate(X_test, y_test, X_train, y_train)
print(rounded_dict(stats))
# {'Gmean': 0.987032, 'F1': 0.955702, 'Precision': 0.930275, 'Recall': 0.982558, 'TP': 1014, 'TN': 8892, 'FP': 76, 'FN': 18}
target_update_period = 10_000
target_update_tau = 1
n_step_update = 4
layers = [Dense(250, activation="relu"),
Dense(2, activation=None)]
learning_rate = 0.00025 # Learning rate
gamma = 0.1 # Discount factor
min_epsilon = 0.01 # Minimal and final chance of choosing random action
decay_episodes = 100_000 # Number of episodes to decay from 1.0 to `min_epsilon`
imb_ratio = 0.1 # Imbalance rate
min_class = [0] # Minority classes
maj_class = [1] # Majority classes
X_train, y_train, X_test, y_test = load_imdb(config=(5_000, 500))
X_train, y_train, X_test, y_test, X_val, y_val = get_train_test_val(X_train, y_train, X_test, y_test,
min_class, maj_class, val_frac=0.1, imb_ratio=imb_ratio, imb_test=False)
model = TrainDDQN(episodes, warmup_steps, learning_rate, gamma, min_epsilon, decay_episodes, target_update_period=target_update_period,
target_update_tau=target_update_tau, batch_size=batch_size, collect_steps_per_episode=collect_steps_per_episode,
memory_length=memory_length, collect_every=collect_every, n_step_update=n_step_update)
model.compile_model(X_train, y_train, layers, imb_ratio=imb_ratio)
model.train(X_test, y_test, "Gmean")
stats = model.evaluate(X_test, y_test, X_train, y_train)
print(rounded_dict(stats))
# {'Gmean': 0.286451, 'F1': 0.152846, 'Precision': 0.4967, 'Recall': 0.09032, 'TP': 1129, 'TN': 11356, 'FP': 1144, 'FN': 11371}