def train(hparams):
# init exp and track all the parameters from the HyperOptArgumentParser
exp = Experiment(name='dense_model', save_dir='/some/path', autosave=False)
exp.add_argparse_meta(hparams)
# define tensorflow graph
x = tf.placeholder(dtype=tf.int32, name='x')
y = tf.placeholder(dtype=tf.int32, name='y')
out = x * y
sess = tf.Session()
# Run the tf op
for train_step in range(0, 100):
output = sess.run(out, feed_dict={x: hparams.x_val, y: hparams.y_val})
exp.add_metric_row({'fake_err': output})
# save exp when we're done
exp.save()
def main_trainer(hparams):
print_params(hparams)
exp = Experiment(name=hparams.tt_name,
debug=hparams.debug,
autosave=False,
description=hparams.tt_description,
save_dir=hparams.tt_save_path)
exp.add_argparse_meta(hparams)
# fit model
val_scores = []
best_score = 0
for trial_nb in range(hparams.nb_trials):
data = dataset_loader.IndividualSequencesData(
hparams.data_path, y_labels=hparams.y_labels.split(','))
X, Y, lengths = flatten_data(data.train_x_y)
# fit
model = hmm.GaussianHMM(n_components=hparams.nb_components,
n_iter=hparams.nb_hmm_iters)
model.fit(X, lengths)
val_X, val_Y, lengths = flatten_data(data.val_x_y)
Y_hat = model.predict(val_X, lengths)
val_score = np.equal(Y_hat, val_Y).sum() / float(len(Y_hat))
# save model
if val_score > best_score:
best_score = val_score
save_model(model, hparams, exp, trial_nb)
val_scores.append(val_score)
exp.add_metric_row({'val_acc': val_score, 'trail_nb': trial_nb})
mean_val_acc = np.mean(val_scores)
exp.add_metric_row({'final_val_acc': mean_val_acc})
exp.save()
def main_trainer(hparams):
print_params(hparams)
exp = Experiment(name=hparams.tt_name,
debug=hparams.debug,
autosave=False,
description=hparams.tt_description,
save_dir=hparams.tt_save_path)
exp.add_argparse_meta(hparams)
# init data loader
# fit model
val_scores, train_scores = [], []
best_score = 0
for trial_nb in range(hparams.nb_trials):
data = SequentialReadingsData(window_size=hparams.time_steps, data_path=hparams.data_path, flatten_x=True)
clf = RandomForestClassifier(n_estimators=hparams.nb_estimators)
clf.fit(data.train_x, data.train_y)
train_score = clf.score(data.train_x, data.train_y)
val_score = clf.score(data.val_x, data.val_y)
# save model when we have a better one
if val_score > best_score:
best_score = val_score
save_model(clf, hparams, exp, trial_nb)
train_scores.append(train_score)
val_scores.append(val_score)
exp.add_metric_row({'val_acc': val_score, 'train_acc': train_score, 'trail_nb': trial_nb})
mean_val_acc = np.mean(val_scores)
mean_train_acc = np.mean(train_scores)
exp.add_metric_row({'final_val_acc': mean_val_acc, 'final_train_acc': mean_train_acc})
exp.save()
def main_trainer(hparams):
print_params(hparams)
full_exp = Experiment(name=hparams.tt_name + '_overall',
debug=hparams.debug,
autosave=False,
description=hparams.tt_description,
save_dir=hparams.tt_save_path)
full_exp.add_argparse_meta(hparams)
# fit model
val_scores, train_scores = [], []
best_acc = 0
best_loss = 0
best_trial_nb = 0
for trial_nb in range(hparams.nb_trials):
exp = Experiment(name=hparams.tt_name,
debug=hparams.debug,
autosave=False,
description=hparams.tt_description,
save_dir=hparams.tt_save_path)
exp.add_argparse_meta(hparams)
data = SequentialReadingsData(window_size=hparams.time_steps,
data_path=hparams.data_path)
val_loss, val_acc, history = fit_feedforward(hparams, exp, trial_nb,
data)
log_history(history.history, exp)
exp.add_metric_row({
'final_val_acc': val_acc,
'final_train_acc': val_loss
})
exp.save()
full_exp.add_metric_row({
'val_acc': val_acc,
'val_loss': val_loss,
'trial_nb': trial_nb
})
# save model when we have a better one
if val_acc > best_acc:
best_acc = val_acc
best_loss = val_loss
best_trial_nb = trial_nb
val_scores.append(val_acc)
mean_val_acc = np.mean(val_scores)
full_exp.add_metric_row({
'final_val_acc': mean_val_acc,
'best_val_loss': best_loss,
'best_val_acc': best_acc,
'best_trial_nb': best_trial_nb
})
full_exp.save()
parser.add_opt_argument_range(
"--noise_factor", default=1, type=float, start=0, end=1, nb_samples=5
)
hparams = parser.parse_args()
env = gym.make("Pendulum-v0")
state_size = int(np.prod(env.observation_space.shape))
action_size = int(np.prod(env.action_space.shape))
S, A = state_size, action_size
H = hparams.hidden_size
exp.add_metric_row({"S": state_size, "A": action_size, "H": hparams.hidden_size})
class Critic(nn.Module):
def __init__(self):
super().__init__()
self.fc1 = Linear(S + A, H)
self.fc2 = Linear(H, H)
self.out = Linear(H, 1)
def forward(self, s, a):
q = torch.cat((s, a), dim=1)
q = relu(self.fc1(q))
q = relu(self.fc2(q))
q = self.out(q)
return q
