def just_train_on_dataset_up_to_T(dat, exs, pybml_ho, sess, T):
train_fd, valid_fd = utils.feed_dicts(dat, exs)
# print('train_feed:', train_fd) # DEBUG
sess.run(pybml_ho.outergradient.initialization)
tr_acc, v_acc = [[] for _ in range(T)], [[] for _ in range(T)]
for ex in exs:
# ex.model.initialize(session=sess)
for t in range(T):
sess.run(
ex.optimizers["apply_updates"],
feed_dict={ex.x: train_fd[ex.x], ex.y: train_fd[ex.y]},
)
tr_acc[t].append(
sess.run(
ex.scores["accuracy"],
feed_dict={ex.x: train_fd[ex.x], ex.y: train_fd[ex.y]},
)
)
v_acc[t].append(
sess.run(
ex.scores["accuracy"],
feed_dict={ex.x: valid_fd[ex.x_], ex.y: valid_fd[ex.y_]},
)
)
return tr_acc, v_acc
def just_train_on_dataset(dat, exs, pybml_ho, sess, T):
train_fd, valid_fd = feed_dicts(dat, exs)
# print('train_feed:', train_fd) # DEBUG
sess.run(pybml_ho.outergradient.initialization)
tr_acc, v_acc = [], []
for ex in exs:
[
sess.run(ex.optimizers['apply_updates'],
feed_dict={
ex.x: train_fd[ex.x],
ex.y: train_fd[ex.y]
}) for _ in range(T)
]
tr_acc.append(
sess.run(ex.scores['accuracy'],
feed_dict={
ex.x: train_fd[ex.x],
ex.y: train_fd[ex.y]
}))
v_acc.append(
sess.run(ex.scores['accuracy'],
feed_dict={
ex.x: valid_fd[ex.x_],
ex.y: valid_fd[ex.y_]
}))
return tr_acc, v_acc
def meta_train(
exp_dir,
metasets,
exs,
pybml_ho,
saver,
sess,
n_test_episodes,
MBS,
seed,
resume,
T,
n_meta_iterations,
print_interval,
save_interval,
):
# use workers to fill the batches queues (is it worth it?)
result_path = os.path.join(exp_dir, "results.pickle")
tf.global_variables_initializer().run(session=sess)
n_test_batches = n_test_episodes // MBS
rand = dl.get_rand_state(seed)
results = {
"train_train": {"mean": [], "std": []},
"train_test": {"mean": [], "std": []},
"test_test": {"mean": [], "std": []},
"valid_test": {"mean": [], "std": []},
"outer_losses": {"mean": [], "std": []},
"learning_rate": [],
"iterations": [],
"episodes": [],
"time": [],
"alpha": [],
}
resume_itr = 0
if resume:
model_file = tf.train.latest_checkpoint(exp_dir)
if model_file:
print("Restoring results from " + result_path)
results = load_obj(result_path)
ind1 = model_file.index("model")
resume_itr = int(model_file[ind1 + 5 :]) + 1
print("Restoring model weights from " + model_file)
saver.restore(sess, model_file)
""" Meta-Train """
train_batches = BatchQueueMock(metasets.train, 1, MBS, rand)
valid_batches = BatchQueueMock(metasets.validation, n_test_batches, MBS, rand)
test_batches = BatchQueueMock(metasets.test, n_test_batches, MBS, rand)
start_time = time.time()
print(
"\nIteration quantities: train_train acc, train_test acc, valid_test, acc"
" test_test acc mean(std) over %d episodes" % n_test_episodes
)
with sess.as_default():
inner_losses = []
for meta_it in range(resume_itr, n_meta_iterations):
tr_fd, v_fd = utils.feed_dicts(train_batches.get_all_batches()[0], exs)
pybml_ho.run(tr_fd, v_fd)
duration = time.time() - start_time
results["time"].append(duration)
outer_losses = []
for _, ex in enumerate(exs):
outer_losses.append(
sess.run(
ex.errors["validation"], boml.utils.merge_dicts(tr_fd, v_fd)
)
)
outer_losses_moments = (np.mean(outer_losses), np.std(outer_losses))
results["outer_losses"]["mean"].append(outer_losses_moments[0])
results["outer_losses"]["std"].append(outer_losses_moments[1])
if meta_it % print_interval == 0 or meta_it == n_meta_iterations - 1:
results["iterations"].append(meta_it)
results["episodes"].append(meta_it * MBS)
if "alpha" in pybml_ho.param_dict.keys():
alpha_moment = pybml_ho.param_dict["alpha"].eval()
print("alpha_itr" + str(meta_it) + ": ", alpha_moment)
results["alpha"].append(alpha_moment)
if "s" in pybml_ho.param_dict.keys():
s = sess.run(["s:0"])[0]
print("s: {}".format(s))
if "t" in pybml_ho.param_dict.keys():
t = sess.run(["t:0"])[0]
print("t: {}".format(t))
train_result = accuracy_on(train_batches, exs, pybml_ho, sess, T)
test_result = accuracy_on(test_batches, exs, pybml_ho, sess, T)
valid_result = accuracy_on(valid_batches, exs, pybml_ho, sess, T)
train_train = (np.mean(train_result[0]), np.std(train_result[0]))
train_test = (np.mean(train_result[1]), np.std(train_result[1]))
valid_test = (np.mean(valid_result[1]), np.std(valid_result[1]))
test_test = (np.mean(test_result[1]), np.std(test_result[1]))
results["train_train"]["mean"].append(train_train[0])
results["train_test"]["mean"].append(train_test[0])
results["valid_test"]["mean"].append(valid_test[0])
results["test_test"]["mean"].append(test_test[0])
results["train_train"]["std"].append(train_train[1])
results["train_test"]["std"].append(train_test[1])
results["valid_test"]["std"].append(valid_test[1])
results["test_test"]["std"].append(test_test[1])
results["inner_losses"] = inner_losses
print("mean outer losses: {}".format(outer_losses_moments[0]))
print(
"it %d, ep %d (%.5fs): %.5f, %.5f, %.5f, %.5f"
% (
meta_it,
meta_it * MBS,
duration,
train_train[0],
train_test[0],
valid_test[0],
test_test[0],
)
)
lr = sess.run(["lr:0"])[0]
print("lr: {}".format(lr))
# do_plot(logdir, results)
if meta_it % save_interval == 0 or meta_it == n_meta_iterations - 1:
saver.save(sess, exp_dir + "/model" + str(meta_it))
save_obj(result_path, results)
start_time = time.time()
return results
def meta_train(exp_dir, metasets, exs, pybml_ho, saver, sess, n_test_episodes,
MBS, seed, resume, T, n_meta_iterations, print_interval,
save_interval):
# use workers to fill the batches queues (is it worth it?)
result_path = os.path.join(exp_dir, 'results.pickle')
tf.global_variables_initializer().run(session=sess)
n_test_batches = n_test_episodes // MBS
rand = dl.get_rand_state(seed)
results = {
'train_train': {
'mean': [],
'std': []
},
'train_test': {
'mean': [],
'std': []
},
'test_test': {
'mean': [],
'std': []
},
'valid_test': {
'mean': [],
'std': []
},
'outer_losses': {
'mean': [],
'std': []
},
'learning_rate': [],
'iterations': [],
'episodes': [],
'time': [],
'alpha': []
}
resume_itr = 0
if resume:
model_file = tf.train.latest_checkpoint(exp_dir)
if model_file:
print("Restoring results from " + result_path)
results = load_obj(result_path)
ind1 = model_file.index('model')
resume_itr = int(model_file[ind1 + 5:]) + 1
print("Restoring model weights from " + model_file)
saver.restore(sess, model_file)
''' Meta-Train '''
train_batches = BatchQueueMock(metasets.train, 1, MBS, rand)
valid_batches = BatchQueueMock(metasets.validation, n_test_batches, MBS,
rand)
test_batches = BatchQueueMock(metasets.test, n_test_batches, MBS, rand)
start_time = time.time()
print(
'\nIteration quantities: train_train acc, train_test acc, valid_test, acc'
' test_test acc mean(std) over %d episodes' % n_test_episodes)
with sess.as_default():
inner_losses = []
for meta_it in range(resume_itr, n_meta_iterations):
tr_fd, v_fd = feed_dicts(train_batches.get_all_batches()[0], exs)
pybml_ho.run(tr_fd, v_fd)
duration = time.time() - start_time
results['time'].append(duration)
outer_losses = []
for _, ex in enumerate(exs):
outer_losses.append(
sess.run(ex.errors['validation'],
boml.utils.merge_dicts(tr_fd, v_fd)))
outer_losses_moments = (np.mean(outer_losses),
np.std(outer_losses))
results['outer_losses']['mean'].append(outer_losses_moments[0])
results['outer_losses']['std'].append(outer_losses_moments[1])
if meta_it % print_interval == 0 or meta_it == n_meta_iterations - 1:
results['iterations'].append(meta_it)
results['episodes'].append(meta_it * MBS)
if 'alpha' in pybml_ho.param_dict.keys():
alpha_moment = pybml_ho.param_dict['alpha'].eval()
print('alpha_itr' + str(meta_it) + ': ', alpha_moment)
results['alpha'].append(alpha_moment)
if 's' in pybml_ho.param_dict.keys():
s = sess.run(["s:0"])[0]
print('s: {}'.format(s))
if 't' in pybml_ho.param_dict.keys():
t = sess.run(["t:0"])[0]
print('t: {}'.format(t))
train_result = accuracy_on(train_batches, exs, pybml_ho, sess,
T)
test_result = accuracy_on(test_batches, exs, pybml_ho, sess, T)
valid_result = accuracy_on(valid_batches, exs, pybml_ho, sess,
T)
train_train = (np.mean(train_result[0]),
np.std(train_result[0]))
train_test = (np.mean(train_result[1]),
np.std(train_result[1]))
valid_test = (np.mean(valid_result[1]),
np.std(valid_result[1]))
test_test = (np.mean(test_result[1]), np.std(test_result[1]))
results['train_train']['mean'].append(train_train[0])
results['train_test']['mean'].append(train_test[0])
results['valid_test']['mean'].append(valid_test[0])
results['test_test']['mean'].append(test_test[0])
results['train_train']['std'].append(train_train[1])
results['train_test']['std'].append(train_test[1])
results['valid_test']['std'].append(valid_test[1])
results['test_test']['std'].append(test_test[1])
results['inner_losses'] = inner_losses
print('mean outer losses: {}'.format(outer_losses_moments[0]))
print('it %d, ep %d (%.5fs): %.5f, %.5f, %.5f, %.5f' %
(meta_it, meta_it * MBS, duration, train_train[0],
train_test[0], valid_test[0], test_test[0]))
lr = sess.run(["lr:0"])[0]
print('lr: {}'.format(lr))
# do_plot(logdir, results)
if meta_it % save_interval == 0 or meta_it == n_meta_iterations - 1:
saver.save(sess, exp_dir + '/model' + str(meta_it))
save_obj(result_path, results)
start_time = time.time()
return results