def evaluate_parameters(optimizer_type, learning_rate, k_way_train_type,
n_convs, dropout, early_stop_patience):
print(optimizer_type, learning_rate, k_way_train_type, n_convs,
dropout, early_stop_patience)
if optimizer_type == 'adam':
optimizer = tf.keras.optimizers.Adam(lr=learning_rate)
elif optimizer_type == 'rmsprop':
optimizer = tf.keras.optimizers.RMSprop(lr=learning_rate)
else:
raise ValueError('Unsupported optimizer_type')
if k_way_train_type == 'large':
cur_k_train = k_way_train
elif k_way_train_type == 'same':
cur_k_train = k_way_test
else:
raise ValueError('Unsupported k value')
img_fn = augmented_img_pipeline_fn
result = evaluate_fashion_few_shot(
train_df=experiment_train_df,
val_df=experiment_val_df,
test_df=val_df,
n_shot=n_shot,
k_way_test=k_way_test,
k_way_train=cur_k_train,
n_queries_train=n_queries_train,
n_queries_test=n_queries_test,
lr=learning_rate,
eps_per_epoch=eps_per_epoch,
n_epochs=n_epochs,
test_eps=eps_per_epoch,
img_shape=img_shape,
embedding_fn=lambda x: build_embedding_model(x, n_convs, dropout),
img_pipeline_fn=img_fn,
opt=optimizer,
patience=early_stop_patience)
result['optimizer'] = optimizer_type
return result['test_loss']
def few_shot_optimize(train_df,
val_df,
test_df,
n_shot,
k_way_test,
k_way_train=config.K_WAY_TRAIN,
n_queries_train=config.N_QUERIES_TRAIN,
n_queries_test=config.N_QUERIES_TEST,
eps_per_epoch=config.EPS_PER_EPOCH,
n_epochs=config.N_EPOCHS,
test_eps=config.TEST_EPS,
img_shape=config.IMG_SHAPE):
experiment_val_classes = set(
np.random.choice(train_df.class_name.unique(), size=16, replace=False))
experiment_train_df = train_df[~train_df.class_name.
isin(experiment_val_classes)]
experiment_val_df = train_df[train_df.class_name.isin(
experiment_val_classes)]
dimensions = [
skopt.space.Categorical(name='optimizer_type',
categories=('adam', 'rmsprop')),
skopt.space.Real(name='learning_rate', low=1e-3, high=3),
skopt.space.Categorical(name='k_way_train_type',
categories=('large', 'same')),
skopt.space.Integer(name='n_convs', low=4, high=6),
skopt.space.Real(name='dropout', low=0.0, high=0.2),
skopt.space.Integer(name='early_stop_patience', low=1, high=5)
]
@skopt.utils.use_named_args(dimensions)
def evaluate_parameters(optimizer_type, learning_rate, k_way_train_type,
n_convs, dropout, early_stop_patience):
print(optimizer_type, learning_rate, k_way_train_type, n_convs,
dropout, early_stop_patience)
if optimizer_type == 'adam':
optimizer = tf.keras.optimizers.Adam(lr=learning_rate)
elif optimizer_type == 'rmsprop':
optimizer = tf.keras.optimizers.RMSprop(lr=learning_rate)
else:
raise ValueError('Unsupported optimizer_type')
if k_way_train_type == 'large':
cur_k_train = k_way_train
elif k_way_train_type == 'same':
cur_k_train = k_way_test
else:
raise ValueError('Unsupported k value')
img_fn = augmented_img_pipeline_fn
result = evaluate_fashion_few_shot(
train_df=experiment_train_df,
val_df=experiment_val_df,
test_df=val_df,
n_shot=n_shot,
k_way_test=k_way_test,
k_way_train=cur_k_train,
n_queries_train=n_queries_train,
n_queries_test=n_queries_test,
lr=learning_rate,
eps_per_epoch=eps_per_epoch,
n_epochs=n_epochs,
test_eps=eps_per_epoch,
img_shape=img_shape,
embedding_fn=lambda x: build_embedding_model(x, n_convs, dropout),
img_pipeline_fn=img_fn,
opt=optimizer,
patience=early_stop_patience)
result['optimizer'] = optimizer_type
return result['test_loss']
res = skopt.gp_minimize(evaluate_parameters,
dimensions,
n_calls=10,
n_random_starts=5)
best_opt, best_lr, best_k_way_type, best_convs, best_dropout, best_patience = res.x
print(res.x)
if best_opt == 'adam':
opt = tf.keras.optimizers.Adam(lr=best_lr)
elif best_opt == 'rmsprop':
opt = tf.keras.optimizers.RMSprop(lr=best_lr)
else:
raise ValueError('oops')
result = evaluate_fashion_few_shot(
train_df=train_df,
val_df=val_df,
test_df=test_df,
n_shot=n_shots,
k_way_test=k_way_test,
k_way_train=k_way_train if best_k_way_type == 'large' else k_way_test,
n_queries_train=n_queries_train,
n_queries_test=n_queries_test,
lr=best_lr,
eps_per_epoch=eps_per_epoch,
n_epochs=n_epochs,
test_eps=test_eps,
img_shape=img_shape,
img_pipeline_fn=augmented_img_pipeline_fn,
opt=opt,
patience=best_patience)
result['opt'] = best_opt
return result
from few_shot.experiments.fashion import evaluate_fashion_few_shot
from few_shot.dataset.fashion import fashion_dfs
from few_shot.dataset.image_pipeline import class_augmentation_fn
if __name__ == '__main__':
np.random.seed(23)
tf.random.set_random_seed(29)
train_df, val_df, test_df = fashion_dfs()
results = []
for n_shots, k_way_test in itertools.product(config.SHOTS,
config.TEST_K_WAY):
print(f'Running fashion experiment {n_shots}-shot, {k_way_test} way')
assert k_way_test <= test_df.class_name.nunique()
result = evaluate_fashion_few_shot(
train_df=train_df,
val_df=val_df,
test_df=test_df,
n_shot=n_shots,
k_way_test=k_way_test,
augment=True,
post_processing_fn=class_augmentation_fn(config.K_WAY_TRAIN),
patience=30)
results.append(result)
df = pd.DataFrame.from_records(results)
print(df)
df.to_csv('fashion_new_augmentation_results.csv')
import numpy as np
import pandas as pd
import tensorflow as tf
import few_shot.experiments.fashion.config as config
from few_shot.experiments.fashion import evaluate_fashion_few_shot
from few_shot.dataset.fashion import fashion_dfs
if __name__ == '__main__':
np.random.seed(23)
tf.random.set_random_seed(29)
train_df, val_df, test_df = fashion_dfs()
results = []
for n_shots, k_way_test in itertools.product(config.SHOTS,
config.TEST_K_WAY):
print(f'Running fashion experiment {n_shots}-shot, {k_way_test} way')
assert k_way_test <= test_df.class_name.nunique()
result = evaluate_fashion_few_shot(train_df=train_df,
val_df=val_df,
test_df=test_df,
n_shot=n_shots,
k_way_test=k_way_test,
patience=20)
results.append(result)
df = pd.DataFrame.from_records(results)
print(df)
df.to_csv('fashion_default_params_no_early_stop_results.csv')
from few_shot.experiments.fashion import evaluate_fashion_few_shot
from few_shot.dataset.fashion import fashion_dfs
from few_shot.dataset.image_pipeline import augmented_img_pipeline_fn
if __name__ == '__main__':
np.random.seed(23)
tf.random.set_random_seed(29)
train_df, val_df, test_df = fashion_dfs()
results = []
for n_shots, k_way_test in itertools.product(config.SHOTS,
config.TEST_K_WAY):
print(f'Running fashion experiment {n_shots}-shot, {k_way_test} way')
assert k_way_test <= test_df.class_name.nunique()
result = evaluate_fashion_few_shot(
train_df=train_df,
val_df=val_df,
test_df=test_df,
n_shot=n_shots,
k_way_test=k_way_test,
img_pipeline_fn=augmented_img_pipeline_fn,
patience=40,
restore_best_weights=True)
results.append(result)
df = pd.DataFrame.from_records(results)
print(df)
df.to_csv('fashion_augmentation_no_early_stopping_results.csv')
from few_shot.model import build_embedding_model
if __name__ == '__main__':
np.random.seed(23)
tf.random.set_random_seed(29)
train_df, val_df, test_df = fashion_dfs()
results = []
for n_shots, k_way_test in itertools.product(config.SHOTS,
config.TEST_K_WAY):
print(f'Running fashion experiment {n_shots}-shot, {k_way_test} way')
assert k_way_test <= test_df.class_name.nunique()
result = evaluate_fashion_few_shot(
train_df=train_df,
val_df=val_df,
test_df=test_df,
n_shot=n_shots,
k_way_test=k_way_test,
embedding_fn=lambda x: build_embedding_model(x, 5, 0.1),
img_pipeline_fn=augmented_img_pipeline_fn,
patience=5,
restore_best_weights=True,
reduce_lr_on_plateau=True)
results.append(result)
df = pd.DataFrame.from_records(results)
print(df)
df.to_csv('fashion_augmentation_only_5_convs_results.csv')
from few_shot.experiments.fashion import config
from few_shot.experiments.fashion import evaluate_fashion_few_shot
from few_shot.dataset.fashion import fashion_dfs
from few_shot.dataset.image_pipeline import augmented_img_pipeline_fn
if __name__ == '__main__':
np.random.seed(23)
tf.random.set_random_seed(29)
train_df, val_df, test_df = fashion_dfs()
results = []
for n_shots, k_way_test in itertools.product(config.SHOTS,
config.TEST_K_WAY):
print(f'Running fashion experiment {n_shots}-shot, {k_way_test} way')
assert k_way_test <= test_df.class_name.nunique()
result = evaluate_fashion_few_shot(
train_df=train_df,
val_df=val_df,
test_df=test_df,
n_shot=n_shots,
k_way_test=k_way_test,
img_pipeline_fn=augmented_img_pipeline_fn)
results.append(result)
df = pd.DataFrame.from_records(results)
print(df)
df.to_csv('fashion_augmentation_only_results.csv')