def make_smoothing(X_train, n_train, args):
X_s = None
X_s_path = os.path.join(args.data_path, 'X_s.npy')
do_smoothing = True
if os.path.isfile(X_s_path):
print "\nLoading smoothed data ..."
X_s = np.load(X_s_path)
print "Checking augmented data ..."
if len(X_s) == n_train:
do_smoothing = False
if do_smoothing:
print "\nSmoothing data ..."
X_m = X_train.mean(axis=0)
X_train -= X_m
with Stopwatch(verbose=True) as s:
[U, s, Vh] = svd(X_train,
full_matrices=False,
compute_uv=True,
overwrite_a=True,
check_finite=False)
s[-1000:] = 0.
X_s = U.dot(np.diag(s).dot(Vh))
X_s += X_m
# save to disk
np.save(X_s_path, X_s)
print "\n"
return X_s
def make_mlp(xxx_todo_changeme, xxx_todo_changeme1, xxx_todo_changeme2,
xxx_todo_changeme3, args):
(X_train, y_train) = xxx_todo_changeme
(X_val, y_val) = xxx_todo_changeme1
(X_test, y_test) = xxx_todo_changeme2
(W, hb) = xxx_todo_changeme3
dense_params = {}
if W is not None and hb is not None:
dense_params['weights'] = (W, hb)
# define and initialize MLP model
mlp = Sequential([
Dense(args.n_hidden,
input_shape=(784, ),
kernel_regularizer=regularizers.l2(args.mlp_l2),
kernel_initializer=glorot_uniform(seed=1111),
**dense_params),
Activation('sigmoid'),
Dense(10, kernel_initializer=glorot_uniform(seed=2222)),
Activation('softmax'),
])
mlp.compile(optimizer=MultiAdam(lr=0.001,
lr_multipliers={
'dense_1': args.mlp_lrm[0],
'dense_2': args.mlp_lrm[1]
}),
loss='categorical_crossentropy',
metrics=['accuracy'])
# train and evaluate classifier
with Stopwatch(verbose=True) as s:
early_stopping = EarlyStopping(monitor=args.mlp_val_metric,
patience=12,
verbose=2)
reduce_lr = ReduceLROnPlateau(monitor=args.mlp_val_metric,
factor=0.2,
verbose=2,
patience=6,
min_lr=1e-5)
callbacks = [early_stopping, reduce_lr]
try:
mlp.fit(X_train,
one_hot(y_train, n_classes=10),
epochs=args.mlp_epochs,
batch_size=args.mlp_batch_size,
shuffle=False,
validation_data=(X_val, one_hot(y_val, n_classes=10)),
callbacks=callbacks)
except KeyboardInterrupt:
pass
y_pred = mlp.predict(X_test)
y_pred = unhot(one_hot_decision_function(y_pred), n_classes=10)
print("Test accuracy: {:.4f}".format(accuracy_score(y_test, y_pred)))
# save predictions, targets, and fine-tuned weights
np.save(args.mlp_save_prefix + 'y_pred.npy', y_pred)
np.save(args.mlp_save_prefix + 'y_test.npy', y_test)
W_finetuned, _ = mlp.layers[0].get_weights()
np.save(args.mlp_save_prefix + 'W_finetuned.npy', W_finetuned)
def make_augmentation(X_train, y_train, n_train, args):
X_aug = None
X_aug_path = os.path.join(args.data_path, 'X_aug.npy')
y_train = y_train.tolist() * 10
RNG(seed=1337).shuffle(y_train)
augment = True
if os.path.isfile(X_aug_path):
print "\nLoading augmented data ..."
X_aug = np.load(X_aug_path)
print "Checking augmented data ..."
if len(X_aug) == 10 * n_train:
augment = False
if augment:
print "\nAugmenting data ..."
s = Stopwatch(verbose=True).start()
X_aug = np.zeros((10 * n_train, 32, 32, 3), dtype=np.float32)
X_train = im_unflatten(X_train)
X_aug[:n_train] = X_train
for i in xrange(n_train):
for k, offset in enumerate(((1, 0), (-1, 0), (0, 1), (0, -1))):
img = X_train[i].copy()
X_aug[(k + 1) * n_train + i] = shift(img, offset=offset)
for i in xrange(5 * n_train):
X_aug[5 * n_train + i] = horizontal_mirror(X_aug[i].copy())
# shuffle once again
RNG(seed=1337).shuffle(X_aug)
# convert to 'uint8' type to save disk space
X_aug *= 255.
X_aug = X_aug.astype('uint8')
# flatten to (10 * `n_train`, 3072) shape
X_aug = im_flatten(X_aug)
# save to disk
np.save(X_aug_path, X_aug)
s.elapsed()
print "\n"
return X_aug, y_train
BN(),
Activation('relu'),
Dropout(args.mlp_dropout, seed=4444),
Dense(10, kernel_initializer=glorot_uniform(seed=5555)),
Activation('softmax'),
])
mlp.compile(optimizer=MultiAdam(lr=0.001,
lr_multipliers={
'dense_1': args.mlp_lrm[0],
'dense_2': args.mlp_lrm[1]
}),
loss='categorical_crossentropy',
metrics=['accuracy'])
# train and evaluate classifier
with Stopwatch(verbose=True) as s:
early_stopping = EarlyStopping(monitor=args.mlp_val_metric,
patience=12,
verbose=2)
reduce_lr = ReduceLROnPlateau(monitor=args.mlp_val_metric,
factor=0.2,
verbose=2,
patience=6,
min_lr=1e-5)
callbacks = [early_stopping, reduce_lr]
try:
mlp.fit(X_train,
one_hot(y_train, n_classes=10),
epochs=args.mlp_epochs,
batch_size=args.mlp_batch_size,
shuffle=False,