def _cae_ocsvm_experiment(dataset_load_fn, dataset_name, single_class_ind,
gpu_q):
# gpu_to_use = gpu_q.get()
# os.environ["CUDA_VISIBLE_DEVICES"] = gpu_to_use
(x_train, y_train), (x_test, y_test) = dataset_load_fn()
print('data_shape', x_train.shape)
n_channels = x_train.shape[get_channels_axis()]
input_side = x_train.shape[2] # channel side will always be at shape[2]
enc = conv_encoder(input_side, n_channels)
dec = conv_decoder(input_side, n_channels)
# print(input_side)
# print(dec.summary())
x_in = Input(shape=x_train.shape[1:])
x_rec = dec(enc(x_in))
cae = Model(x_in, x_rec)
cae.compile('adam', 'mse')
x_train_task = x_train[y_train.flatten() == single_class_ind]
x_test_task = x_test[y_test.flatten(
) == single_class_ind] # This is just for visual monitoring
cae.fit(x=x_train_task,
y=x_train_task,
batch_size=128,
epochs=200,
validation_data=(x_test_task, x_test_task))
x_train_task_rep = enc.predict(x_train_task, batch_size=128)
if dataset_name in LARGE_DATASET_NAMES: # OC-SVM is quadratic on the number of examples, so subsample training set
subsample_inds = np.random.choice(len(x_train_task_rep),
2500,
replace=False)
x_train_task_rep_temp = x_train_task_rep[subsample_inds]
x_test_rep = enc.predict(x_test, batch_size=128)
pg = ParameterGrid({
'nu': np.linspace(0.1, 0.9, num=9),
'gamma': np.logspace(-7, 2, num=10, base=2)
})
results = Parallel(n_jobs=PARALLEL_N_JOBS)(delayed(
_train_ocsvm_and_score)(d, x_train_task_rep_temp, y_test.flatten() ==
single_class_ind, x_test_rep) for d in pg)
best_params, best_auc_score = max(zip(pg, results), key=lambda t: t[-1])
print(best_params)
best_ocsvm = OneClassSVM(**best_params).fit(x_train_task_rep)
scores = best_ocsvm.decision_function(x_test_rep)
labels = y_test.flatten() == single_class_ind
res_file_name = '{}_cae-oc-svm_{}_{}.npz'.format(
dataset_name, get_class_name_from_index(single_class_ind,
dataset_name),
datetime.datetime.now().strftime('%Y-%m-%d-%H%M'))
res_file_path = os.path.join(RESULTS_DIR, dataset_name, res_file_name)
save_roc_pr_curve_data(scores, labels, res_file_path)
def _adgan_experiment(dataset_load_fn, dataset_name, single_class_ind, gpu_q):
gpu_to_use = gpu_q.get()
os.environ["CUDA_VISIBLE_DEVICES"] = gpu_to_use
(x_train, y_train), (x_test, y_test) = dataset_load_fn()
if len(x_test) > 5000:
# subsample x_test due to runtime complexity
chosen_inds = np.random.choice(len(x_test), 5000, replace=False)
x_test = x_test[chosen_inds]
y_test = y_test[chosen_inds]
n_channels = x_train.shape[get_channels_axis()]
input_side = x_train.shape[2] # image side will always be at shape[2]
critic = conv_encoder(input_side,
n_channels,
representation_dim=1,
representation_activation='linear')
noise_size = 256
generator = conv_decoder(input_side,
n_channels=n_channels,
representation_dim=noise_size)
def prior_gen(b_size):
return np.random.normal(size=(b_size, noise_size))
batch_size = 128
epochs = 100
x_train_task = x_train[y_train.flatten() == single_class_ind]
def data_gen(b_size):
chosen_inds = np.random.choice(len(x_train_task),
b_size,
replace=False)
return x_train_task[chosen_inds]
adgan.train_wgan_with_grad_penalty(prior_gen,
generator,
data_gen,
critic,
batch_size,
epochs,
grad_pen_coef=20)
scores = adgan.scores_from_adgan_generator(x_test, prior_gen, generator)
labels = y_test.flatten() == single_class_ind
res_file_name = '{}_adgan_{}_{}.npz'.format(
dataset_name, get_class_name_from_index(single_class_ind,
dataset_name),
datetime.now().strftime('%Y-%m-%d-%H%M'))
res_file_path = os.path.join(RESULTS_DIR, dataset_name, res_file_name)
save_roc_pr_curve_data(scores, labels, res_file_path)
gpu_q.put(gpu_to_use)
def _dagmm_experiment(dataset_load_fn, dataset_name, single_class_ind, gpu_q):
gpu_to_use = gpu_q.get()
os.environ["CUDA_VISIBLE_DEVICES"] = gpu_to_use
(x_train, y_train), (x_test, y_test) = dataset_load_fn()
n_channels = x_train.shape[get_channels_axis()]
input_side = x_train.shape[2] # image side will always be at shape[2]
enc = conv_encoder(input_side, n_channels, representation_dim=5,
representation_activation='linear')
dec = conv_decoder(input_side, n_channels=n_channels, representation_dim=enc.output_shape[-1])
n_components = 3
estimation = Sequential([Dense(64, activation='tanh', input_dim=enc.output_shape[-1] + 2), Dropout(0.5),
Dense(10, activation='tanh'), Dropout(0.5),
Dense(n_components, activation='softmax')]
)
batch_size = 256
epochs = 200
lambda_diag = 0.0005
lambda_energy = 0.01
dagmm_mdl = dagmm.create_dagmm_model(enc, dec, estimation, lambda_diag)
dagmm_mdl.compile('adam', ['mse', lambda y_true, y_pred: lambda_energy*y_pred])
x_train_task = x_train[y_train.flatten() == single_class_ind]
x_test_task = x_test[y_test.flatten() == single_class_ind] # This is just for visual monitoring
dagmm_mdl.fit(x=x_train_task, y=[x_train_task, np.zeros((len(x_train_task), 1))], # second y is dummy
batch_size=batch_size,
epochs=epochs,
validation_data=(x_test_task, [x_test_task, np.zeros((len(x_test_task), 1))]),
# verbose=0
)
energy_mdl = Model(dagmm_mdl.input, dagmm_mdl.output[-1])
scores = -energy_mdl.predict(x_test, batch_size)
scores = scores.flatten()
if not np.all(np.isfinite(scores)):
min_finite = np.min(scores[np.isfinite(scores)])
scores[~np.isfinite(scores)] = min_finite - 1
labels = y_test.flatten() == single_class_ind
res_file_name = '{}_dagmm_{}_{}.npz'.format(dataset_name,
get_class_name_from_index(single_class_ind, dataset_name),
datetime.now().strftime('%Y-%m-%d-%H%M'))
res_file_path = os.path.join(RESULTS_DIR, dataset_name, res_file_name)
save_roc_pr_curve_data(scores, labels, res_file_path)
gpu_q.put(gpu_to_use)