def main(_):
n_concept = 5
n_cluster = 5
n = 60000
n0 = int(n * 0.8)
pretrain = True
# Loads data.
x, y, concept = toy_helper.load_xyconcept(n, pretrain)
if not pretrain:
x_train = x[:n0, :]
x_val = x[n0:, :]
y_train = y[:n0, :]
y_val = y[n0:, :]
all_feature_dense = np.load('all_feature_dense.npy')
f_train = all_feature_dense[:n0, :]
f_val = all_feature_dense[n0:, :]
# Loads model.
if not pretrain:
dense2, predict, _ = toy_helper.load_model(x_train,
y_train,
x_val,
y_val,
pretrain=pretrain)
else:
dense2, predict, _ = toy_helper.load_model(_,
_,
_,
_,
pretrain=pretrain)
# Loads concepts.
concept_arraynew = np.load('concept_arraynew.npy')
concept_arraynew2 = np.load('concept_arraynew2.npy')
for n_concept in range(1, 10):
print(n_concept)
# Discovers concept with true cluster.
finetuned_model_pr = ipca.ipca_model(concept_arraynew2,
dense2,
predict,
f_train,
y_train,
f_val,
y_val,
n_concept,
comp1=True)
num_epoch = 5
for _ in range(num_epoch):
finetuned_model_pr.fit(f_train,
y_train,
batch_size=50,
epochs=10,
verbose=True,
validation_data=(f_val, y_val))
# Evaluates groupacc and get concept_matrix.
_, _ = ipca.get_groupacc(finetuned_model_pr,
concept_arraynew2,
f_train,
f_val,
concept,
n_concept,
n_cluster,
n0,
verbose=False)
# Discovers concepts with self-discovered clusters.
finetuned_model_pr = ipca.ipca_model(concept_arraynew,
dense2,
predict,
f_train,
y_train,
f_val,
y_val,
n_concept,
comp1=True)
num_epoch = 5
for _ in range(num_epoch):
finetuned_model_pr.fit(f_train,
y_train,
batch_size=50,
epochs=10,
verbose=True,
validation_data=(f_val, y_val))
_, _ = ipca.get_groupacc(finetuned_model_pr,
concept_arraynew,
f_train,
f_val,
concept,
n_concept,
n_cluster,
n0,
verbose=False)
for n_concept in range(1, 10):
print(n_concept)
concept_matrix_ace = toy_helper.get_ace_concept(
concept_arraynew, dense2, predict, f_val, n_concept)
finetuned_model_pr_ace = ipca.ipca_model(concept_arraynew,
dense2,
predict,
f_train,
y_train,
f_val,
y_val,
n_concept,
verbose=True,
epochs=0,
metric='accuracy')
finetuned_model_pr_ace.layers[-5].set_weights([concept_matrix_ace])
print(finetuned_model_pr_ace.evaluate(f_val, y_val))
_, _ = ipca.get_groupacc(finetuned_model_pr_ace,
concept_arraynew,
f_train,
f_val,
concept,
n_concept,
n_cluster,
n0,
verbose=False)
concept_matrix_ace2 = toy_helper.get_ace_concept(
concept_arraynew2, dense2, predict, f_val, n_concept)
finetuned_model_pr_ace2 = ipca.ipca_model(concept_arraynew2,
dense2,
predict,
f_train,
y_train,
f_val,
y_val,
n_concept,
verbose=True,
epochs=0,
metric='accuracy')
finetuned_model_pr_ace2.layers[-5].set_weights([concept_matrix_ace2])
print(finetuned_model_pr_ace2.evaluate(f_val, y_val))
_, _ = ipca.get_groupacc(finetuned_model_pr_ace2,
concept_arraynew2,
f_train,
f_val,
concept,
n_concept,
n_cluster,
n0,
verbose=False)
concept_matrix_pca = toy_helper.get_pca_concept(f_train, n_concept)
finetuned_model_pr_pca = ipca.ipca_model(concept_arraynew,
dense2,
predict,
f_train,
y_train,
f_val,
y_val,
n_concept,
verbose=True,
epochs=0,
metric='accuracy')
finetuned_model_pr_pca.layers[-5].set_weights([concept_matrix_pca])
print(finetuned_model_pr_pca.evaluate(f_val, y_val))
_, _ = ipca.get_groupacc(finetuned_model_pr_pca,
concept_arraynew,
f_train,
f_val,
concept,
n_concept,
n_cluster,
n0,
verbose=False)
y_train_logit, y_val_logit, y_train, \
y_val, f_train, f_val, dense2, predict = awa_helper.load_data(TRAIN_DIR,
SIZE,
BATCH_SIZE,
pretrained,
noise=0.0)
concept_arraynew, concept_arraynew_active, \
concept_list, active_list = awa_helper.load_conceptarray()
finetuned_model_pr = ipca.ipca_model(concept_arraynew_active,
dense2,
predict,
f_train,
y_train_logit,
f_val,
y_val_logit,
n_concept,
verbose=True,
epochs=150,
metric='accuracy')
num_epoch = 50
for _ in range(num_epoch):
finetuned_model_pr.fit(f_train,
y_train_logit,
batch_size=100,
epochs=10,
verbose=1,
validation_data=(f_val, y_val_logit))
def main(_):
n_concept = 5
n_cluster = 5
n = 60000
n0 = int(n * 0.8)
pretrain = True
# Loads data.
x, y, concept = toy_helper.load_xyconcept(n, pretrain)
if not pretrain:
x_train = x[:n0, :]
x_val = x[n0:, :]
y_train = y[:n0, :]
y_val = y[n0:, :]
all_feature_dense = np.load('all_feature_dense.npy')
f_train = all_feature_dense[:n0, :]
f_val = all_feature_dense[n0:, :]
# Loads model
if not pretrain:
dense2, predict, _ = toy_helper.load_model(
x_train, y_train, x_val, y_val, pretrain=pretrain)
else:
dense2, predict, _ = toy_helper.load_model(_, _, _, _, pretrain=pretrain)
# Loads concept
concept_arraynew = np.load('concept_arraynew.npy')
concept_arraynew2 = np.load('concept_arraynew2.npy')
# Returns discovered concepts with true clusters
finetuned_model_pr = ipca.ipca_model(concept_arraynew2, dense2, predict,
f_train, y_train, f_val, y_val,
n_concept)
num_epoch = 5
for _ in range(num_epoch):
finetuned_model_pr.fit(
f_train,
y_train,
batch_size=50,
epochs=10,
verbose=True,
validation_data=(f_val, y_val))
# Evaluates groupacc and get concept_matrix
concept_matrix, _ = ipca.get_groupacc(
finetuned_model_pr,
concept_arraynew2,
f_train,
f_val,
concept,
n_concept,
n_cluster,
n0,
verbose=False)
# Saves concept matrix
with open('concept_matrix_sup.pickle', 'wb') as handle:
pickle.dump(concept_matrix, handle, protocol=pickle.HIGHEST_PROTOCOL)
# Plots nearest neighbors
feature_sp1 = np.load('feature_sp1.npy')
segment_sp1 = np.load('segment_sp1.npy')
feature_sp1_1000 = feature_sp1[:1000]
segment_sp1_1000 = segment_sp1[:1000]
ipca.plot_nearestneighbor(concept_matrix, feature_sp1_1000, segment_sp1_1000)
# Discovered concepts with self-discovered clusters.
finetuned_model_pr = ipca.ipca_model(concept_arraynew, dense2, predict,
f_train, y_train, f_val, y_val,
n_concept)
num_epoch = 5
for _ in range(num_epoch):
finetuned_model_pr.fit(
f_train,
y_train,
batch_size=50,
epochs=10,
verbose=True,
validation_data=(f_val, y_val))
concept_matrix, _ = toy_helper.get_groupacc(
finetuned_model_pr,
concept_arraynew,
f_train,
f_val,
concept,
n_concept,
n_cluster,
n0,
verbose=False)
# Saves concept matrix.
with open('concept_matrix_unsup.pickle', 'wb') as handle:
pickle.dump(concept_matrix, handle, protocol=pickle.HIGHEST_PROTOCOL)
# Plots nearest neighbors.
toy_helper.plot_nearestneighbor(concept_matrix,
feature_sp1_1000, segment_sp1_1000)