def normalize_datasets(datasets):
datasets[0], mean, std, _ = normalize(datasets[0], remove_constant=False)
datasets[1], _, _, _ = normalize(datasets[1],
mean,
std,
remove_constant=False)
datasets[2], _, _, _ = normalize(datasets[2],
mean,
std,
remove_constant=False)
if zshot:
datasets[3], _, _, _ = normalize(datasets[3],
mean,
std,
remove_constant=False)
return datasets
if len(data.shape) > 2:
shape = reduce(mul, data.shape[1:], 1)
data = data.reshape([-1, shape])
return data
data = load_data()
# Get training data
n_train = int(len(data) * train_fract)
data = data[:n_train] #reuse var to save memory
if gaps:
gap_ids = np.load(os.path.join(data_dir, 'gap_ids.npy'))
data = np.delete(data, gap_ids, 0)
# Fit model to normalized training data
data, m, s, fs = normalize(data)
rng = np.random.RandomState(123)
model = PCA(n_components=n_components, random_state=rng)
model.fit(data)
# Encode all (normalized) data
data = load_data()
data, _, _, _ = normalize(data, m, s, fs)
data = model.transform(data) # codes, reuse var
expl_var = np.array(np.sum(model.explained_variance_ratio_))
# Save results
codes_dir = os.path.join(data_dir, 'codes/')
f_name = os.path.join(codes_dir, "pca_{0}_{1}v".format(gaps, n_components))
np.savez_compressed(f_name, codes=data, expl_var=expl_var)
print("PCA values saved to {1}.".format(f_name))
def main():
causal = False
if causal:
config_files = [
"VAEConv2D_v2_CausalDsprite_ber_shape2_scale5_ld2",
"VAEConv2D_v2_CausalDsprite_ber_shape2_scale5_ld3",
"VAEConv2D_v2_CausalDsprite_ber_shape2_scale5_ld4",
"VAEConv2D_v2_CausalDsprite_ber_shape2_scale5_ld6",
"VAEConv2D_v2_CausalDsprite_ber_shape2_scale5_ld10"
]
model_names = ["C-ld-2", "C-ld-3", "C-ld-4", "C-ld-6", "C-ld-10"]
#checkpoint_files=["vae_checkpoint"+ str(i) + ".pth" for i in range(50)]
checkpoint_files = ["vae_lastCheckpoint.pth"]
dataset_zip = np.load(
'datasets/dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz')
print('Keys in the dataset:', dataset_zip.keys())
imgs = dataset_zip['imgs']
#get the idxs
data_sets = []
data_sets_true = []
d_sprite_idx, X_true_data, _ = caus_utils.calc_dsprite_idxs(
num_samples=10000,
seed=12345,
constant_factor=[0, 0],
causal=causal,
color=0,
shape=2,
scale=5)
X_true_data = np.array(X_true_data)[:, :2]
#add a random g dimesnoion?
#r_gen_data=[]
#for r in range(len(X_true_data)):
# r_gen_data.append(random.uniform(0, 1))
#X_true_data=np.array(X_true_data)
#r_gen_data=np.array(r_gen_data)
#print(X_true_data.shape)
#print(r_gen_data.shape)
#X_true_data=np.c_[ X_true_data, r_gen_data]
X_data = caus_utils.make_dataset_d_sprite(d_sprite_dataset=imgs,
dsprite_idx=d_sprite_idx,
img_size=256)
print(len(X_data))
X_data_t = X_data[8500:]
X_data = X_data[:8500]
X_true_data_t = X_true_data[8500:]
X_true_data = X_true_data[:8500]
if not causal:
# the custem girls dataset
#config_files=["VAE_NonCausalDsprite_ber_shape2_scale5_ld2","VAE_NonCausalDsprite_ber_shape2_scale5_ld3","VAE_NonCausalDsprite_ber_shape2_scale5_ld4"
#,"VAE_NonCausalDsprite_ber_shape2_scale5_ld6","VAE_NonCausalDsprite_ber_shape2_scale5_ld10"]
#model_names=["NC_ds_ld2","NC_ds_ld3","NC_ds_ld4","NC_ds_ld6","NC_ds_ld10",]
config_files = [
"VAEConv2D_v2_NonCausalDsprite_ber_shape2_scale5_ld2",
"VAEConv2D_v2_NonCausalDsprite_ber_shape2_scale5_ld3",
"VAEConv2D_v2_NonCausalDsprite_ber_shape2_scale5_ld4",
"VAEConv2D_v2_NonCausalDsprite_ber_shape2_scale5_ld6",
"VAEConv2D_v2_NonCausalDsprite_ber_shape2_scale5_ld10"
]
model_names = ["NC-ld-2", "NC-ld-3", "NC-ld-4", "NC-ld-6", "NC-ld-10"]
#checkpoint_files=["vae_checkpoint"+ str(i) + ".pth" for i in range(50)]
checkpoint_files = ["vae_lastCheckpoint.pth"]
dataset_zip = np.load(
'datasets/dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz')
print('Keys in the dataset:', dataset_zip.keys())
imgs = dataset_zip['imgs']
#get the idxs
data_sets = []
data_sets_true = []
d_sprite_idx, X_true_data, _ = caus_utils.calc_dsprite_idxs(
num_samples=20000,
seed=12345,
constant_factor=[0, 0, 0],
causal=causal,
color=0,
shape=2,
scale=5)
d_sprite_idx_t, X_true_data_t, _ = caus_utils.calc_dsprite_idxs(
num_samples=1500,
seed=54321,
constant_factor=[0, 0, 0],
causal=causal,
color=0,
shape=2,
scale=5)
X_data = caus_utils.make_dataset_d_sprite(d_sprite_dataset=imgs,
dsprite_idx=d_sprite_idx,
img_size=256)
X_data_t = caus_utils.make_dataset_d_sprite(d_sprite_dataset=imgs,
dsprite_idx=d_sprite_idx_t,
img_size=256)
#split in train and test
X_data_train = np.array(X_data)
X_data_test = np.array(X_data_t)
X_true_data_train, _, _, _ = normalize(np.array(X_true_data),
remove_constant=False)
X_true_data_test, _, _, _ = normalize(np.array(X_true_data_t),
remove_constant=False)
#swap to see
#X_data_train[:,[0, 2]] = X_data_train[:,[2, 0]]
#X_data_test[:,[0, 2]] = X_data_test[:,[2, 0]]
#X_true_data_train[:,[0, 1]] = X_true_data_train[:,[1, 0]]
#X_true_data_test[:,[0, 1]] = X_true_data_test[:,[1, 0]]
zs_train_all = []
zs_test_all = []
for config_file in config_files:
for checkpoint_file in checkpoint_files:
print("obtaining representations (zs)")
print(config_file)
print(checkpoint_file)
zs_train = obtain_representation(X_data_train, config_file,
checkpoint_file)
#plt.hist(zs_train, bins='auto')
#plt.show()
print(config_file)
print("mean: " + str(np.mean(zs_train, axis=0)))
print("std: " + str(np.std(zs_train, axis=0)))
print("min: " + str(np.min(zs_train, axis=0)))
print("max: " + str(np.max(zs_train, axis=0)))
zs_test = obtain_representation(X_data_test, config_file,
checkpoint_file)
zs_train_n, _, _, _ = normalize(zs_train, remove_constant=False)
zs_test_n, _, _, _ = normalize(zs_test, remove_constant=False)
zs_train_all.append(zs_train_n)
zs_test_all.append(zs_test_n)
lasso(z_encodes_train=zs_train_all,
z_encodes_test=zs_test_all,
gt_labels_train=X_true_data_train,
gt_labels_test=X_true_data_test,
model_names=model_names)
#random_forest(z_encodes_train=zs_train_all,z_encodes_test=zs_test_all,gt_labels_train=X_true_data_train,gt_labels_test=X_true_data_test,model_names=model_names)
print("DONZO!")
def main():
causal = False
nn = MLPRegressor(hidden_layer_sizes=(10, ),
activation='relu',
solver='adam',
alpha=0.001,
batch_size='auto',
learning_rate='constant',
learning_rate_init=0.01,
power_t=0.5,
max_iter=1000,
shuffle=True,
random_state=9,
tol=0.0001,
verbose=False,
warm_start=False,
momentum=0.9,
nesterovs_momentum=True,
early_stopping=False,
validation_fraction=0.1,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08)
if causal:
config_file = "VAEConv2D_v2_CausalDsprite_ber_shape2_scale5_ld3"
model_name = "C-ld-3"
#checkpoint_files=["vae_checkpoint"+ str(i) + ".pth" for i in range(50)]
checkpoint_file = "vae_lastCheckpoint.pth"
dataset_zip = np.load(
'datasets/dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz')
print('Keys in the dataset:', dataset_zip.keys())
imgs = dataset_zip['imgs']
#get the idxs
data_sets = []
data_sets_true = []
#find corrsbonding axes X
z_x = -1
d_sprite_idx, X_true_data, _ = caus_utils.calc_dsprite_idxs(
num_samples=1000,
seed=7,
constant_factor=[1, 0],
causal=causal,
color=0,
shape=2,
scale=5)
X_data = caus_utils.make_dataset_d_sprite(d_sprite_dataset=imgs,
dsprite_idx=d_sprite_idx,
img_size=256)
zs_train = obtain_representation(X_data, config_file, checkpoint_file)
zs_train = np.array(zs_train)
zs_std = np.std(zs_train, axis=0)
z_x_idx = np.argmin(zs_std)
print("X-CONSTANT: ")
print(zs_std)
#find corrsbonding axes X
z_y = -1
d_sprite_idx, Y_true_data, _ = caus_utils.calc_dsprite_idxs(
num_samples=1000,
seed=7,
constant_factor=[0, 1],
causal=causal,
color=0,
shape=2,
scale=5)
Y_data = caus_utils.make_dataset_d_sprite(d_sprite_dataset=imgs,
dsprite_idx=d_sprite_idx,
img_size=256)
zs_train = obtain_representation(Y_data, config_file, checkpoint_file)
zs_train = np.array(zs_train)
zs_std = np.std(zs_train, axis=0)
z_y_idx = np.argmin(zs_std)
print("Y-CONSTANT: ")
print(zs_std)
if z_y_idx == z_x_idx:
print("z_y: " + str(z_y_idx))
print("z_x: " + str(z_x_idx))
print("Fatal error! could not get unique g to z")
a = 1 / 0
#time to learn some causal relationships!!!!
d_sprite_idx, true_data, _ = caus_utils.calc_dsprite_idxs(
num_samples=10000,
seed=12345,
constant_factor=[0, 0],
causal=causal,
color=0,
shape=2,
scale=5)
data = caus_utils.make_dataset_d_sprite(d_sprite_dataset=imgs,
dsprite_idx=d_sprite_idx,
img_size=256)
zs_data = obtain_representation(data, config_file, checkpoint_file)
#normalize data
true_data_norm, _, _, _ = normalize(np.array(true_data),
remove_constant=False)
zs_data_norm, _, _, _ = normalize(np.array(zs_data),
remove_constant=False)
#train input predictor:
input_x_y = true_data_norm[:, :2]
input_o = np.array(true_data_norm[:, 2])
n = nn.fit(input_x_y[:8500], input_o[:8500])
test = nn.predict(input_x_y[8500:])
input_mse = np.sum(np.square(test - input_o[8500:]))
print("causal input mse: " + str(input_mse))
#train latent predictor:
z_x_y = np.array([zs_data_norm[:, z_x_idx], zs_data_norm[:,
z_y_idx]]).T
z_o_idx_t = [0, 1, 2]
z_o_idx = -1
for i in range(len(z_o_idx_t)):
if not z_o_idx_t[i] == z_y_idx and not z_o_idx_t[i] == z_x_idx:
z_o_idx = z_o_idx_t[i]
z_o = zs_data_norm[:, z_o_idx]
n = nn.fit(z_x_y[:8500], z_o[:8500])
test = nn.predict(z_x_y[8500:])
z_mse = np.sum(np.square(test - z_o[8500:]))
print("causal latent mse: " + str(z_mse))
if not causal:
config_file = "VAEConv2D_v2_NonCausalDsprite_ber_shape2_scale5_ld3"
model_name = "NC-ld-3"
#checkpoint_files=["vae_checkpoint"+ str(i) + ".pth" for i in range(50)]
checkpoint_file = "vae_lastCheckpoint.pth"
dataset_zip = np.load(
'datasets/dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz')
print('Keys in the dataset:', dataset_zip.keys())
imgs = dataset_zip['imgs']
#get the idxs
data_sets = []
data_sets_true = []
#find corrsbonding axes X
z_x = -1
d_sprite_idx, X_true_data, _ = caus_utils.calc_dsprite_idxs(
num_samples=1000,
seed=7,
constant_factor=[1, 0, 0],
causal=causal,
color=0,
shape=2,
scale=5)
X_data = caus_utils.make_dataset_d_sprite(d_sprite_dataset=imgs,
dsprite_idx=d_sprite_idx,
img_size=256)
zs_train = obtain_representation(X_data, config_file, checkpoint_file)
zs_train = np.array(zs_train)
zs_std = np.std(zs_train, axis=0)
z_x_idx = np.argmin(zs_std)
print("X-CONSTANT: ")
print(zs_std)
#find corrsbonding axes X
z_y = -1
d_sprite_idx, Y_true_data, _ = caus_utils.calc_dsprite_idxs(
num_samples=1000,
seed=7,
constant_factor=[0, 1, 0],
causal=causal,
color=0,
shape=2,
scale=5)
Y_data = caus_utils.make_dataset_d_sprite(d_sprite_dataset=imgs,
dsprite_idx=d_sprite_idx,
img_size=256)
zs_train = obtain_representation(Y_data, config_file, checkpoint_file)
zs_train = np.array(zs_train)
zs_std = np.std(zs_train, axis=0)
z_y_idx = np.argmin(zs_std)
print("Y-CONSTANT: ")
print(zs_std)
#find corrsbonding axes O
z_y = -1
d_sprite_idx, O_true_data, _ = caus_utils.calc_dsprite_idxs(
num_samples=1000,
seed=7,
constant_factor=[0, 0, 1],
causal=causal,
color=0,
shape=2,
scale=5)
O_data = caus_utils.make_dataset_d_sprite(d_sprite_dataset=imgs,
dsprite_idx=d_sprite_idx,
img_size=256)
zs_train = obtain_representation(O_data, config_file, checkpoint_file)
zs_train = np.array(zs_train)
zs_std = np.std(zs_train, axis=0)
z_o_idx = np.argmin(zs_std)
print("O-CONSTANT: ")
print(zs_std)
if z_y_idx == z_x_idx or z_y_idx == z_o_idx or z_o_idx == z_x_idx:
print("z_y: " + str(z_y_idx))
print("z_x: " + str(z_x_idx))
print("z_o: " + str(z_o_idx))
print("Fatal error! could not get unique g to z")
a = 1 / 0
#time to learn some causal relationships!!!!
d_sprite_idx, true_data, _ = caus_utils.calc_dsprite_idxs(
num_samples=10000,
seed=12345,
constant_factor=[0, 0, 0],
causal=causal,
color=0,
shape=2,
scale=5)
data = caus_utils.make_dataset_d_sprite(d_sprite_dataset=imgs,
dsprite_idx=d_sprite_idx,
img_size=256)
zs_data = obtain_representation(data, config_file, checkpoint_file)
#normalize data
true_data_norm, _, _, _ = normalize(np.array(true_data),
remove_constant=False)
zs_data_norm, _, _, _ = normalize(np.array(zs_data),
remove_constant=False)
#train input predictor:
input_x_y = true_data_norm[:, :2]
input_o = np.array(true_data_norm[:, 2])
n = nn.fit(input_x_y[:8500], input_o[:8500])
test = nn.predict(input_x_y[8500:])
input_mse = np.sum(np.square(test - input_o[8500:]))
print("Non-causal input mse: " + str(input_mse))
#train latent predictor:
z_x_y = np.array([zs_data_norm[:, z_x_idx], zs_data_norm[:,
z_y_idx]]).T
z_o = zs_data_norm[:, z_o_idx]
n = nn.fit(z_x_y[:8500], z_o[:8500])
test = nn.predict(z_x_y[8500:])
z_mse = np.sum(np.square(test - z_o[8500:]))
print("Non-causal latent mse: " + str(z_mse))