def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
data_file = join(RAW_DATA_DIR, "ComputerVision", "dSprites",
"dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz")
# It is already in the range [0, 1]
with np.load(data_file, encoding="latin1") as f:
x_train = f['imgs']
# 3 shape * 6 scale * 40 rotation * 32 pos X * 32 pos Y
y_train = f['latents_classes']
x_train = np.expand_dims(x_train.astype(np.float32), axis=-1)
num_train = len(x_train)
print("num_train: {}".format(num_train))
# =====================================
# Instantiate model
# =====================================
if args.enc_dec_model == "1Konny":
encoder = Encoder_1Konny(args.z_dim, stochastic=True)
decoder = Decoder_1Konny()
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support enc_dec_model='{}'!".format(
args.enc_dec_model))
model = FactorVAE([64, 64, 1],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
'Dz_tc_loss_coeff': args.Dz_tc_loss_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
save_dir = make_dir_if_not_exist(
join(args.save_dir, "{}_{}".format(args.enc_dec_model, args.run)))
# =====================================
np.set_printoptions(threshold=np.nan,
linewidth=1000,
precision=5,
suppress=True)
num_samples = args.num_samples
print("num_samples: {}".format(num_samples))
# Compute representations
# ================================= #
z_data_file = join(save_dir, "z_data.npz")
if not exists(z_data_file):
all_z_samples = []
all_z_mean = []
all_z_stddev = []
print("")
print("Compute all_z_mean, all_z_stddev and all_attrs!")
count = 0
for batch_ids in iterate_data(num_train,
10 * args.batch_size,
shuffle=False):
x = x_train[batch_ids]
z_samples, z_mean, z_stddev = sess.run(model.get_output(
['z1_gen', 'z_mean', 'z_stddev']),
feed_dict={
model.is_train: False,
model.x_ph: x
})
all_z_samples.append(z_samples)
all_z_mean.append(z_mean)
all_z_stddev.append(z_stddev)
count += len(batch_ids)
print("\rProcessed {} samples!".format(count), end="")
print()
all_z_samples = np.concatenate(all_z_samples, axis=0)
all_z_mean = np.concatenate(all_z_mean, axis=0)
all_z_stddev = np.concatenate(all_z_stddev, axis=0)
np.savez_compressed(z_data_file,
all_z_samples=all_z_samples,
all_z_mean=all_z_mean,
all_z_stddev=all_z_stddev)
else:
print("{} exists. Load data from file!".format(z_data_file))
with np.load(z_data_file, "r") as f:
all_z_samples = f['all_z_samples']
all_z_mean = f['all_z_mean']
all_z_stddev = f['all_z_stddev']
# ================================= #
all_z_samples = np.reshape(all_z_samples, [3, 6, 40, 32, 32, -1])
all_z_mean = np.reshape(all_z_mean, [3, 6, 40, 32, 32, -1])
all_z_stddev = np.reshape(all_z_stddev, [3, 6, 40, 32, 32, -1])
if args.gpu_support == 'cupy':
print("Use cupy instead of numpy!")
results = MIG_4_dSprites_cupy(all_z_samples,
all_z_mean,
all_z_stddev,
version=1,
batch_size=10,
num_samples=num_samples,
gpu=args.gpu_id)
else:
results = MIG_4_dSprites(all_z_samples,
all_z_mean,
all_z_stddev,
num_samples=num_samples,
version=1,
batch_size=200)
result_file = join(save_dir,
"results[num_samples={}].npz".format(num_samples))
np.savez_compressed(result_file, **results)
f.close()
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
celebA_loader = TFCelebAWithAttrLoader(root_dir=args.celebA_root_dir)
img_height, img_width = args.celebA_resize_size, args.celebA_resize_size
celebA_loader.build_transformation_flow_tf(
*celebA_loader.get_transform_fns("1Konny",
resize_size=args.celebA_resize_size))
num_train = celebA_loader.num_train_data
# =====================================
# Instantiate model
# =====================================
if args.activation == "relu":
activation = tf.nn.relu
elif args.activation == "leaky_relu":
activation = tf.nn.leaky_relu
else:
raise ValueError("Do not support '{}' activation!".format(
args.activation))
if args.enc_dec_model == "1Konny":
# assert args.z_dim == 65, "For 1Konny, z_dim must be 65. Found {}!".format(args.z_dim)
encoder = Encoder_1Konny(args.z_dim,
stochastic=True,
activation=activation)
decoder = Decoder_1Konny([img_height, img_width, 3],
activation=activation,
output_activation=tf.nn.sigmoid)
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support encoder/decoder model '{}'!".format(
args.enc_dec_model))
model = FactorVAE([img_height, img_width, 3],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
# save_dir = remove_dir_if_exist(join(args.save_dir, "FactorVAE_{}".format(args.run)), ask_4_permission=True)
# save_dir = make_dir_if_not_exist(save_dir)
save_dir = make_dir_if_not_exist(
join(args.save_dir, "FactorVAE_{}".format(args.run)))
# =====================================
np.set_printoptions(threshold=np.nan,
linewidth=1000,
precision=3,
suppress=True)
num_bins = args.num_bins
bin_limits = tuple([float(s) for s in args.bin_limits.split(";")])
data_proportion = args.data_proportion
f = open(join(
save_dir, 'log[bins={},bin_limits={},data={}].txt'.format(
num_bins, bin_limits, data_proportion)),
mode='w')
print_ = functools.partial(print_both, file=f)
result_file = join(
args.interpretability_metrics_dir, "FactorVAE_{}".format(args.run),
"results[bins={},bin_limits={},data={}].npz".format(
num_bins, bin_limits, data_proportion))
results = np.load(result_file, "r")
print_("")
print_("num_bins: {}".format(num_bins))
print_("bin_limits: {}".format(bin_limits))
print_("data_proportion: {}".format(data_proportion))
# Plotting
# =========================================== #
# seed = 389
# num_samples = 30
seed = 398
num_samples = 1
ids = list(range(seed, seed + num_samples))
print_("\nids: {}".format(ids))
data = celebA_loader.sample_images_from_dataset(sess, 'train', ids)
span = 3
points_one_side = 5
attr_names = celebA_loader.attributes
print_("#attrs: {}".format(len(attr_names)))
print_("attr_names: {}".format(attr_names))
print_("results.keys: {}".format(list(results.keys())))
# (z_dim, num_attrs)
MI_ids_sorted = results['MI_ids_sorted']
MI_sorted = results['MI_sorted']
MI_gap_y = results['MI_gap_y']
H_y = results['H_y_4_diff_z'][:, 0]
assert MI_ids_sorted.shape[1] == len(attr_names) == len(MI_gap_y) == len(H_y), \
"MI_ids_sorted.shape: {}, len(attr_names): {}, len(MI_gap_y): {}, len(H_y): {}".format(
MI_ids_sorted.shape, len(attr_names), len(MI_gap_y), len(H_y))
print_("\nShow RMIG!")
for i in range(len(attr_names)):
print("{}: RMIG: {:.4f}, RMIG (unnorm): {:.4f}, H: {:.4f}".format(
attr_names[i], MI_gap_y[i], MI_gap_y[i] * H_y[i], H_y[i]))
print_("\nShow JEMMIG!")
H_z_y = results['H_z_y']
MI_z_y = results['MI_z_y']
ids_sorted_by_MI = np.argsort(MI_z_y, axis=0)[::-1]
MI_z_y_sorted = np.take_along_axis(MI_z_y, ids_sorted_by_MI, axis=0)
H_z_y_sorted = np.take_along_axis(H_z_y, ids_sorted_by_MI, axis=0)
H_diff = H_z_y_sorted[0, :] - MI_z_y_sorted[0, :]
JEMMIG_unnorm = H_diff + MI_z_y_sorted[1, :]
JEMMIG_norm = JEMMIG_unnorm / (np.log(num_bins) + H_y)
for i in range(len(attr_names)):
print(
"{}: JEMMIG: {:.4f}, JEMMIG (unnorm): {:.4f}, H_diff: {:.4f}, I2: {:.4f}, top 2 latents: z{}, z{}"
.format(attr_names[i], JEMMIG_norm[i], JEMMIG_unnorm[i], H_diff[i],
MI_z_y_sorted[1, i], ids_sorted_by_MI[0, i],
ids_sorted_by_MI[1, i]))
# Plot JEMMIG/RMIG of all attributes
# ========================================= #
import matplotlib.pyplot as plt
import matplotlib
from matplotlib.backends.backend_pdf import PdfPages
font = {'family': 'normal', 'size': 12}
matplotlib.rc('font', **font)
width = 0.3
plt.bar(np.arange(len(attr_names)),
JEMMIG_norm,
width=width,
align='center',
edgecolor='black',
label='JEMMIG')
plt.bar(np.arange(len(attr_names)) + width,
MI_gap_y,
width=width,
align='center',
edgecolor='black',
label='RMIG')
plt.xticks(np.arange(len(attr_names)) + 0.5 * width,
attr_names,
rotation=90)
plt.xlabel("model")
plt.ylabel("JEMMIG/RMIG")
# subplot_adjust = {'left': 0.08, 'right': 0.99, 'bottom': 0.17, 'top': 0.98}
# plt.subplots_adjust(**subplot_adjust)
plt.gcf().set_size_inches(12, 3)
save_dir = make_dir_if_not_exist(save_dir)
save_file = join(save_dir, "JEMMIG_tc_beta.pdf")
with PdfPages(save_file) as pdf_file:
plt.savefig(pdf_file, format='pdf')
plt.show()
plt.close()
# ========================================= #
# Uncomment if you want
'''
for n in range(len(ids)):
for k in range(len(attr_names)):
MI_ids_top10 = MI_ids_sorted[:10, k]
MI_top10 = MI_sorted[:10, k]
print("Plot top 10 latents for factor '{}'!".format(attr_names[k]))
img_file = join(save_dir, "x_train[{}][attr={}][bins={},bin_limits={},data={}].png".
format(ids[n], attr_names[k], num_bins, bin_limits, data_proportion))
model.cond_all_latents_traverse_v2(img_file, sess, data[n],
z_comps=MI_ids_top10,
z_comp_labels=["z[{}] ({:.4f})".format(comp, mi)
for comp, mi in zip(MI_ids_top10, MI_top10)],
span=span, points_1_side=points_one_side,
hl_x=True,
font_size=9,
title="{} (MI gap={:.4f}, H={:.4f})".format(
attr_names[k], MI_gap_y[k], H_y[k]),
title_font_scale=1.5,
subplot_adjust={'left': 0.16, 'right': 0.99,
'bottom': 0.01, 'top': 0.95},
size_inches=(6.5, 5.2),
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
#'''
# Top 5 only
for n in range(len(ids)):
for k in range(len(attr_names)):
MI_ids_top10 = MI_ids_sorted[:5, k]
MI_top10 = MI_sorted[:5, k]
print("Plot top 5 latents for factor '{}'!".format(attr_names[k]))
img_file = join(
save_dir, "train{}_attr={}_bins={}_data={}.png".format(
ids[n], attr_names[k], num_bins, data_proportion))
model.cond_all_latents_traverse_v2(
img_file,
sess,
data[n],
z_comps=MI_ids_top10,
z_comp_labels=[
"z[{}] ({:.4f})".format(comp, mi)
for comp, mi in zip(MI_ids_top10, MI_top10)
],
span=span,
points_1_side=points_one_side,
hl_x=True,
font_size=9,
title="{} (MI gap={:.4f}, H={:.4f})".format(
attr_names[k], MI_gap_y[k], H_y[k]),
title_font_scale=1.5,
subplot_adjust={
'left': 0.16,
'right': 0.99,
'bottom': 0.005,
'top': 0.93
},
size_inches=(6.5, 2.8),
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
'''
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
celebA_loader = TFCelebALoader(root_dir=args.celebA_root_dir)
img_height, img_width = args.celebA_resize_size, args.celebA_resize_size
celebA_loader.build_transformation_flow_tf(
*celebA_loader.get_transform_fns("1Konny", resize_size=args.celebA_resize_size))
num_train = celebA_loader.num_train_data
# =====================================
# Instantiate model
# =====================================
if args.activation == "relu":
activation = tf.nn.relu
elif args.activation == "leaky_relu":
activation = tf.nn.leaky_relu
else:
raise ValueError("Do not support '{}' activation!".format(args.activation))
if args.enc_dec_model == "1Konny":
# assert args.z_dim == 65, "For 1Konny, z_dim must be 65. Found {}!".format(args.z_dim)
encoder = Encoder_1Konny(args.z_dim, stochastic=True, activation=activation)
decoder = Decoder_1Konny([img_height, img_width, 3], activation=activation,
output_activation=tf.nn.sigmoid)
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support encoder/decoder model '{}'!".format(args.enc_dec_model))
model = FactorVAE([img_height, img_width, 3], args.z_dim,
encoder=encoder, decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True, gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
# save_dir = remove_dir_if_exist(join(args.save_dir, "FactorVAE_{}".format(args.run)), ask_4_permission=True)
# save_dir = make_dir_if_not_exist(save_dir)
save_dir = make_dir_if_not_exist(join(args.save_dir, "FactorVAE_{}".format(args.run)))
# =====================================
np.set_printoptions(threshold=np.nan, linewidth=1000, precision=3, suppress=True)
num_bins = args.num_bins
data_proportion = args.data_proportion
bin_limits = tuple([float(s) for s in args.bin_limits.split(";")])
top_k = args.top_k
f = open(join(save_dir, 'log[bins={},bin_limits={},data={}].txt'.
format(num_bins, bin_limits, data_proportion)), mode='w')
print_ = functools.partial(print_both, file=f)
result_file = join(args.informativeness_metrics_dir, "FactorVAE_{}".format(args.run),
'results[bins={},bin_limits={},data={}].npz'.
format(num_bins, bin_limits, data_proportion))
results = np.load(result_file, "r")
print_("")
print_("num_bins: {}".format(num_bins))
print_("bin_limits: {}".format(bin_limits))
print_("data_proportion: {}".format(data_proportion))
print_("top_k: {}".format(top_k))
# Plotting
# =========================================== #
# seed = 389
# num_samples = 30
seed = 398
num_samples = 1
ids = list(range(seed, seed + num_samples))
print_("\nids: {}".format(ids))
data = celebA_loader.sample_images_from_dataset(sess, 'train', ids)
span = 3
points_one_side = 5
print_("sorted_MI: {}".format(results["sorted_MI_z_x"]))
print_("sorted_z_ids: {}".format(results["sorted_z_comps"]))
print_("sorted_norm_MI: {}".format(results["sorted_norm_MI_z_x"]))
print_("sorted_norm_z_ids: {}".format(results["sorted_norm_z_comps"]))
top_MI = results["sorted_MI_z_x"][:top_k]
top_z_ids = results["sorted_z_comps"][:top_k]
top_norm_MI = results["sorted_norm_MI_z_x"][:top_k]
top_norm_z_ids = results["sorted_norm_z_comps"][:top_k]
print("Matplotlib font size: {}".format(matplotlib.rcParams['font.size'],))
for n in range(len(ids)):
if top_k == 10:
print("Plot conditional all comps z traverse with train sample {}!".format(ids[n]))
img_file = join(save_dir, "x_train[{}][bins={},bin_limits={},data={}].png".
format(ids[n], num_bins, bin_limits, data_proportion))
model.cond_all_latents_traverse_v2(img_file, sess, data[n],
z_comps=top_z_ids,
z_comp_labels=["z[{}] ({:.2f})".format(comp, mi)
for comp, mi in zip(top_z_ids, top_MI)],
span=span, points_1_side=points_one_side,
hl_x=True,
font_size=matplotlib.rcParams['font.size'],
subplot_adjust={'left': 0.15, 'right': 0.99,
'bottom': 0.01, 'top': 0.99},
size_inches=(6.3, 4.9),
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
img_file = join(save_dir, "x_train[{}][bins={},bin_limits={},data={},norm].png".
format(ids[n], num_bins, bin_limits, data_proportion))
model.cond_all_latents_traverse_v2(img_file, sess, data[n],
z_comps=top_norm_z_ids,
z_comp_labels=["z[{}] ({:.2f})".format(comp, mi)
for comp, mi in zip(top_norm_z_ids, top_norm_MI)],
span=span, points_1_side=points_one_side,
hl_x=True,
font_size=matplotlib.rcParams['font.size'],
subplot_adjust={'left': 0.15, 'right': 0.99,
'bottom': 0.01, 'top': 0.99},
size_inches=(6.3, 4.9),
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
else:
print("Plot conditional all comps z traverse with train sample {}!".format(ids[n]))
img_file = join(save_dir, "x_train[{}][bins={},bin_limits={},data={}].png".
format(ids[n], num_bins, bin_limits, data_proportion))
model.cond_all_latents_traverse_v2(img_file, sess, data[n],
z_comps=top_z_ids,
z_comp_labels=["z[{}] ({:.2f})".format(comp, mi)
for comp, mi in zip(top_z_ids, top_MI)],
span=span, points_1_side=points_one_side,
hl_x=True,
font_size=5,
subplot_adjust={'left': 0.19, 'right': 0.99,
'bottom': 0.01, 'top': 0.99},
size_inches=(2.98, 9.85),
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
img_file = join(save_dir, "x_train[{}][bins={},bin_limits={},data={},norm].png".
format(ids[n], num_bins, bin_limits, data_proportion))
model.cond_all_latents_traverse_v2(img_file, sess, data[n],
z_comps=top_norm_z_ids,
z_comp_labels=["z[{}] ({:.2f})".format(comp, mi)
for comp, mi in zip(top_norm_z_ids, top_norm_MI)],
span=span, points_1_side=points_one_side,
hl_x=True,
font_size=5,
subplot_adjust={'left': 0.19, 'right': 0.99,
'bottom': 0.01, 'top': 0.99},
size_inches=(2.98, 9.85),
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# =========================================== #
f.close()
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
data_file = join(RAW_DATA_DIR, "ComputerVision", "dSprites",
"dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz")
# It is already in the range [0, 1]
with np.load(data_file, encoding="latin1") as f:
x_train = f['imgs']
# 3 shape * 6 scale * 40 rotation * 32 pos X * 32 pos Y
y_train = f['latents_classes'][:, 1:]
x_train = np.expand_dims(x_train.astype(np.float32), axis=-1)
num_train = len(x_train)
print("num_train: {}".format(num_train))
print("y_train[:10]: {}".format(y_train[:10]))
# =====================================
# Instantiate model
# =====================================
if args.enc_dec_model == "1Konny":
encoder = Encoder_1Konny(args.z_dim, stochastic=True)
decoder = Decoder_1Konny()
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support enc_dec_model='{}'!".format(args.enc_dec_model))
model = FactorVAE([64, 64, 1], args.z_dim,
encoder=encoder, decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True, gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
'Dz_tc_loss_coeff': args.Dz_tc_loss_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
save_dir = make_dir_if_not_exist(join(args.save_dir, "{}_{}".format(args.enc_dec_model, args.run)))
# =====================================
np.set_printoptions(threshold=np.nan, linewidth=1000, precision=5, suppress=True)
num_bins = args.num_bins
data_proportion = args.data_proportion
num_data = int(data_proportion * num_train)
assert num_data == num_train, "For dSprites, you must use all data!"
# file
f = open(join(save_dir, 'log[bins={},data={}].txt'.
format(num_bins, data_proportion)), mode='w')
# print function
print_ = functools.partial(print_both, file=f)
print_("num_bins: {}".format(num_bins))
print_("data_proportion: {}".format(data_proportion))
# Compute representations
# ================================= #
z_data_file = join(save_dir, "z_data[data={}].npz".format(data_proportion))
if not exists(z_data_file):
all_z_mean = []
all_z_stddev = []
print("")
print("Compute all_z_mean, all_z_stddev and all_attrs!")
count = 0
for batch_ids in iterate_data(num_data, 10 * args.batch_size, shuffle=False):
x = x_train[batch_ids]
z_mean, z_stddev = sess.run(
model.get_output(['z_mean', 'z_stddev']),
feed_dict={model.is_train: False, model.x_ph: x})
all_z_mean.append(z_mean)
all_z_stddev.append(z_stddev)
count += len(batch_ids)
print("\rProcessed {} samples!".format(count), end="")
print()
all_z_mean = np.concatenate(all_z_mean, axis=0)
all_z_stddev = np.concatenate(all_z_stddev, axis=0)
np.savez_compressed(z_data_file, all_z_mean=all_z_mean,
all_z_stddev=all_z_stddev)
else:
print("{} exists. Load data from file!".format(z_data_file))
with np.load(z_data_file, "r") as f:
all_z_mean = f['all_z_mean']
all_z_stddev = f['all_z_stddev']
print_("")
print_("all_z_mean.shape: {}".format(all_z_mean.shape))
print_("all_z_stddev.shape: {}".format(all_z_stddev.shape))
# ================================= #
# Transpose and compute MIG score
# ================================= #
assert len(all_z_mean) == len(y_train)
# (num_latents, num_samples)
all_z_mean = np.transpose(all_z_mean, [1, 0])
print_("")
print_("all_z_mean.shape: {}".format(all_z_mean.shape))
y_train = np.transpose(y_train, [1, 0])
print_("")
print_("y_train.shape: {}".format(y_train.shape))
# All
# --------------------------------- #
result_all = compute_mig(all_z_mean, y_train, is_discrete_z=False,
is_discrete_y=True, num_bins=num_bins)
# (num_latents, num_factors)
MI_gap_y = result_all['MI_gap_y']
attr_ids_sorted = np.argsort(MI_gap_y, axis=0)[::-1].tolist()
MI_gap_y_sorted = MI_gap_y[attr_ids_sorted].tolist()
print_("")
print_("MIG: {}".format(result_all['MIG']))
print_("Sorted factors:\n{}".format(list(zip(attr_ids_sorted, MI_gap_y_sorted))))
save_file = join(save_dir, "results[bins={},data={}].npz".format(num_bins, data_proportion))
np.savez_compressed(save_file, **result_all)
# --------------------------------- #
# ================================= #
f.close()
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
data_file = join(RAW_DATA_DIR, "ComputerVision", "dSprites",
"dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz")
# It is already in the range [0, 1]
with np.load(data_file, encoding="latin1") as f:
x_train = f['imgs']
x_train = np.reshape(x_train, [3, 6, 40, 32, 32, 64, 64, 1])
# =====================================
# Instantiate model
# =====================================
if args.enc_dec_model == "1Konny":
encoder = Encoder_1Konny(args.z_dim, stochastic=True)
decoder = Decoder_1Konny()
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support enc_dec_model='{}'!".format(
args.enc_dec_model))
model = FactorVAE([64, 64, 1],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
'Dz_tc_loss_coeff': args.Dz_tc_loss_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
save_dir = make_dir_if_not_exist(
join(args.save_dir, "{}_{}".format(args.enc_dec_model, args.run)))
# =====================================
np.set_printoptions(threshold=np.nan,
linewidth=1000,
precision=3,
suppress=True)
result_file = join(args.JEMMIG_sampling_dir,
"{}_{}".format(args.enc_dec_model, args.run),
"results[num_samples={}].npz".format(args.num_samples))
results = np.load(result_file, "r")
print("results.keys: {}".format(list(results.keys())))
# Plotting
# =========================================== #
data = [
x_train[0, 3, 20, 16, 16], x_train[1, 3, 20, 16, 16], x_train[2, 3, 20,
16, 16]
]
gt_factors = ['Shape', 'Scale', 'Rotation', 'Pos_x', 'Pos_y']
ids_sorted = results['id_sorted']
MI_zi_yk_sorted = results['MI_zi_yk_sorted']
H_zi_yk_sorted = results['H_zi_yk_sorted']
H_yk = results['H_yk']
RMIG_yk = results['RMIG_yk']
RMIG_norm_yk = results['RMIG_norm_yk']
JEMMIG_yk = results['JEMMIG_yk']
print("MI_zi_yk_sorted:\n{}".format(MI_zi_yk_sorted))
print("\nShow RMIG!")
for k in range(len(gt_factors)):
print(
"{}, RMIG: {:.4f}, RMIG (norm): {:.4f}, H: {:.4f}, I1: {:.4f}, I2: {:.4f}"
.format(gt_factors[k], RMIG_yk[k], RMIG_norm_yk[k], H_yk[k],
MI_zi_yk_sorted[0, k], MI_zi_yk_sorted[1, k]))
print("\nShow JEMMIG!")
for k in range(len(gt_factors)):
print(
"{}, JEMMIG: {:.4f}, H1: {:.4f}, H1-I1: {:.4f}, I2: {:.4f}, top2 ids: z{}, z{}"
.format(gt_factors[k], JEMMIG_yk[k], H_zi_yk_sorted[0, k],
H_zi_yk_sorted[0, k] - MI_zi_yk_sorted[0, k],
MI_zi_yk_sorted[1, k], ids_sorted[0, k], ids_sorted[1, k]))
span = 3
points_one_side = 5
for n in range(len(data)):
for k in range(len(gt_factors)):
print("x={}, y={}!".format(n, gt_factors[k]))
img_file = join(
save_dir,
"{}-x{}_num_samples={}].png".format(gt_factors[k], n,
args.num_samples))
ids_top3 = ids_sorted[:3, k]
MI_top3 = MI_zi_yk_sorted[:3, k]
model.cond_all_latents_traverse_v2(
img_file,
sess,
data[n],
z_comps=ids_top3,
z_comp_labels=[
"z[{}] ({:.4f})".format(comp, mi)
for comp, mi in zip(ids_top3, MI_top3)
],
span=span,
points_1_side=points_one_side,
hl_x=True,
font_size=9,
title="{} (RMIG={:.4f}, JEMMIG={:.4f}, H={:.4f})".format(
gt_factors[k], RMIG_yk[k], JEMMIG_yk[k], H_yk[k]),
title_font_scale=1.5,
subplot_adjust={
'left': 0.16,
'right': 0.99,
'bottom': 0.01,
'top': 0.88
},
size_inches=(6.2, 1.7),
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
f.close()
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
celebA_loader = TFCelebALoader(root_dir=args.celebA_root_dir)
img_height, img_width = args.celebA_resize_size, args.celebA_resize_size
celebA_loader.build_transformation_flow_tf(
*celebA_loader.get_transform_fns("1Konny",
resize_size=args.celebA_resize_size))
num_train = celebA_loader.num_train_data
# =====================================
# Instantiate model
# =====================================
if args.activation == "relu":
activation = tf.nn.relu
elif args.activation == "leaky_relu":
activation = tf.nn.leaky_relu
else:
raise ValueError("Do not support '{}' activation!".format(
args.activation))
if args.enc_dec_model == "1Konny":
# assert args.z_dim == 65, "For 1Konny, z_dim must be 65. Found {}!".format(args.z_dim)
encoder = Encoder_1Konny(args.z_dim,
stochastic=True,
activation=activation)
decoder = Decoder_1Konny([img_height, img_width, 3],
activation=activation,
output_activation=tf.nn.sigmoid)
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support encoder/decoder model '{}'!".format(
args.enc_dec_model))
model = FactorVAE([img_height, img_width, 3],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
save_dir = remove_dir_if_exist(join(args.save_dir,
"FactorVAE_{}".format(args.run)),
ask_4_permission=False)
save_dir = make_dir_if_not_exist(save_dir)
# save_dir = make_dir_if_not_exist(join(args.save_dir, "FactorVAE_{}".format(args.run)))
# =====================================
np.set_printoptions(threshold=np.nan,
linewidth=1000,
precision=3,
suppress=True)
f = open(join(save_dir, 'log.txt'), mode='w')
print_ = functools.partial(print_both, file=f)
# z gaussian stddev
# ======================================= #
all_z_mean = []
all_z_stddev = []
count = 0
for batch_ids in iterate_data(int(0.05 * num_train),
10 * args.batch_size,
shuffle=False):
x = celebA_loader.sample_images_from_dataset(sess, 'train', batch_ids)
z_mean, z_stddev = sess.run(model.get_output(['z_mean', 'z_stddev']),
feed_dict={
model.is_train: False,
model.x_ph: x
})
all_z_mean.append(z_mean)
all_z_stddev.append(z_stddev)
count += len(batch_ids)
print("\rProcessed {} samples!".format(count), end="")
print()
all_z_mean = np.concatenate(all_z_mean, axis=0)
all_z_stddev = np.concatenate(all_z_stddev, axis=0)
# ======================================= #
z_std_error = np.std(all_z_mean, axis=0, ddof=0)
z_sorted_comps = np.argsort(z_std_error)[::-1]
top10_z_comps = z_sorted_comps[:10]
print_("")
print_("z_std_error: {}".format(z_std_error))
print_("z_sorted_std_error: {}".format(z_std_error[z_sorted_comps]))
print_("z_sorted_comps: {}".format(z_sorted_comps))
print_("top10_z_comps: {}".format(top10_z_comps))
z_stddev_mean = np.mean(all_z_stddev, axis=0)
info_z_comps = [
idx for idx in range(len(z_stddev_mean)) if z_stddev_mean[idx] < 0.4
]
print_("info_z_comps: {}".format(info_z_comps))
print_("len(info_z_comps): {}".format(len(info_z_comps)))
# Plotting
# =========================================== #
seed = 389
num_samples = 30
ids = list(range(seed, seed + num_samples))
print("\nids: {}".format(ids))
data = celebA_loader.sample_images_from_dataset(sess, 'train', ids)
span = 3
points_one_side = 5
# span = 8
# points_one_side = 12
for n in range(len(ids)):
print("Plot conditional all comps z traverse with train sample {}!".
format(ids[n]))
img_file = join(save_dir,
"x_train[{}]_[span={}]_hl.png".format(ids[n], span))
# model.cond_all_latents_traverse_v2(img_file, sess, data[n],
# z_comps=top10_z_comps,
# z_comp_labels=None,
# span=span, points_1_side=points_one_side,
# hl_x=True,
# batch_size=args.batch_size,
# dec_output_2_img_func=binary_float_to_uint8)
img_file = join(
save_dir,
"x_train[{}]_[span={}]_hl_labeled.png".format(ids[n], span))
model.cond_all_latents_traverse_v2(
img_file,
sess,
data[n],
z_comps=top10_z_comps,
z_comp_labels=["z[{}]".format(comp) for comp in top10_z_comps],
span=span,
points_1_side=points_one_side,
hl_x=True,
subplot_adjust={
'left': 0.09,
'right': 0.98,
'bottom': 0.02,
'top': 0.98
},
size_inches=(6, 5),
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# img_file = join(save_dir, "x_train[{}]_[span={}].png".format(ids[n], span))
# model.cond_all_latents_traverse_v2(img_file, sess, data[n],
# z_comps=top10_z_comps,
# z_comp_labels=None,
# span=span, points_1_side=points_one_side,
# hl_x=False,
# batch_size=args.batch_size,
# dec_output_2_img_func=binary_float_to_uint8)
#
# img_file = join(save_dir, "x_train[{}]_[span={}]_labeled.png".format(ids[n], span))
# model.cond_all_latents_traverse_v2(img_file, sess, data[n],
# z_comps=top10_z_comps,
# z_comp_labels=["z[{}]".format(comp) for comp in top10_z_comps],
# span=span, points_1_side=points_one_side,
# hl_x=False,
# subplot_adjust={'left': 0.09, 'right': 0.98, 'bottom': 0.02, 'top': 0.98},
# size_inches=(6, 5),
# batch_size=args.batch_size,
# dec_output_2_img_func=binary_float_to_uint8)
img_file = join(
save_dir, "x_train[{}]_[span={}]_info_hl.png".format(ids[n], span))
model.cond_all_latents_traverse_v2(
img_file,
sess,
data[n],
z_comps=info_z_comps,
z_comp_labels=None,
span=span,
points_1_side=points_one_side,
hl_x=True,
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# =========================================== #
f.close()
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
celebA_loader = TFCelebALoader(root_dir=args.celebA_root_dir)
img_height, img_width = args.celebA_resize_size, args.celebA_resize_size
celebA_loader.build_transformation_flow_tf(
*celebA_loader.get_transform_fns("1Konny",
resize_size=args.celebA_resize_size))
# =====================================
# Instantiate model
# =====================================
if args.activation == "relu":
activation = tf.nn.relu
elif args.activation == "leaky_relu":
activation = tf.nn.leaky_relu
else:
raise ValueError("Do not support '{}' activation!".format(
args.activation))
if args.enc_dec_model == "1Konny":
assert args.z_dim == 65, "For 1Konny, z_dim must be 65. Found {}!".format(
args.z_dim)
encoder = Encoder_1Konny(args.z_dim,
stochastic=True,
activation=activation)
decoder = Decoder_1Konny([img_height, img_width, 3],
activation=activation,
output_activation=tf.nn.sigmoid)
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support encoder/decoder model '{}'!".format(
args.enc_dec_model))
model = FactorVAE([img_height, img_width, 3],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
# =====================================
# Reconstruct
# ======================================= #
seed = 341
rs = np.random.RandomState(seed)
ids = rs.choice(celebA_loader.num_test_data, size=15)
x = celebA_loader.sample_images_from_dataset(sess, 'test', ids)
save_dir = make_dir_if_not_exist(join(args.save_dir, args.run))
img_file = join(save_dir, 'x_test.png')
save_img_block(img_file, binary_float_to_uint8(np.expand_dims(x, axis=0)))
img_file = join(save_dir, 'recx_test_1.png')
model.reconstruct_images(img_file,
sess,
x,
block_shape=[1, len(ids)],
batch_size=-1,
show_original_images=False,
dec_output_2_img_func=binary_float_to_uint8)
img_file = join(save_dir, 'recx_test_2.png')
model.reconstruct_images(img_file,
sess,
x,
block_shape=[1, len(ids)],
batch_size=-1,
show_original_images=True,
dec_output_2_img_func=binary_float_to_uint8)
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
data_file = join(RAW_DATA_DIR, "ComputerVision", "dSprites",
"dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz")
# It is already in the range [0, 1]
with np.load(data_file, encoding="latin1") as f:
x_train = f['imgs']
# 3 shape * 6 scale * 40 rotation * 32 pos X * 32 pos Y
y_train = f['latents_classes']
x_train = np.expand_dims(x_train.astype(np.float32), axis=-1)
num_train = len(x_train)
print("num_train: {}".format(num_train))
print("y_train[:10]: {}".format(y_train[:10]))
# =====================================
# Instantiate model
# =====================================
if args.enc_dec_model == "1Konny":
encoder = Encoder_1Konny(args.z_dim, stochastic=True)
decoder = Decoder_1Konny()
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support enc_dec_model='{}'!".format(args.enc_dec_model))
model = FactorVAE([64, 64, 1], args.z_dim,
encoder=encoder, decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True, gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
'Dz_tc_loss_coeff': args.Dz_tc_loss_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
save_dir = make_dir_if_not_exist(join(args.save_dir, "{}_{}".format(args.enc_dec_model, args.run)))
# =====================================
np.set_printoptions(threshold=np.nan, linewidth=1000, precision=5, suppress=True)
num_bins = args.num_bins
bin_limits = tuple([float(s) for s in args.bin_limits.split(";")])
data_proportion = args.data_proportion
num_data = int(data_proportion * num_train)
assert num_data == num_train, "For dSprites, you must use all data!"
eps = 1e-8
# file
f = open(join(save_dir, 'log[bins={},bin_limits={},data={}].txt'.
format(num_bins, bin_limits, data_proportion)), mode='w')
# print function
print_ = functools.partial(print_both, file=f)
print_("num_bins: {}".format(num_bins))
print_("bin_limits: {}".format(bin_limits))
print_("data_proportion: {}".format(data_proportion))
# Compute bins
# ================================= #
print_("")
print_("bin_limits: {}".format(bin_limits))
assert len(bin_limits) == 2 and bin_limits[0] < bin_limits[1], "bin_limits={}".format(bin_limits)
bins = np.linspace(bin_limits[0], bin_limits[1], num_bins + 1, endpoint=True)
print_("bins: {}".format(bins))
assert len(bins) == num_bins + 1
bin_widths = [bins[b] - bins[b - 1] for b in range(1, len(bins))]
print_("bin_widths: {}".format(bin_widths))
assert len(bin_widths) == num_bins, "len(bin_widths)={} while num_bins={}!".format(len(bin_widths), num_bins)
assert np.all(np.greater(bin_widths, 0)), "bin_widths: {}".format(bin_widths)
bin_centers = [(bins[b] + bins[b - 1]) * 0.5 for b in range(1, len(bins))]
print_("bin_centers: {}".format(bin_centers))
assert len(bin_centers) == num_bins, "len(bin_centers)={} while num_bins={}!".format(len(bin_centers), num_bins)
# ================================= #
# Compute representations
# ================================= #
z_data_file = join(save_dir, "z_data[data={}].npz".format(data_proportion))
if not exists(z_data_file):
all_z_mean = []
all_z_stddev = []
print("")
print("Compute all_z_mean, all_z_stddev and all_attrs!")
count = 0
for batch_ids in iterate_data(num_data, 10 * args.batch_size, shuffle=False):
x = x_train[batch_ids]
z_mean, z_stddev = sess.run(
model.get_output(['z_mean', 'z_stddev']),
feed_dict={model.is_train: False, model.x_ph: x})
all_z_mean.append(z_mean)
all_z_stddev.append(z_stddev)
count += len(batch_ids)
print("\rProcessed {} samples!".format(count), end="")
print()
all_z_mean = np.concatenate(all_z_mean, axis=0)
all_z_stddev = np.concatenate(all_z_stddev, axis=0)
np.savez_compressed(z_data_file, all_z_mean=all_z_mean,
all_z_stddev=all_z_stddev)
else:
print("{} exists. Load data from file!".format(z_data_file))
with np.load(z_data_file, "r") as f:
all_z_mean = f['all_z_mean']
all_z_stddev = f['all_z_stddev']
# ================================= #
print_("")
all_Q_z_cond_x = []
for i in range(args.z_dim):
print_("\nCompute all_Q_z{}_cond_x!".format(i))
all_Q_s_cond_x = []
for batch_ids in iterate_data(len(all_z_mean), 500, shuffle=False, include_remaining=True):
# (batch_size, num_bins)
q_s_cond_x = normal_density(np.expand_dims(bin_centers, axis=0),
mean=np.expand_dims(all_z_mean[batch_ids, i], axis=-1),
stddev=np.expand_dims(all_z_stddev[batch_ids, i], axis=-1))
# (batch_size, num_bins)
max_q_s_cond_x = np.max(q_s_cond_x, axis=-1)
# print("\nmax_q_s_cond_x: {}".format(np.sort(max_q_s_cond_x)))
# (batch_size, num_bins)
deter_s_cond_x = at_bin(all_z_mean[batch_ids, i], bins).astype(np.float32)
# (batch_size, num_bins)
Q_s_cond_x = q_s_cond_x * np.expand_dims(bin_widths, axis=0)
Q_s_cond_x = Q_s_cond_x / np.maximum(np.sum(Q_s_cond_x, axis=1, keepdims=True), eps)
# print("sort(sum(Q_s_cond_x)) (before): {}".format(np.sort(np.sum(Q_s_cond_x, axis=-1))))
Q_s_cond_x = np.where(np.expand_dims(np.less(max_q_s_cond_x, 1e-5), axis=-1),
deter_s_cond_x, Q_s_cond_x)
# print("sort(sum(Q_s_cond_x)) (after): {}".format(np.sort(np.sum(Q_s_cond_x, axis=-1))))
all_Q_s_cond_x.append(Q_s_cond_x)
# (num_samples, num_bins)
all_Q_s_cond_x = np.concatenate(all_Q_s_cond_x, axis=0)
assert np.all(all_Q_s_cond_x >= 0), "'all_Q_s_cond_x' contains negative values. " \
"sorted_all_Q_s_cond_x[:30]:\n{}!".format(np.sort(all_Q_s_cond_x[:30], axis=None))
assert len(all_Q_s_cond_x) == num_train
all_Q_z_cond_x.append(all_Q_s_cond_x)
# (z_dim, num_samples, num_bins)
all_Q_z_cond_x = np.asarray(all_Q_z_cond_x, dtype=np.float32)
print_("all_Q_z_cond_x.shape: {}".format(all_Q_z_cond_x.shape))
print_("sum(all_Q_z_cond_x)[:, :10]:\n{}".format(np.sum(all_Q_z_cond_x, axis=-1)[:, :10]))
# (z_dim, num_bins)
Q_z = np.mean(all_Q_z_cond_x, axis=1)
log_Q_z = np.log(np.clip(Q_z, eps, 1-eps))
print_("Q_z.shape: {}".format(Q_z.shape))
print_("sum(Q_z): {}".format(np.sum(Q_z, axis=-1)))
# (z_dim, )
H_z = -np.sum(Q_z * log_Q_z, axis=-1)
# Factors
gt_factors = ['shape', 'scale', 'rotation', 'pos_x', 'pos_y']
gt_num_values = [3, 6, 40, 32, 32]
MI_z_y = np.zeros([args.z_dim, len(gt_factors)], dtype=np.float32)
H_z_y = np.zeros([args.z_dim, len(gt_factors)], dtype=np.float32)
ids_sorted = np.zeros([args.z_dim, len(gt_factors)], dtype=np.int32)
MI_z_y_sorted = np.zeros([args.z_dim, len(gt_factors)], dtype=np.float32)
H_z_y_sorted = np.zeros([args.z_dim, len(gt_factors)], dtype=np.float32)
H_y = []
RMIG = []
JEMMI = []
for k, (factor, num_values) in enumerate(zip(gt_factors, gt_num_values)):
print_("\n#" + "=" * 50 + "#")
print_("The {}-th gt factor '{}' has {} values!".format(k, factor, num_values))
print_("")
# (num_samples, num_categories)
# NOTE: We must use k+1 to account for the 'color' attribute, which is always white
all_Q_yk_cond_x = one_hot(y_train[:, k+1], num_categories=num_values, dtype=np.float32)
print_("all_Q_yk_cond_x.shape: {}".format(all_Q_yk_cond_x.shape))
# (num_categories)
Q_yk = np.mean(all_Q_yk_cond_x, axis=0)
log_Q_yk = np.log(np.clip(Q_yk, eps, 1-eps))
print_("Q_yk.shape: {}".format(Q_yk.shape))
H_yk = -np.sum(Q_yk * log_Q_yk)
print_("H_yk: {}".format(H_yk))
H_y.append(H_yk)
Q_z_yk = np.zeros([args.z_dim, num_bins, num_values], dtype=np.float32)
# Compute I(zi, yk)
for i in range(args.z_dim):
print_("\n#" + "-" * 50 + "#")
all_Q_zi_cond_x = all_Q_z_cond_x[i]
assert len(all_Q_zi_cond_x) == len(all_Q_yk_cond_x) == num_train, \
"all_Q_zi_cond_x.shape: {}, all_Q_yk_cond_x.shape: {}".format(
all_Q_zi_cond_x.shape, all_Q_yk_cond_x.shape)
# (num_bins, num_categories)
Q_zi_yk = np.matmul(np.transpose(all_Q_zi_cond_x, axes=[1, 0]), all_Q_yk_cond_x)
Q_zi_yk = Q_zi_yk / num_train
print_("np.sum(Q_zi_yk): {}".format(np.sum(Q_zi_yk)))
Q_zi_yk = Q_zi_yk / np.maximum(np.sum(Q_zi_yk), eps)
print_("np.sum(Q_zi_yk) (normalized): {}".format(np.sum(Q_zi_yk)))
assert np.all(Q_zi_yk >= 0), "'Q_zi_yk' contains negative values. " \
"sorted_Q_zi_yk[:10]:\n{}!".format(np.sort(Q_zi_yk, axis=None))
# (num_bins, num_categories)
log_Q_zi_yk = np.log(np.clip(Q_zi_yk, eps, 1 - eps))
print_("")
print_("Q_zi (default): {}".format(Q_z[i]))
print_("Q_zi (sum of Q_zi_yk over yk): {}".format(np.sum(Q_zi_yk, axis=-1)))
print_("")
print_("Q_yk (default): {}".format(Q_yk))
print_("Q_yk (sum of Q_zi_yk over zi): {}".format(np.sum(Q_zi_yk, axis=0)))
MI_zi_yk = Q_zi_yk * (log_Q_zi_yk -
np.expand_dims(log_Q_z[i], axis=-1) -
np.expand_dims(log_Q_yk, axis=0))
MI_zi_yk = np.sum(MI_zi_yk)
H_zi_yk = -np.sum(Q_zi_yk * log_Q_zi_yk)
Q_z_yk[i] = Q_zi_yk
MI_z_y[i, k] = MI_zi_yk
H_z_y[i, k] = H_zi_yk
print_("#" + "-" * 50 + "#")
# Print statistics for all z
print_("")
print_("MI_z_yk:\n{}".format(MI_z_y[:, k]))
print_("H_z_yk:\n{}".format(H_z_y[:, k]))
print_("H_z:\n{}".format(H_z))
print_("H_yk:\n{}".format(H_yk))
# Compute RMIG and JEMMI
ids_yk_sorted = np.argsort(MI_z_y[:, k], axis=0)[::-1]
MI_z_yk_sorted = np.take_along_axis(MI_z_y[:, k], ids_yk_sorted, axis=0)
H_z_yk_sorted = np.take_along_axis(H_z_y[:, k], ids_yk_sorted, axis=0)
RMIG_yk = np.divide(MI_z_yk_sorted[0] - MI_z_yk_sorted[1], H_yk)
JEMMI_yk = np.divide(H_z_yk_sorted[0] - MI_z_yk_sorted[0] + MI_z_yk_sorted[1],
H_yk + np.log(num_bins))
ids_sorted[:, k] = ids_yk_sorted
MI_z_y_sorted[:, k] = MI_z_yk_sorted
H_z_y_sorted[:, k] = H_z_yk_sorted
RMIG.append(RMIG_yk)
JEMMI.append(JEMMI_yk)
print_("")
print_("ids_sorted: {}".format(ids_sorted))
print_("MI_z_yk_sorted: {}".format(MI_z_yk_sorted))
print_("RMIG_yk: {}".format(RMIG_yk))
print_("JEMMI_yk: {}".format(JEMMI_yk))
z_yk_prob_file = join(save_dir, "z_yk_prob_4_{}[bins={},bin_limits={},data={}].npz".
format(factor, num_bins, bin_limits, data_proportion))
np.savez_compressed(z_yk_prob_file, Q_z_yk=Q_z_yk)
print_("#" + "=" * 50 + "#")
results = {
"MI_z_y": MI_z_y,
"H_z_y": H_z_y,
"ids_sorted": ids_sorted,
"MI_z_y_sorted": MI_z_y_sorted,
"H_z_y_sorted": H_z_y_sorted,
"H_z": H_z,
"H_y": np.asarray(H_y, dtype=np.float32),
"RMIG": np.asarray(RMIG, dtype=np.float32),
"JEMMI": np.asarray(JEMMI, dtype=np.float32),
}
result_file = join(save_dir, "results[bins={},bin_limits={},data={}].npz".
format(num_bins, bin_limits, data_proportion))
np.savez_compressed(result_file, **results)
f.close()
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
data_file = join(RAW_DATA_DIR, "ComputerVision", "dSprites",
"dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz")
# It is already in the range [0, 1]
with np.load(data_file, encoding="latin1") as f:
x_train = f['imgs']
# 3 shape * 6 scale * 40 rotation * 32 pos X * 32 pos Y
y_train = f['latents_classes']
x_train = np.expand_dims(x_train.astype(np.float32), axis=-1)
num_train = len(x_train)
print("num_train: {}".format(num_train))
print("y_train[:10]: {}".format(y_train[:10]))
# =====================================
# Instantiate model
# =====================================
if args.enc_dec_model == "1Konny":
encoder = Encoder_1Konny(args.z_dim, stochastic=True)
decoder = Decoder_1Konny()
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support enc_dec_model='{}'!".format(
args.enc_dec_model))
model = FactorVAE([64, 64, 1],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
'Dz_tc_loss_coeff': args.Dz_tc_loss_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
save_dir = make_dir_if_not_exist(
join(args.save_dir, "{}_{}".format(args.enc_dec_model, args.run)))
# =====================================
np.set_printoptions(threshold=np.nan,
linewidth=1000,
precision=5,
suppress=True)
num_bins = args.num_bins
bin_limits = tuple([float(s) for s in args.bin_limits.split(";")])
data_proportion = args.data_proportion
num_data = int(data_proportion * num_train)
assert num_data == num_train, "For dSprites, you must use all data!"
eps = 1e-8
# file
f = open(join(
save_dir, 'log[bins={},bin_limits={},data={}].txt'.format(
num_bins, bin_limits, data_proportion)),
mode='w')
# print function
print_ = functools.partial(print_both, file=f)
print_("num_bins: {}".format(num_bins))
print_("bin_limits: {}".format(bin_limits))
print_("data_proportion: {}".format(data_proportion))
# Compute bins
# ================================= #
print_("")
print_("bin_limits: {}".format(bin_limits))
assert len(bin_limits) == 2 and bin_limits[0] < bin_limits[
1], "bin_limits={}".format(bin_limits)
bins = np.linspace(bin_limits[0],
bin_limits[1],
num_bins + 1,
endpoint=True)
print_("bins: {}".format(bins))
assert len(bins) == num_bins + 1
bin_widths = [bins[b] - bins[b - 1] for b in range(1, len(bins))]
print_("bin_widths: {}".format(bin_widths))
assert len(bin_widths
) == num_bins, "len(bin_widths)={} while num_bins={}!".format(
len(bin_widths), num_bins)
assert np.all(np.greater(bin_widths,
0)), "bin_widths: {}".format(bin_widths)
bin_centers = [(bins[b] + bins[b - 1]) * 0.5 for b in range(1, len(bins))]
print_("bin_centers: {}".format(bin_centers))
assert len(bin_centers
) == num_bins, "len(bin_centers)={} while num_bins={}!".format(
len(bin_centers), num_bins)
# ================================= #
# Compute representations
# ================================= #
z_data_file = join(save_dir, "z_data[data={}].npz".format(data_proportion))
if not exists(z_data_file):
all_z_mean = []
all_z_stddev = []
print("")
print("Compute all_z_mean, all_z_stddev and all_attrs!")
count = 0
for batch_ids in iterate_data(num_data,
10 * args.batch_size,
shuffle=False):
x = x_train[batch_ids]
z_mean, z_stddev = sess.run(model.get_output(
['z_mean', 'z_stddev']),
feed_dict={
model.is_train: False,
model.x_ph: x
})
all_z_mean.append(z_mean)
all_z_stddev.append(z_stddev)
count += len(batch_ids)
print("\rProcessed {} samples!".format(count), end="")
print()
all_z_mean = np.concatenate(all_z_mean, axis=0)
all_z_stddev = np.concatenate(all_z_stddev, axis=0)
np.savez_compressed(z_data_file,
all_z_mean=all_z_mean,
all_z_stddev=all_z_stddev)
else:
print("{} exists. Load data from file!".format(z_data_file))
with np.load(z_data_file, "r") as f:
all_z_mean = f['all_z_mean']
all_z_stddev = f['all_z_stddev']
# ================================= #
# Compute mutual information
# ================================= #
H_z = []
H_z_cond_x = []
MI_z_x = []
norm_MI_z_x = []
Q_z_cond_x = []
for i in range(args.z_dim):
print_("")
print_("Compute I(z{}, x)!".format(i))
# Q_s_cond_x
all_Q_s_cond_x = []
for batch_ids in iterate_data(len(all_z_mean),
500,
shuffle=False,
include_remaining=True):
# (batch_size, num_bins)
q_s_cond_x = normal_density(
np.expand_dims(bin_centers, axis=0),
mean=np.expand_dims(all_z_mean[batch_ids, i], axis=-1),
stddev=np.expand_dims(all_z_stddev[batch_ids, i], axis=-1))
# (batch_size, num_bins)
max_q_s_cond_x = np.max(q_s_cond_x, axis=-1)
# print("max_q_s_cond_x: {}".format(np.sort(max_q_s_cond_x)))
# (batch_size, num_bins)
deter_s_cond_x = at_bin(all_z_mean[batch_ids, i],
bins).astype(np.float32)
# (batch_size, num_bins)
Q_s_cond_x = q_s_cond_x * np.expand_dims(bin_widths, axis=0)
Q_s_cond_x = Q_s_cond_x / np.maximum(
np.sum(Q_s_cond_x, axis=1, keepdims=True), eps)
# print("sort(sum(Q_s_cond_x)) (before): {}".format(np.sort(np.sum(Q_s_cond_x, axis=-1))))
Q_s_cond_x = np.where(
np.expand_dims(np.less(max_q_s_cond_x, 1e-5), axis=-1),
deter_s_cond_x, Q_s_cond_x)
# print("sort(sum(Q_s_cond_x)) (after): {}".format(np.sort(np.sum(Q_s_cond_x, axis=-1))))
all_Q_s_cond_x.append(Q_s_cond_x)
all_Q_s_cond_x = np.concatenate(all_Q_s_cond_x, axis=0)
print_("sort(sum(all_Q_s_cond_x))[:10]: {}".format(
np.sort(np.sum(all_Q_s_cond_x, axis=-1), axis=0)[:100]))
assert np.all(all_Q_s_cond_x >= 0), "'all_Q_s_cond_x' contains negative values. " \
"sorted_all_Q_s_cond_x[:30]:\n{}!".format(
np.sort(all_Q_s_cond_x[:30], axis=None))
Q_z_cond_x.append(all_Q_s_cond_x)
H_zi_cond_x = -np.mean(np.sum(
all_Q_s_cond_x * np.log(np.maximum(all_Q_s_cond_x, eps)), axis=1),
axis=0)
# Q_s
Q_s = np.mean(all_Q_s_cond_x, axis=0)
print_("Q_s: {}".format(Q_s))
print_("sum(Q_s): {}".format(sum(Q_s)))
assert np.all(Q_s >= 0), "'Q_s' contains negative values. " \
"sorted_Q_s[:10]:\n{}!".format(np.sort(Q_s, axis=None))
Q_s = Q_s / np.sum(Q_s, axis=0)
print_("sum(Q_s) (normalized): {}".format(sum(Q_s)))
H_zi = -np.sum(Q_s * np.log(np.maximum(Q_s, eps)), axis=0)
MI_zi_x = H_zi - H_zi_cond_x
normalized_MI_zi_x = (1.0 * MI_zi_x) / (H_zi + eps)
print_("H_zi: {}".format(H_zi))
print_("H_zi_cond_x: {}".format(H_zi_cond_x))
print_("MI_zi_x: {}".format(MI_zi_x))
print_("normalized_MI_zi_x: {}".format(normalized_MI_zi_x))
H_z.append(H_zi)
H_z_cond_x.append(H_zi_cond_x)
MI_z_x.append(MI_zi_x)
norm_MI_z_x.append(normalized_MI_zi_x)
H_z = np.asarray(H_z, dtype=np.float32)
H_z_cond_x = np.asarray(H_z_cond_x, dtype=np.float32)
MI_z_x = np.asarray(MI_z_x, dtype=np.float32)
norm_MI_z_x = np.asarray(norm_MI_z_x, dtype=np.float32)
print_("")
print_("H_z: {}".format(H_z))
print_("H_z_cond_x: {}".format(H_z_cond_x))
print_("MI_z_x: {}".format(MI_z_x))
print_("norm_MI_z_x: {}".format(norm_MI_z_x))
sorted_z_comps = np.argsort(MI_z_x, axis=0)[::-1]
sorted_MI_z_x = np.take_along_axis(MI_z_x, sorted_z_comps, axis=0)
print_("sorted_MI_z_x: {}".format(sorted_MI_z_x))
print_("sorted_z_comps: {}".format(sorted_z_comps))
sorted_norm_z_comps = np.argsort(norm_MI_z_x, axis=0)[::-1]
sorted_norm_MI_z_x = np.take_along_axis(norm_MI_z_x,
sorted_norm_z_comps,
axis=0)
print_("sorted_norm_MI_z_x: {}".format(sorted_norm_MI_z_x))
print_("sorted_norm_z_comps: {}".format(sorted_norm_z_comps))
result_file = join(
save_dir, 'results[bins={},bin_limits={},data={}].npz'.format(
num_bins, bin_limits, data_proportion))
np.savez_compressed(result_file,
H_z=H_z,
H_z_cond_x=H_z_cond_x,
MI_z_x=MI_z_x,
norm_MI_z_x=norm_MI_z_x,
sorted_MI_z_x=sorted_MI_z_x,
sorted_z_comps=sorted_z_comps,
sorted_norm_MI_z_x=sorted_norm_MI_z_x,
sorted_norm_z_comps=sorted_norm_z_comps)
Q_z_cond_x = np.asarray(Q_z_cond_x, dtype=np.float32)
z_prob_file = join(
save_dir, 'z_prob[bins={},bin_limits={},data={}].npz'.format(
num_bins, bin_limits, data_proportion))
np.savez_compressed(z_prob_file, Q_z_cond_x=Q_z_cond_x)
# ================================= #
f.close()
def main(args):
np.set_printoptions(threshold=np.nan, linewidth=1000, precision=3)
# =====================================
# Preparation
# =====================================
data_file = os.path.join(RAW_DATA_DIR, "ComputerVision", "dSprites",
"dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz")
# It is already in the range [0, 1]
with np.load(data_file, encoding="latin1") as f:
x_train = f['imgs']
x_train = np.expand_dims(x_train.astype(np.float32), axis=-1)
num_train = len(x_train)
print("x_train: {}".format(num_train))
args.output_dir = os.path.join(args.output_dir, args.enc_dec_model,
args.run)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
else:
if args.force_rm_dir:
import shutil
shutil.rmtree(args.output_dir, ignore_errors=True)
print("Removed '{}'".format(args.output_dir))
else:
raise ValueError("Output directory '{}' existed. 'force_rm_dir' "
"must be set to True!".format(args.output_dir))
os.mkdir(args.output_dir)
save_args(os.path.join(args.output_dir, 'config.json'), args)
# pp.pprint(args.__dict__)
# =====================================
# Instantiate models
# =====================================
if args.enc_dec_model == "1Konny":
encoder = Encoder_1Konny(args.z_dim, stochastic=True)
decoder = Decoder_1Konny()
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support enc_dec_model='{}'!".format(
args.enc_dec_model))
model = FactorVAE([64, 64, 1],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
'Dz_tc_loss_coeff': args.Dz_tc_loss_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_list()
# SimpleParamPrinter.print_all_params_tf_slim()
loss = model.get_loss()
train_params = model.get_train_params()
opt_Dz = tf.train.AdamOptimizer(learning_rate=args.lr_Dz,
beta1=args.beta1_Dz,
beta2=args.beta2_Dz)
opt_vae = tf.train.AdamOptimizer(learning_rate=args.lr_vae,
beta1=args.beta1_vae,
beta2=args.beta2_vae)
with tf.control_dependencies(model.get_all_update_ops()):
train_op_Dz = opt_Dz.minimize(loss=loss['Dz_loss'],
var_list=train_params['Dz_loss'])
train_op_D = train_op_Dz
train_op_vae = opt_vae.minimize(loss=loss['vae_loss'],
var_list=train_params['vae_loss'])
# =====================================
# TF Graph Handler
asset_dir = make_dir_if_not_exist(os.path.join(args.output_dir, "asset"))
img_gen_dir = make_dir_if_not_exist(os.path.join(asset_dir, "img_gen"))
img_rec_dir = make_dir_if_not_exist(os.path.join(asset_dir, "img_rec"))
img_itpl_dir = make_dir_if_not_exist(os.path.join(asset_dir, "img_itpl"))
log_dir = make_dir_if_not_exist(os.path.join(args.output_dir, "log"))
train_log_file = os.path.join(log_dir, "train.log")
summary_dir = make_dir_if_not_exist(
os.path.join(args.output_dir, "summary_tf"))
model_dir = make_dir_if_not_exist(os.path.join(args.output_dir,
"model_tf"))
train_helper = SimpleTrainHelper(
log_dir=summary_dir,
save_dir=model_dir,
max_to_keep=3,
max_to_keep_best=1,
)
# =====================================
# =====================================
# Training Loop
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
train_helper.initialize(sess,
init_variables=True,
create_summary_writer=True)
Dz_fetch_keys = [
'Dz_loss', 'Dz_tc_loss', 'Dz_loss_normal', 'Dz_loss_factor',
'Dz_avg_prob_normal', 'Dz_avg_prob_factor', 'gp0_z_tc'
]
D_fetch_keys = Dz_fetch_keys
vae_fetch_keys = ['vae_loss', 'rec_x', 'kld_loss', 'tc_loss']
global_step = 0
for epoch in range(args.epochs):
for batch_ids in iterate_data(num_train, args.batch_size,
shuffle=True):
global_step += 1
x = x_train[batch_ids]
z = np.random.normal(size=[len(x), args.z_dim])
batch_ids_2 = np.random.choice(num_train,
size=len(batch_ids)).tolist()
xa = x_train[batch_ids_2]
for i in range(args.D_steps):
_, Dm = sess.run(
[train_op_D,
model.get_output(D_fetch_keys, as_dict=True)],
feed_dict={
model.is_train: True,
model.x_ph: x,
model.z_ph: z,
model.xa_ph: xa
})
for i in range(args.vae_steps):
_, VAEm = sess.run(
[
train_op_vae,
model.get_output(vae_fetch_keys, as_dict=True)
],
feed_dict={
model.is_train: True,
model.x_ph: x,
model.z_ph: z,
model.xa_ph: xa
})
if global_step % args.save_freq == 0:
train_helper.save(sess, global_step)
if global_step % args.log_freq == 0:
log_str = "\n[FactorVAE/{}/{} (dSprites)], Epoch[{}/{}], Step {}".format(
args.enc_dec_model, args.run, epoch, args.epochs, global_step) + \
"\nvae_loss: {:.4f}, Dz_loss: {:.4f}, Dz_tc_loss: {:.4f}".format(
VAEm['vae_loss'], Dm['Dz_loss'], Dm['Dz_tc_loss']) + \
"\nrec_x: {:.4f}, kld_loss: {:.4f}, tc_loss: {:.4f}".format(
VAEm['rec_x'], VAEm['kld_loss'], VAEm['tc_loss']) + \
"\nDz_loss_normal: {:.4f}, Dz_loss_factor: {:.4f}".format(
Dm['Dz_loss_normal'], Dm['Dz_loss_factor']) + \
"\nDz_avg_prob_normal: {:.4f}, Dz_avg_prob_factor: {:.4f}".format(
Dm['Dz_avg_prob_normal'], Dm['Dz_avg_prob_factor']) + \
"\ngp0_z_tc_coeff: {:.4f}, gp0_z_tc: {:.4f}".format(args.gp0_z_tc_coeff, Dm['gp0_z_tc'])
print(log_str)
with open(train_log_file, "a") as f:
f.write(log_str)
f.write("\n")
f.close()
train_helper.add_summary(
custom_tf_scalar_summary('vae_loss',
VAEm['vae_loss'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('rec_x',
VAEm['rec_x'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('kld_loss',
VAEm['kld_loss'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('tc_loss',
VAEm['tc_loss'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('Dz_tc_loss',
Dm['Dz_tc_loss'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('Dz_loss_normal',
Dm['Dz_loss_normal'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('Dz_loss_factor',
Dm['Dz_loss_factor'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('Dz_prob_normal',
Dm['Dz_avg_prob_normal'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('Dz_prob_factor',
Dm['Dz_avg_prob_factor'],
prefix='train'), global_step)
if global_step % args.viz_gen_freq == 0:
# Generate images
# ------------------------- #
z = np.random.normal(size=[64, args.z_dim])
img_file = os.path.join(img_gen_dir,
'step[%d]_gen_test.png' % global_step)
model.generate_images(
img_file,
sess,
z,
block_shape=[8, 8],
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# ------------------------- #
if global_step % args.viz_rec_freq == 0:
# Reconstruct images
# ------------------------- #
x = x_train[np.random.choice(num_train, size=64,
replace=False)]
img_file = os.path.join(img_rec_dir,
'step[%d]_rec_test.png' % global_step)
model.reconstruct_images(
img_file,
sess,
x,
block_shape=[8, 8],
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# ------------------------- #
if global_step % args.viz_itpl_freq == 0:
# Interpolate images
# ------------------------- #
x1 = x_train[np.random.choice(num_train,
size=12,
replace=False)]
x2 = x_train[np.random.choice(num_train,
size=12,
replace=False)]
img_file = os.path.join(img_itpl_dir,
'step[%d]_itpl_test.png' % global_step)
model.interpolate_images(
img_file,
sess,
x1,
x2,
num_itpl_points=12,
batch_on_row=True,
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# ------------------------- #
# Last save
train_helper.save(sess, global_step)
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
celebA_loader = TFCelebAWithAttrLoader(root_dir=args.celebA_root_dir)
img_height, img_width = args.celebA_resize_size, args.celebA_resize_size
celebA_loader.build_transformation_flow_tf(
*celebA_loader.get_transform_fns("1Konny",
resize_size=args.celebA_resize_size))
# =====================================
# Instantiate model
# =====================================
if args.activation == "relu":
activation = tf.nn.relu
elif args.activation == "leaky_relu":
activation = tf.nn.leaky_relu
else:
raise ValueError("Do not support '{}' activation!".format(
args.activation))
if args.enc_dec_model == "1Konny":
# assert args.z_dim == 65, "For 1Konny, z_dim must be 65. Found {}!".format(args.z_dim)
encoder = Encoder_1Konny(args.z_dim,
stochastic=True,
activation=activation)
decoder = Decoder_1Konny([img_height, img_width, 3],
activation=activation,
output_activation=tf.nn.sigmoid)
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support encoder/decoder model '{}'!".format(
args.enc_dec_model))
model = FactorVAE([img_height, img_width, 3],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
# save_dir = remove_dir_if_exist(join(args.save_dir, "FactorVAE_{}".format(args.run)), ask_4_permission=False)
# save_dir = make_dir_if_not_exist(save_dir)
save_dir = make_dir_if_not_exist(
join(args.save_dir, "FactorVAE_{}".format(args.run)))
# =====================================
np.set_printoptions(threshold=np.nan,
linewidth=1000,
precision=3,
suppress=True)
num_bins = args.num_bins
bin_limits = tuple([float(s) for s in args.bin_limits.split(";")])
data_proportion = args.data_proportion
num_data = int(data_proportion * celebA_loader.num_train_data)
top_k = args.top_k
eps = 1e-8
# file
f = open(join(
save_dir, 'log[bins={},bin_limits={},data={}].txt'.format(
num_bins, bin_limits, data_proportion)),
mode='w')
# print function
print_ = functools.partial(print_both, file=f)
print_("num_bins: {}".format(num_bins))
print_("bin_limits: {}".format(bin_limits))
print_("data_proportion: {}".format(data_proportion))
print_("top_k: {}".format(top_k))
# Compute bins
# ================================= #
print_("")
print_("bin_limits: {}".format(bin_limits))
assert len(bin_limits) == 2 and bin_limits[0] < bin_limits[
1], "bin_limits={}".format(bin_limits)
bins = np.linspace(bin_limits[0],
bin_limits[1],
num_bins + 1,
endpoint=True)
print_("bins: {}".format(bins))
assert len(bins) == num_bins + 1
bin_widths = [bins[b] - bins[b - 1] for b in range(1, len(bins))]
print_("bin_widths: {}".format(bin_widths))
assert len(bin_widths
) == num_bins, "len(bin_widths)={} while num_bins={}!".format(
len(bin_widths), num_bins)
assert np.all(np.greater(bin_widths,
0)), "bin_widths: {}".format(bin_widths)
bin_centers = [(bins[b] + bins[b - 1]) * 0.5 for b in range(1, len(bins))]
print_("bin_centers: {}".format(bin_centers))
assert len(bin_centers
) == num_bins, "len(bin_centers)={} while num_bins={}!".format(
len(bin_centers), num_bins)
# ================================= #
# Compute representations
# ================================= #
z_data_file = join(args.informativeness_metrics_dir,
"FactorVAE_{}".format(args.run),
"z_data[data={}].npz".format(data_proportion))
with np.load(z_data_file, "r") as f:
all_z_mean = f['all_z_mean']
all_z_stddev = f['all_z_stddev']
print_("")
print_("all_z_mean.shape: {}".format(all_z_mean.shape))
print_("all_z_stddev.shape: {}".format(all_z_stddev.shape))
# ================================= #
# Compute the mutual information
# ================================= #
mi_file = join(
args.informativeness_metrics_dir, "FactorVAE_{}".format(args.run),
'results[bins={},bin_limits={},data={}].npz'.format(
num_bins, bin_limits, data_proportion))
with np.load(mi_file, "r") as f:
sorted_MI_z_x = f['sorted_MI_z_x']
sorted_z_ids = f['sorted_z_comps']
H_z = f['H_z']
if top_k > 0:
top_MI = sorted_MI_z_x[:top_k]
top_z_ids = sorted_z_ids[:top_k]
bot_MI = sorted_MI_z_x[-top_k:]
bot_z_ids = sorted_z_ids[-top_k:]
top_bot_MI = np.concatenate([top_MI, bot_MI], axis=0)
top_bot_z_ids = np.concatenate([top_z_ids, bot_z_ids], axis=0)
print_("top MI: {}".format(top_MI))
print_("top_z_ids: {}".format(top_z_ids))
print_("bot MI: {}".format(bot_MI))
print_("bot_z_ids: {}".format(bot_z_ids))
else:
top_bot_MI = sorted_MI_z_x
top_bot_z_ids = sorted_z_ids
# ================================= #
H_z1z2_mean_mat = np.full(
[len(top_bot_z_ids), len(top_bot_z_ids)], -1, dtype=np.float32)
MI_z1z2_mean_mat = np.full(
[len(top_bot_z_ids), len(top_bot_z_ids)], -1, dtype=np.float32)
H_z1z2_mean = []
MI_z1z2_mean = []
z1z2_ids = []
# Compute the mutual information
# ================================= #
for i in range(len(top_bot_z_ids)):
z_idx1 = top_bot_z_ids[i]
H_s1 = H_z[z_idx1]
for j in range(i + 1, len(top_bot_z_ids)):
z_idx2 = top_bot_z_ids[j]
H_s2 = H_z[z_idx2]
print_("")
print_("Compute MI(z{}_mean, z{}_mean)!".format(z_idx1, z_idx2))
s1s2_mean_counter = np.zeros([num_bins, num_bins], dtype=np.int32)
for batch_ids in iterate_data(len(all_z_mean),
100,
shuffle=False,
include_remaining=True):
s1 = at_bin(all_z_mean[batch_ids, z_idx1], bins, one_hot=False)
s2 = at_bin(all_z_mean[batch_ids, z_idx2], bins, one_hot=False)
for s1_, s2_ in zip(s1, s2):
s1s2_mean_counter[s1_, s2_] += 1
# I(s1, s2) = Q(s1, s2) * (log Q(s1, s2) - log Q(s1) log Q(s2))
# ---------------------------------- #
Q_s1s2_mean = (s1s2_mean_counter *
1.0) / np.sum(s1s2_mean_counter).astype(np.float32)
log_Q_s1s2_mean = np.log(np.maximum(Q_s1s2_mean, eps))
H_s1s2_mean = -np.sum(Q_s1s2_mean * log_Q_s1s2_mean)
MI_s1s2_mean = H_s1 + H_s2 - H_s1s2_mean
print_("H_s1: {}".format(H_s1))
print_("H_s2: {}".format(H_s2))
print_("H_s1s2_mean: {}".format(H_s1s2_mean))
print_("MI_s1s2_mean: {}".format(MI_s1s2_mean))
H_z1z2_mean.append(H_s1s2_mean)
MI_z1z2_mean.append(MI_s1s2_mean)
z1z2_ids.append((z_idx1, z_idx2))
H_z1z2_mean_mat[i, j] = H_s1s2_mean
H_z1z2_mean_mat[j, i] = H_s1s2_mean
MI_z1z2_mean_mat[i, j] = MI_s1s2_mean
MI_z1z2_mean_mat[j, i] = MI_s1s2_mean
H_z1z2_mean = np.asarray(H_z1z2_mean, dtype=np.float32)
MI_z1z2_mean = np.asarray(MI_z1z2_mean, dtype=np.float32)
z1z2_ids = np.asarray(z1z2_ids, dtype=np.int32)
result_file = join(
save_dir, "results[bins={},bin_limits={},data={},k={}].npz".format(
num_bins, bin_limits, data_proportion, top_k))
results = {
'H_z1z2_mean': H_z1z2_mean,
'MI_z1z2_mean': MI_z1z2_mean,
'H_z1z2_mean_mat': H_z1z2_mean_mat,
'MI_z1z2_mean_mat': MI_z1z2_mean_mat,
'z1z2_ids': z1z2_ids,
}
np.savez_compressed(result_file, **results)
# ================================= #
f.close()
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
celebA_loader = TFCelebAWithAttrLoader(root_dir=args.celebA_root_dir)
img_height, img_width = args.celebA_resize_size, args.celebA_resize_size
celebA_loader.build_transformation_flow_tf(
*celebA_loader.get_transform_fns("1Konny", resize_size=args.celebA_resize_size))
# =====================================
# Instantiate model
# =====================================
if args.activation == "relu":
activation = tf.nn.relu
elif args.activation == "leaky_relu":
activation = tf.nn.leaky_relu
else:
raise ValueError("Do not support '{}' activation!".format(args.activation))
if args.enc_dec_model == "1Konny":
# assert args.z_dim == 65, "For 1Konny, z_dim must be 65. Found {}!".format(args.z_dim)
encoder = Encoder_1Konny(args.z_dim, stochastic=True, activation=activation)
decoder = Decoder_1Konny([img_height, img_width, 3], activation=activation,
output_activation=tf.nn.sigmoid)
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support encoder/decoder model '{}'!".format(args.enc_dec_model))
model = FactorVAE([img_height, img_width, 3], args.z_dim,
encoder=encoder, decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True, gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
# save_dir = remove_dir_if_exist(join(args.save_dir, "FactorVAE_{}".format(args.run)), ask_4_permission=True)
# save_dir = make_dir_if_not_exist(save_dir)
save_dir = make_dir_if_not_exist(join(args.save_dir, "FactorVAE_{}".format(args.run)))
# =====================================
np.set_printoptions(threshold=np.nan, linewidth=1000, precision=4, suppress=True)
num_bins = args.num_bins
bin_limits = tuple([float(s) for s in args.bin_limits.split(";")])
data_proportion = args.data_proportion
num_data = int(data_proportion * celebA_loader.num_train_data)
eps = 1e-8
# file
f = open(join(save_dir, 'log[bins={},bin_limits={},data={}].txt'.
format(num_bins, bin_limits, data_proportion)), mode='w')
# print function
print_ = functools.partial(print_both, file=f)
print_("num_bins: {}".format(num_bins))
print_("bin_limits: {}".format(bin_limits))
print_("data_proportion: {}".format(data_proportion))
# Compute representations
# ================================= #
z_data_file = join(save_dir, "z_data[data={}].npz".format(data_proportion))
if not exists(z_data_file):
all_z_mean = []
all_z_stddev = []
print("")
print("Compute all_z_mean and all_z_stddev!")
count = 0
for batch_ids in iterate_data(num_data, 10 * args.batch_size, shuffle=False):
x = celebA_loader.sample_images_from_dataset(sess, 'train', batch_ids)
z_mean, z_stddev = sess.run(
model.get_output(['z_mean', 'z_stddev']),
feed_dict={model.is_train: False, model.x_ph: x})
all_z_mean.append(z_mean)
all_z_stddev.append(z_stddev)
count += len(batch_ids)
print("\rProcessed {} samples!".format(count), end="")
print()
all_z_mean = np.concatenate(all_z_mean, axis=0)
all_z_stddev = np.concatenate(all_z_stddev, axis=0)
np.savez_compressed(z_data_file, all_z_mean=all_z_mean, all_z_stddev=all_z_stddev)
else:
print("{} exists. Load data from file!".format(z_data_file))
with np.load(z_data_file, "r") as f:
all_z_mean = f['all_z_mean']
all_z_stddev = f['all_z_stddev']
print_("")
print_("all_z_mean.shape: {}".format(all_z_mean.shape))
print_("all_z_stddev.shape: {}".format(all_z_stddev.shape))
# ================================= #
# Compute bins
# ================================= #
print_("")
print_("bin_limits: {}".format(bin_limits))
assert len(bin_limits) == 2 and bin_limits[0] < bin_limits[1], "bin_limits={}".format(bin_limits)
bins = np.linspace(bin_limits[0], bin_limits[1], num_bins + 1, endpoint=True)
print_("bins: {}".format(bins))
assert len(bins) == num_bins + 1
bin_widths = [bins[b] - bins[b - 1] for b in range(1, len(bins))]
print_("bin_widths: {}".format(bin_widths))
assert len(bin_widths) == num_bins, "len(bin_widths)={} while num_bins={}!".format(len(bin_widths), num_bins)
assert np.all(np.greater(bin_widths, 0)), "bin_widths: {}".format(bin_widths)
bin_centers = [(bins[b] + bins[b - 1]) * 0.5 for b in range(1, len(bins))]
print_("bin_centers: {}".format(bin_centers))
assert len(bin_centers) == num_bins, "len(bin_centers)={} while num_bins={}!".format(len(bin_centers), num_bins)
# ================================= #
# Compute mutual information
# ================================= #
H_z = []
H_z_cond_x = []
MI_z_x = []
norm_MI_z_x = []
Q_z_cond_x = []
for i in range(args.z_dim):
print_("")
print_("Compute I(z{}, x)!".format(i))
# Q_s_cond_x
all_Q_s_cond_x = []
for batch_ids in iterate_data(len(all_z_mean), 500, shuffle=False, include_remaining=True):
# (batch_size, num_bins)
q_s_cond_x = normal_density(np.expand_dims(bin_centers, axis=0),
mean=np.expand_dims(all_z_mean[batch_ids, i], axis=-1),
stddev=np.expand_dims(all_z_stddev[batch_ids, i], axis=-1))
# (batch_size, num_bins)
max_q_s_cond_x = np.max(q_s_cond_x, axis=-1)
# print("\nmax_q_s_cond_x: {}".format(np.sort(max_q_s_cond_x)))
# (batch_size, num_bins)
deter_s_cond_x = at_bin(all_z_mean[batch_ids, i], bins).astype(np.float32)
# (batch_size, num_bins)
Q_s_cond_x = q_s_cond_x * np.expand_dims(bin_widths, axis=0)
Q_s_cond_x = Q_s_cond_x / np.maximum(np.sum(Q_s_cond_x, axis=1, keepdims=True), eps)
# print("sort(sum(Q_s_cond_x)) (before): {}".format(np.sort(np.sum(Q_s_cond_x, axis=-1))))
Q_s_cond_x = np.where(np.expand_dims(np.less(max_q_s_cond_x, 1e-5), axis=-1),
deter_s_cond_x, Q_s_cond_x)
# print("sort(sum(Q_s_cond_x)) (after): {}".format(np.sort(np.sum(Q_s_cond_x, axis=-1))))
all_Q_s_cond_x.append(Q_s_cond_x)
all_Q_s_cond_x = np.concatenate(all_Q_s_cond_x, axis=0)
print_("sort(sum(all_Q_s_cond_x))[:10]: {}".format(
np.sort(np.sum(all_Q_s_cond_x, axis=-1), axis=0)[:100]))
assert np.all(all_Q_s_cond_x >= 0), "'all_Q_s_cond_x' contains negative values. " \
"sorted_all_Q_s_cond_x[:30]:\n{}!".format(np.sort(all_Q_s_cond_x[:30], axis=None))
Q_z_cond_x.append(all_Q_s_cond_x)
H_zi_cond_x = -np.mean(np.sum(all_Q_s_cond_x * np.log(np.maximum(all_Q_s_cond_x, eps)), axis=1), axis=0)
# Q_s
Q_s = np.mean(all_Q_s_cond_x, axis=0)
print_("Q_s: {}".format(Q_s))
print_("sum(Q_s): {}".format(sum(Q_s)))
assert np.all(Q_s >= 0), "'Q_s' contains negative values. " \
"sorted_Q_s[:10]:\n{}!".format(np.sort(Q_s, axis=None))
Q_s = Q_s / np.sum(Q_s, axis=0)
print_("sum(Q_s) (normalized): {}".format(sum(Q_s)))
H_zi = -np.sum(Q_s * np.log(np.maximum(Q_s, eps)), axis=0)
MI_zi_x = H_zi - H_zi_cond_x
normalized_MI_zi_x = (1.0 * MI_zi_x) / (H_zi + eps)
print_("H_zi: {}".format(H_zi))
print_("H_zi_cond_x: {}".format(H_zi_cond_x))
print_("MI_zi_x: {}".format(MI_zi_x))
print_("normalized_MI_zi_x: {}".format(normalized_MI_zi_x))
H_z.append(H_zi)
H_z_cond_x.append(H_zi_cond_x)
MI_z_x.append(MI_zi_x)
norm_MI_z_x.append(normalized_MI_zi_x)
H_z = np.asarray(H_z, dtype=np.float32)
H_z_cond_x = np.asarray(H_z_cond_x, dtype=np.float32)
MI_z_x = np.asarray(MI_z_x, dtype=np.float32)
norm_MI_z_x = np.asarray(norm_MI_z_x, dtype=np.float32)
print_("")
print_("H_z: {}".format(H_z))
print_("H_z_cond_x: {}".format(H_z_cond_x))
print_("MI_z_x: {}".format(MI_z_x))
print_("norm_MI_z_x: {}".format(norm_MI_z_x))
sorted_z_comps = np.argsort(MI_z_x, axis=0)[::-1]
sorted_MI_z_x = np.take_along_axis(MI_z_x, sorted_z_comps, axis=0)
print_("sorted_MI_z_x: {}".format(sorted_MI_z_x))
print_("sorted_z_comps: {}".format(sorted_z_comps))
sorted_norm_z_comps = np.argsort(norm_MI_z_x, axis=0)[::-1]
sorted_norm_MI_z_x = np.take_along_axis(norm_MI_z_x, sorted_norm_z_comps, axis=0)
print_("sorted_norm_MI_z_x: {}".format(sorted_norm_MI_z_x))
print_("sorted_norm_z_comps: {}".format(sorted_norm_z_comps))
result_file = join(save_dir, 'results[bins={},bin_limits={},data={}].npz'.
format(num_bins, bin_limits, data_proportion))
np.savez_compressed(result_file,
H_z=H_z, H_z_cond_x=H_z_cond_x, MI_z_x=MI_z_x, norm_MI_z_x=norm_MI_z_x,
sorted_MI_z_x=sorted_MI_z_x, sorted_z_comps=sorted_z_comps,
sorted_norm_MI_z_x=sorted_norm_MI_z_x,
sorted_norm_z_comps=sorted_norm_z_comps)
Q_z_cond_x = np.asarray(Q_z_cond_x, dtype=np.float32)
z_prob_file = join(save_dir, 'z_prob[bins={},bin_limits={},data={}].npz'.
format(num_bins, bin_limits, data_proportion))
np.savez_compressed(z_prob_file, Q_z_cond_x=Q_z_cond_x)
def main(args):
# =====================================
# Load config
# =====================================
with open(os.path.join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Preparation
# =====================================
data_file = os.path.join(RAW_DATA_DIR, "ComputerVision", "dSprites",
"dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz")
# It is already in the range [0, 1]
with np.load(data_file, encoding="latin1") as f:
x_train = f['imgs']
x_train = np.expand_dims(x_train.astype(np.float32), axis=-1)
# =====================================
# Instantiate models
# =====================================
encoder = Encoder_1Konny(args.z_dim, stochastic=True)
decoder = Decoder_1Konny()
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
model = FactorVAE([64, 64, 1],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_list()
# =====================================
# TF Graph Handler
asset_dir = make_dir_if_not_exist(os.path.join(args.output_dir, "asset"))
img_eval = remove_dir_if_exist(os.path.join(asset_dir, "img_eval"),
ask_4_permission=False)
img_eval = make_dir_if_not_exist(img_eval)
img_x_rec = make_dir_if_not_exist(os.path.join(img_eval, "x_rec"))
img_z_rand_2_traversal = make_dir_if_not_exist(
os.path.join(img_eval, "z_rand_2_traversal"))
img_z_cond_all_traversal = make_dir_if_not_exist(
os.path.join(img_eval, "z_cond_all_traversal"))
img_z_cond_1_traversal = make_dir_if_not_exist(
os.path.join(img_eval, "z_cond_1_traversal"))
img_z_corr = make_dir_if_not_exist(os.path.join(img_eval, "z_corr"))
img_z_dist = make_dir_if_not_exist(os.path.join(img_eval, "z_dist"))
img_z_stat_dist = make_dir_if_not_exist(
os.path.join(img_eval, "z_stat_dist"))
img_rec_error_dist = make_dir_if_not_exist(
os.path.join(img_eval, "rec_error_dist"))
model_dir = make_dir_if_not_exist(os.path.join(args.output_dir,
"model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# =====================================
# =====================================
# Training Loop
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
# Load model
train_helper.load(sess, load_step=args.load_step)
#'''
# Reconstruction
# ======================================= #
seed = 389
x = x_train[np.arange(seed, seed + 64)]
img_file = os.path.join(img_x_rec, 'x_rec_train.png')
model.reconstruct_images(img_file,
sess,
x,
block_shape=[8, 8],
batch_size=-1,
dec_output_2_img_func=binary_float_to_uint8)
# ======================================= #
# z random/conditional traversal
# ======================================= #
# Plot z cont with z cont
z_zero = np.zeros([args.z_dim], dtype=np.float32)
z_rand = np.random.randn(args.z_dim)
z_start, z_stop = -4, 4
num_points = 8
for i in range(args.z_dim):
for j in range(i + 1, args.z_dim):
print("Plot random 2 comps z traverse with {} and {} components!".
format(i, j))
img_file = os.path.join(img_z_rand_2_traversal,
'z[{},{},zero].png'.format(i, j))
model.rand_2_latents_traverse(
img_file,
sess,
default_z=z_zero,
z_comp1=i,
start1=z_start,
stop1=z_stop,
num_points1=num_points,
z_comp2=j,
start2=z_start,
stop2=z_stop,
num_points2=num_points,
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
img_file = os.path.join(img_z_rand_2_traversal,
'z[{},{},rand].png'.format(i, j))
model.rand_2_latents_traverse(
img_file,
sess,
default_z=z_rand,
z_comp1=i,
start1=z_start,
stop1=z_stop,
num_points1=num_points,
z_comp2=j,
start2=z_stop,
stop2=z_stop,
num_points2=num_points,
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
seed = 389
z_start, z_stop = -4, 4
num_itpl_points = 8
for n in range(seed, seed + 30):
print("Plot conditional all comps z traverse with test sample {}!".
format(n))
x = x_train[n]
img_file = os.path.join(img_z_cond_all_traversal,
'x_train{}.png'.format(n))
model.cond_all_latents_traverse(
img_file,
sess,
x,
start=z_start,
stop=z_stop,
num_itpl_points=num_itpl_points,
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
seed = 64
z_start, z_stop = -4, 4
num_itpl_points = 8
print("Plot conditional 1 comp z traverse!")
for i in range(args.z_dim):
x = x_train[seed:seed + 64]
img_file = os.path.join(
img_z_cond_1_traversal,
'x_train[{},{}]_z{}.png'.format(seed, seed + 64, i))
model.cond_1_latent_traverse(
img_file,
sess,
x,
z_comp=i,
start=z_start,
stop=z_stop,
num_itpl_points=num_itpl_points,
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# ======================================= #
# '''
# z correlation matrix
# ======================================= #
data = x_train
all_z = []
for batch_ids in iterate_data(len(data), args.batch_size, shuffle=False):
x = data[batch_ids]
z = model.encode(sess, x)
assert len(
z.shape) == 2 and z.shape[1] == args.z_dim, "z.shape: {}".format(
z.shape)
all_z.append(z)
all_z = np.concatenate(all_z, axis=0)
print("Start plotting!")
plot_corrmat_with_histogram(os.path.join(img_z_corr, "corr_mat.png"),
all_z)
plot_comp_dist(os.path.join(img_z_dist, 'z_{}'), all_z, x_lim=(-5, 5))
print("Done!")
# ======================================= #
# '''
# z gaussian stddev
# ======================================= #
print("\nPlot z mean and stddev!")
data = x_train
all_z_mean = []
all_z_stddev = []
for batch_ids in iterate_data(len(data), args.batch_size, shuffle=False):
x = data[batch_ids]
z_mean, z_stddev = sess.run(model.get_output(['z_mean', 'z_stddev']),
feed_dict={
model.is_train: False,
model.x_ph: x
})
all_z_mean.append(z_mean)
all_z_stddev.append(z_stddev)
all_z_mean = np.concatenate(all_z_mean, axis=0)
all_z_stddev = np.concatenate(all_z_stddev, axis=0)
plot_comp_dist(os.path.join(img_z_stat_dist, 'z_mean_{}.png'),
all_z_mean,
x_lim=(-5, 5))
plot_comp_dist(os.path.join(img_z_stat_dist, 'z_stddev_{}.png'),
all_z_stddev,
x_lim=(0, 3))
# ======================================= #
# '''
# Decoder sensitivity
# ======================================= #
z_start = -3
z_stop = 3
for i in range(args.z_dim):
print("\nPlot rec error distribution for z component {}!".format(i))
all_z1 = np.array(all_z, copy=True, dtype=np.float32)
all_z2 = np.array(all_z, copy=True, dtype=np.float32)
all_z1[:, i] = z_start
all_z2[:, i] = z_stop
all_x_rec1 = []
all_x_rec2 = []
for batch_ids in iterate_data(len(x_train),
args.batch_size,
shuffle=False):
z1 = all_z1[batch_ids]
z2 = all_z2[batch_ids]
x1 = model.decode(sess, z1)
x2 = model.decode(sess, z2)
all_x_rec1.append(x1)
all_x_rec2.append(x2)
all_x_rec1 = np.concatenate(all_x_rec1, axis=0)
all_x_rec2 = np.concatenate(all_x_rec2, axis=0)
rec_errors = np.sum(np.reshape((all_x_rec1 - all_x_rec2)**2,
[len(x_train), 28 * 28]),
axis=1)
plot_comp_dist(
os.path.join(
img_rec_error_dist,
'rec_error[zi={},{},{}].png'.format(i, z_start, z_stop)),
rec_errors)
def main(args):
# =====================================
# Load config
# =====================================
with open(os.path.join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Preparation
# =====================================
celebA_loader = TFCelebALoader(root_dir=args.celebA_root_dir)
num_train = celebA_loader.num_train_data
num_test = celebA_loader.num_test_data
img_height, img_width = args.celebA_resize_size, args.celebA_resize_size
celebA_loader.build_transformation_flow_tf(
*celebA_loader.get_transform_fns("1Konny",
resize_size=args.celebA_resize_size))
# =====================================
# Instantiate models
# =====================================
# Only use activation for encoder and decoder
if args.activation == "relu":
activation = tf.nn.relu
elif args.activation == "leaky_relu":
activation = tf.nn.leaky_relu
else:
raise ValueError("Do not support '{}' activation!".format(
args.activation))
if args.enc_dec_model == "1Konny":
# assert args.z_dim == 65, "For 1Konny, z_dim must be 65. Found {}!".format(args.z_dim)
encoder = Encoder_1Konny(args.z_dim,
stochastic=True,
activation=activation)
decoder = Decoder_1Konny([img_height, img_width, 3],
activation=activation,
output_activation=tf.nn.sigmoid)
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
elif args.enc_dec_model == "my":
# assert args.z_dim == 150, "For my, z_dim must be 150. Found {}!".format(args.z_dim)
encoder = Encoder_My(args.z_dim,
stochastic=True,
activation=activation)
decoder = Decoder_My([img_height, img_width, 3],
activation=activation,
output_activation=tf.nn.sigmoid)
disc_z = DiscriminatorZ_My(num_outputs=2)
else:
raise ValueError("Do not support encoder/decoder model '{}'!".format(
args.enc_dec_model))
model = FactorVAE([img_height, img_width, 3],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
# 'Dz_tc_loss': args.Dz_tc_loss,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_list()
# =====================================
# TF Graph Handler
asset_dir = make_dir_if_not_exist(os.path.join(args.output_dir, "asset"))
img_eval = remove_dir_if_exist(os.path.join(asset_dir, "img_eval"),
ask_4_permission=False)
img_eval = make_dir_if_not_exist(img_eval)
img_x_gen = make_dir_if_not_exist(os.path.join(img_eval, "x_gen"))
img_x_rec = make_dir_if_not_exist(os.path.join(img_eval, "x_rec"))
img_z_rand_2_traversal = make_dir_if_not_exist(
os.path.join(img_eval, "z_rand_2_traversal"))
img_z_cond_all_traversal = make_dir_if_not_exist(
os.path.join(img_eval, "z_cond_all_traversal"))
img_z_cond_1_traversal = make_dir_if_not_exist(
os.path.join(img_eval, "z_cond_1_traversal"))
img_z_corr = make_dir_if_not_exist(os.path.join(img_eval, "z_corr"))
img_z_dist = make_dir_if_not_exist(os.path.join(img_eval, "z_dist"))
img_z_stat_dist = make_dir_if_not_exist(
os.path.join(img_eval, "z_stat_dist"))
# img_rec_error_dist = make_dir_if_not_exist(os.path.join(img_eval, "rec_error_dist"))
model_dir = make_dir_if_not_exist(os.path.join(args.output_dir,
"model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# =====================================
# =====================================
# Training Loop
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
# Load model
train_helper.load(sess, load_step=args.load_step)
# '''
# Generation
# ======================================= #
z = np.random.randn(64, args.z_dim)
img_file = os.path.join(img_x_gen, 'x_gen_test.png')
model.generate_images(img_file,
sess,
z,
block_shape=[8, 8],
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# ======================================= #
# '''
# '''
# Reconstruction
# ======================================= #
seed = 389
x = celebA_loader.sample_images_from_dataset(sess, 'test',
list(range(seed, seed + 64)))
img_file = os.path.join(img_x_rec, 'x_rec_test.png')
model.reconstruct_images(img_file,
sess,
x,
block_shape=[8, 8],
batch_size=-1,
dec_output_2_img_func=binary_float_to_uint8)
# ======================================= #
# '''
# '''
# z random traversal
# ======================================= #
if args.z_dim <= 5:
print("z_dim = {}, perform random traversal!".format(args.z_dim))
# Plot z cont with z cont
z_zero = np.zeros([args.z_dim], dtype=np.float32)
z_rand = np.random.randn(args.z_dim)
z_start, z_stop = -4, 4
num_points = 8
for i in range(args.z_dim):
for j in range(i + 1, args.z_dim):
print(
"Plot random 2 comps z traverse with {} and {} components!"
.format(i, j))
img_file = os.path.join(img_z_rand_2_traversal,
'z[{},{},zero].png'.format(i, j))
model.rand_2_latents_traverse(
img_file,
sess,
default_z=z_zero,
z_comp1=i,
start1=z_start,
stop1=z_stop,
num_points1=num_points,
z_comp2=j,
start2=z_start,
stop2=z_stop,
num_points2=num_points,
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
img_file = os.path.join(img_z_rand_2_traversal,
'z[{},{},rand].png'.format(i, j))
model.rand_2_latents_traverse(
img_file,
sess,
default_z=z_rand,
z_comp1=i,
start1=z_start,
stop1=z_stop,
num_points1=num_points,
z_comp2=j,
start2=z_stop,
stop2=z_stop,
num_points2=num_points,
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# ======================================= #
# '''
# z conditional traversal (all features + one feature)
# ======================================= #
seed = 389
num_samples = 30
data = celebA_loader.sample_images_from_dataset(
sess, 'train', list(range(seed, seed + num_samples)))
z_start, z_stop = -4, 4
num_itpl_points = 8
for n in range(num_samples):
print("Plot conditional all comps z traverse with test sample {}!".
format(n))
img_file = os.path.join(img_z_cond_all_traversal,
'x_train{}.png'.format(n))
model.cond_all_latents_traverse(
img_file,
sess,
data[n],
start=z_start,
stop=z_stop,
num_itpl_points=num_itpl_points,
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
z_start, z_stop = -4, 4
num_itpl_points = 8
for i in range(args.z_dim):
print("Plot conditional z traverse with comp {}!".format(i))
img_file = os.path.join(
img_z_cond_1_traversal,
'x_train[{},{}]_z{}.png'.format(seed, seed + num_samples, i))
model.cond_1_latent_traverse(
img_file,
sess,
data,
z_comp=i,
start=z_start,
stop=z_stop,
num_itpl_points=num_itpl_points,
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# ======================================= #
# '''
# '''
# z correlation matrix
# ======================================= #
all_z = []
for batch_ids in iterate_data(num_train, args.batch_size, shuffle=False):
x = celebA_loader.sample_images_from_dataset(sess, 'train', batch_ids)
z = model.encode(sess, x)
assert len(
z.shape) == 2 and z.shape[1] == args.z_dim, "z.shape: {}".format(
z.shape)
all_z.append(z)
all_z = np.concatenate(all_z, axis=0)
print("Start plotting!")
plot_corrmat_with_histogram(os.path.join(img_z_corr, "corr_mat.png"),
all_z)
plot_comp_dist(os.path.join(img_z_dist, 'z_{}'), all_z, x_lim=(-5, 5))
print("Done!")
# ======================================= #
# '''
# '''
# z gaussian stddev
# ======================================= #
print("\nPlot z mean and stddev!")
all_z_mean = []
all_z_stddev = []
for batch_ids in iterate_data(num_train, args.batch_size, shuffle=False):
x = celebA_loader.sample_images_from_dataset(sess, 'train', batch_ids)
z_mean, z_stddev = sess.run(model.get_output(['z_mean', 'z_stddev']),
feed_dict={
model.is_train: False,
model.x_ph: x
})
all_z_mean.append(z_mean)
all_z_stddev.append(z_stddev)
all_z_mean = np.concatenate(all_z_mean, axis=0)
all_z_stddev = np.concatenate(all_z_stddev, axis=0)
plot_comp_dist(os.path.join(img_z_stat_dist, 'z_mean_{}.png'),
all_z_mean,
x_lim=(-5, 5))
plot_comp_dist(os.path.join(img_z_stat_dist, 'z_stddev_{}.png'),
all_z_stddev,
x_lim=(-0.5, 3))
def main(args):
# =====================================
# Preparation
# =====================================
celebA_loader = TFCelebALoader(root_dir=args.celebA_root_dir)
num_train = celebA_loader.num_train_data
num_test = celebA_loader.num_test_data
img_height, img_width = args.celebA_resize_size, args.celebA_resize_size
celebA_loader.build_transformation_flow_tf(
*celebA_loader.get_transform_fns("1Konny",
resize_size=args.celebA_resize_size))
args.output_dir = os.path.join(args.output_dir, args.enc_dec_model,
args.run)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
else:
if args.force_rm_dir:
import shutil
shutil.rmtree(args.output_dir, ignore_errors=True)
print("Removed '{}'".format(args.output_dir))
else:
raise ValueError("Output directory '{}' existed. 'force_rm_dir' "
"must be set to True!".format(args.output_dir))
os.mkdir(args.output_dir)
save_args(os.path.join(args.output_dir, 'config.json'), args)
# pp.pprint(args.__dict__)
# =====================================
# Instantiate models
# =====================================
# Only use activation for encoder and decoder
if args.activation == "relu":
activation = tf.nn.relu
elif args.activation == "leaky_relu":
activation = tf.nn.leaky_relu
else:
raise ValueError("Do not support '{}' activation!".format(
args.activation))
if args.enc_dec_model == "1Konny":
# assert args.z_dim == 65, "For 1Konny, z_dim must be 65. Found {}!".format(args.z_dim)
encoder = Encoder_1Konny(args.z_dim,
stochastic=True,
activation=activation)
decoder = Decoder_1Konny([img_height, img_width, 3],
activation=activation,
output_activation=tf.nn.sigmoid)
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support encoder/decoder model '{}'!".format(
args.enc_dec_model))
model = FactorVAE([img_height, img_width, 3],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'Dz_tc_loss': args.Dz_tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_list()
loss = model.get_loss()
train_params = model.get_train_params()
opt_Dz = tf.train.AdamOptimizer(learning_rate=args.lr_Dz,
beta1=args.beta1_Dz,
beta2=args.beta2_Dz)
opt_vae = tf.train.AdamOptimizer(learning_rate=args.lr_vae,
beta1=args.beta1_vae,
beta2=args.beta2_vae)
with tf.control_dependencies(model.get_all_update_ops()):
train_op_Dz = opt_Dz.minimize(loss=loss['Dz_loss'],
var_list=train_params['Dz_loss'])
train_op_D = train_op_Dz
train_op_vae = opt_vae.minimize(loss=loss['vae_loss'],
var_list=train_params['vae_loss'])
# =====================================
# TF Graph Handler
asset_dir = make_dir_if_not_exist(os.path.join(args.output_dir, "asset"))
img_gen_dir = make_dir_if_not_exist(os.path.join(asset_dir, "img_gen"))
img_rec_dir = make_dir_if_not_exist(os.path.join(asset_dir, "img_rec"))
img_itpl_dir = make_dir_if_not_exist(os.path.join(asset_dir, "img_itpl"))
log_dir = make_dir_if_not_exist(os.path.join(args.output_dir, "log"))
train_log_file = os.path.join(log_dir, "train.log")
summary_dir = make_dir_if_not_exist(
os.path.join(args.output_dir, "summary_tf"))
model_dir = make_dir_if_not_exist(os.path.join(args.output_dir,
"model_tf"))
train_helper = SimpleTrainHelper(
log_dir=summary_dir,
save_dir=model_dir,
max_to_keep=3,
max_to_keep_best=1,
)
# =====================================
# =====================================
# Training Loop
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
train_helper.initialize(sess,
init_variables=True,
create_summary_writer=True)
Dz_fetch_keys = [
'Dz_loss', 'Dz_tc_loss', 'Dz_loss_normal', 'Dz_loss_factor',
'Dz_avg_prob_normal', 'Dz_avg_prob_factor', 'gp0_z_tc'
]
D_fetch_keys = Dz_fetch_keys
vae_fetch_keys = ['vae_loss', 'rec_x', 'kld_loss', 'tc_loss']
train_sampler = ContinuousIndexSampler(num_train,
args.batch_size,
shuffle=True)
import math
num_batch_per_epochs = int(math.ceil(num_train / args.batch_size))
global_step = 0
for epoch in range(args.epochs):
for _ in range(num_batch_per_epochs):
global_step += 1
batch_ids = train_sampler.sample_ids()
x = celebA_loader.sample_images_from_dataset(
sess, 'train', batch_ids)
z = np.random.randn(len(x), args.z_dim)
batch_ids_2 = np.random.choice(num_train, size=len(batch_ids))
xa = celebA_loader.sample_images_from_dataset(
sess, 'train', batch_ids_2)
for i in range(args.D_steps):
_, Dm = sess.run(
[train_op_D,
model.get_output(D_fetch_keys, as_dict=True)],
feed_dict={
model.is_train: True,
model.x_ph: x,
model.z_ph: z,
model.xa_ph: xa
})
for i in range(args.vae_steps):
_, VAEm = sess.run(
[
train_op_vae,
model.get_output(vae_fetch_keys, as_dict=True)
],
feed_dict={
model.is_train: True,
model.x_ph: x,
model.z_ph: z,
model.xa_ph: xa
})
if global_step % args.save_freq == 0:
train_helper.save(sess, global_step)
if global_step % args.log_freq == 0:
log_str = "\n[FactorVAE (celebA)/{}, {}]".format(args.enc_dec_model, args.run) + \
"\nEpoch {}/{}, Step {}, vae_loss: {:.4f}, Dz_loss: {:.4f}, Dz_tc_loss: {:.4f}".format(
epoch, args.epochs, global_step, VAEm['vae_loss'], Dm['Dz_loss'], Dm['Dz_tc_loss']) + \
"\nrec_x: {:.4f}, kld_loss: {:.4f}, tc_loss: {:.4f}".format(
VAEm['rec_x'], VAEm['kld_loss'], VAEm['tc_loss']) + \
"\nDz_loss_normal: {:.4f}, Dz_loss_factor: {:.4f}".format(
Dm['Dz_loss_normal'], Dm['Dz_loss_factor']) + \
"\nDz_avg_prob_normal: {:.4f}, Dz_avg_prob_factor: {:.4f}".format(
Dm['Dz_avg_prob_normal'], Dm['Dz_avg_prob_factor']) + \
"\ngp0_z_tc_coeff: {:.4f}, gp0_z_tc: {:.4f}".format(args.gp0_z_tc_coeff, Dm['gp0_z_tc'])
print(log_str)
with open(train_log_file, "a") as f:
f.write(log_str)
f.write("\n")
f.close()
train_helper.add_summary(
custom_tf_scalar_summary('vae_loss',
VAEm['vae_loss'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('rec_x',
VAEm['rec_x'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('kld_loss',
VAEm['kld_loss'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('tc_loss',
VAEm['tc_loss'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('Dz_tc_loss',
Dm['Dz_tc_loss'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('Dz_loss_normal',
Dm['Dz_loss_normal'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('Dz_loss_factor',
Dm['Dz_loss_factor'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('Dz_prob_normal',
Dm['Dz_avg_prob_normal'],
prefix='train'), global_step)
train_helper.add_summary(
custom_tf_scalar_summary('Dz_prob_factor',
Dm['Dz_avg_prob_factor'],
prefix='train'), global_step)
if global_step % args.viz_gen_freq == 0:
# Generate images
# ------------------------- #
z = np.random.randn(64, args.z_dim)
img_file = os.path.join(img_gen_dir,
'step[%d]_gen_test.png' % global_step)
model.generate_images(
img_file,
sess,
z,
block_shape=[8, 8],
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# ------------------------- #
if global_step % args.viz_rec_freq == 0:
# Reconstruct images
# ------------------------- #
x = celebA_loader.sample_images_from_dataset(
sess, 'test',
np.random.choice(num_test, size=64, replace=False))
img_file = os.path.join(img_rec_dir,
'step[%d]_rec_test.png' % global_step)
model.reconstruct_images(
img_file,
sess,
x,
block_shape=[8, 8],
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# ------------------------- #
if global_step % args.viz_itpl_freq == 0:
# Interpolate images
# ------------------------- #
x1 = celebA_loader.sample_images_from_dataset(
sess, 'test',
np.random.choice(num_test, size=12, replace=False))
x2 = celebA_loader.sample_images_from_dataset(
sess, 'test',
np.random.choice(num_test, size=12, replace=False))
img_file = os.path.join(img_itpl_dir,
'step[%d]_itpl_test.png' % global_step)
model.interpolate_images(
img_file,
sess,
x1,
x2,
num_itpl_points=12,
batch_on_row=True,
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
# ------------------------- #
if epoch % 100 == 0:
train_helper.save_separately(
sess,
model_name="model_epoch{}".format(epoch),
global_step=global_step)
# Last save
train_helper.save(sess, global_step)
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
celebA_loader = TFCelebALoader(root_dir=args.celebA_root_dir)
img_height, img_width = args.celebA_resize_size, args.celebA_resize_size
celebA_loader.build_transformation_flow_tf(
*celebA_loader.get_transform_fns("1Konny", resize_size=args.celebA_resize_size))
# =====================================
# Instantiate model
# =====================================
if args.activation == "relu":
activation = tf.nn.relu
elif args.activation == "leaky_relu":
activation = tf.nn.leaky_relu
else:
raise ValueError("Do not support '{}' activation!".format(args.activation))
if args.enc_dec_model == "1Konny":
assert args.z_dim == 65, "For 1Konny, z_dim must be 65. Found {}!".format(args.z_dim)
encoder = Encoder_1Konny(args.z_dim, stochastic=True, activation=activation)
decoder = Decoder_1Konny([img_height, img_width, 3], activation=activation,
output_activation=tf.nn.sigmoid)
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
elif args.enc_dec_model == "my":
assert args.z_dim == 150, "For 1Konny, z_dim must be 150. Found {}!".format(args.z_dim)
encoder = Encoder_My(args.z_dim, stochastic=True, activation=activation)
decoder = Decoder_My([img_height, img_width, 3], activation=activation,
output_activation=tf.nn.sigmoid)
disc_z = DiscriminatorZ_My(num_outputs=2)
else:
raise ValueError("Do not support encoder/decoder model '{}'!".format(args.enc_dec_model))
model = FactorVAE([img_height, img_width, 3], args.z_dim,
encoder=encoder, decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True, gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
save_dir = remove_dir_if_exist(join(args.save_dir, "FactorVAE_{}".format(args.run)), ask_4_permission=False)
save_dir = make_dir_if_not_exist(save_dir)
# =====================================
# z correlation matrix
# ======================================= #
for deterministic in [True, False]:
all_z = []
for batch_ids in iterate_data(celebA_loader.num_train_data, args.batch_size, shuffle=False):
x = celebA_loader.sample_images_from_dataset(sess, 'train', batch_ids)
z = model.encode(sess, x, deterministic=deterministic)
assert len(z.shape) == 2 and z.shape[1] == args.z_dim, "z.shape: {}".format(z.shape)
all_z.append(z)
all_z = np.concatenate(all_z, axis=0)
# plot_corrmat(join(save_dir, "corr_mat[deter={}].png".format(deterministic)), all_z,
# font={'size': 14},
# subplot_adjust={'left': 0.04, 'right': 0.96, 'bottom': 0.02, 'top': 0.98},
# size_inches=(7.2, 6))
plot_corrmat_with_histogram(join(save_dir, "corr_mat_hist[deter={}].png".format(deterministic)), all_z,
font={'size': 14},
subplot_adjust={'left': 0.04, 'right': 0.96, 'bottom': 0.02, 'top': 0.98},
size_inches=(10, 3))
def main(args):
# Load config
# =========================================== #
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =========================================== #
# Load dataset
# =========================================== #
data_file = join(RAW_DATA_DIR, "ComputerVision", "dSprites",
"dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz")
# It is already in the range [0, 1]
with np.load(data_file, encoding="latin1") as f:
x_train = f['imgs']
x_train = np.reshape(x_train, [3, 6, 40, 32, 32, 64, 64, 1])
# =========================================== #
# Build model
# =========================================== #
if args.enc_dec_model == "1Konny":
encoder = Encoder_1Konny(args.z_dim, stochastic=True)
decoder = Decoder_1Konny()
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support enc_dec_model='{}'!".format(
args.enc_dec_model))
model = FactorVAE([64, 64, 1],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
'Dz_tc_loss_coeff': args.Dz_tc_loss_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =========================================== #
# Initialize session
# =========================================== #
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
train_helper.load(sess, load_step=args.load_step)
save_dir = make_dir_if_not_exist(
join(args.save_dir, "{}_{}".format(args.enc_dec_model, args.run)))
# =========================================== #
# Load result file
# =========================================== #
result_file = join(args.SEPIN_dir, "{}_{}".format(args.enc_dec_model,
args.run),
"results[num_samples={}].npz".format(args.num_samples))
results = np.load(result_file, "r")
print("results.keys: {}".format(list(results.keys())))
np.set_printoptions(threshold=np.nan,
linewidth=1000,
precision=3,
suppress=True)
# =========================================== #
# Plotting
# =========================================== #
data = [
x_train[0, 3, 20, 16, 16], x_train[1, 3, 20, 16, 16], x_train[2, 3, 20,
16, 16]
]
gt_factors = ['Shape', 'Scale', 'Rotation', 'Pos_x', 'Pos_y']
# (num_latents,)
MI_zi_x = results['MI_zi_x']
SEP_zi = results['SEP_zi']
ids_sorted = np.argsort(SEP_zi, axis=0)[::-1]
print("")
print("MI_zi_x: {}".format(MI_zi_x))
print("SEP_zi: {}".format(SEP_zi))
print("ids_sorted: {}".format(ids_sorted))
span = 3
points_one_side = 5
for n in range(len(data)):
img_file = join(
save_dir, "sep_x{}_num_samples={}.png".format(n, args.num_samples))
model.cond_all_latents_traverse_v2(
img_file,
sess,
data[n],
z_comps=ids_sorted,
z_comp_labels=[
"z[{}] (SEP={:.4f}, INFO={:.4f})".format(
idx, SEP_zi[idx], MI_zi_x[idx]) for idx in ids_sorted
],
span=span,
points_1_side=points_one_side,
hl_x=True,
font_size=9,
title_font_scale=1.5,
subplot_adjust={
'left': 0.55,
'right': 0.99,
'bottom': 0.01,
'top': 0.99
},
size_inches=(4.0, 1.7),
batch_size=args.batch_size,
dec_output_2_img_func=binary_float_to_uint8)
f.close()
def main(args):
# =====================================
# Load config
# =====================================
with open(join(args.output_dir, 'config.json')) as f:
config = json.load(f)
args.__dict__.update(config)
# =====================================
# Dataset
# =====================================
data_file = join(RAW_DATA_DIR, "ComputerVision", "dSprites",
"dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz")
# It is already in the range [0, 1]
with np.load(data_file, encoding="latin1") as f:
x_train = f['imgs']
# 3 shape * 6 scale * 40 rotation * 32 pos X * 32 pos Y
y_train = f['latents_classes']
x_train = np.expand_dims(x_train.astype(np.float32), axis=-1)
num_train = len(x_train)
print("num_train: {}".format(num_train))
print("y_train[:10]: {}".format(y_train[:10]))
# =====================================
# Instantiate model
# =====================================
if args.enc_dec_model == "1Konny":
encoder = Encoder_1Konny(args.z_dim, stochastic=True)
decoder = Decoder_1Konny()
disc_z = DiscriminatorZ_1Konny(num_outputs=2)
else:
raise ValueError("Do not support enc_dec_model='{}'!".format(
args.enc_dec_model))
model = FactorVAE([64, 64, 1],
args.z_dim,
encoder=encoder,
decoder=decoder,
discriminator_z=disc_z,
rec_x_mode=args.rec_x_mode,
use_gp0_z_tc=True,
gp0_z_tc_mode=args.gp0_z_tc_mode)
loss_coeff_dict = {
'rec_x': args.rec_x_coeff,
'kld_loss': args.kld_loss_coeff,
'tc_loss': args.tc_loss_coeff,
'gp0_z_tc': args.gp0_z_tc_coeff,
'Dz_tc_loss_coeff': args.Dz_tc_loss_coeff,
}
model.build(loss_coeff_dict)
SimpleParamPrinter.print_all_params_tf_slim()
# =====================================
# Load model
# =====================================
config_proto = tf.ConfigProto(allow_soft_placement=True)
config_proto.gpu_options.allow_growth = True
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.9
sess = tf.Session(config=config_proto)
model_dir = make_dir_if_not_exist(join(args.output_dir, "model_tf"))
train_helper = SimpleTrainHelper(log_dir=None, save_dir=model_dir)
# Load model
train_helper.load(sess, load_step=args.load_step)
# =====================================
# Experiments
save_dir = make_dir_if_not_exist(
join(args.save_dir, "{}_{}".format(args.enc_dec_model, args.run)))
# =====================================
np.set_printoptions(threshold=np.nan,
linewidth=1000,
precision=5,
suppress=True)
num_bins = args.num_bins
bin_limits = tuple([float(s) for s in args.bin_limits.split(";")])
data_proportion = args.data_proportion
num_data = int(data_proportion * num_train)
assert num_data == num_train, "For dSprites, you must use all data!"
top_k = args.top_k
eps = 1e-8
# file
f = open(join(
save_dir, 'log[bins={},bin_limits={},data={}].txt'.format(
num_bins, bin_limits, data_proportion)),
mode='w')
# print function
print_ = functools.partial(print_both, file=f)
print_("num_bins: {}".format(num_bins))
print_("bin_limits: {}".format(bin_limits))
print_("data_proportion: {}".format(data_proportion))
print_("top_k: {}".format(top_k))
# Compute bins
# ================================= #
print_("")
print_("bin_limits: {}".format(bin_limits))
assert len(bin_limits) == 2 and bin_limits[0] < bin_limits[
1], "bin_limits={}".format(bin_limits)
bins = np.linspace(bin_limits[0],
bin_limits[1],
num_bins + 1,
endpoint=True)
print_("bins: {}".format(bins))
assert len(bins) == num_bins + 1
bin_widths = [bins[b] - bins[b - 1] for b in range(1, len(bins))]
print_("bin_widths: {}".format(bin_widths))
assert len(bin_widths
) == num_bins, "len(bin_widths)={} while num_bins={}!".format(
len(bin_widths), num_bins)
assert np.all(np.greater(bin_widths,
0)), "bin_widths: {}".format(bin_widths)
bin_centers = [(bins[b] + bins[b - 1]) * 0.5 for b in range(1, len(bins))]
print_("bin_centers: {}".format(bin_centers))
assert len(bin_centers
) == num_bins, "len(bin_centers)={} while num_bins={}!".format(
len(bin_centers), num_bins)
# ================================= #
# Compute representations
# ================================= #
z_data_file = join(args.informativeness_metrics_dir,
"{}_{}".format(args.enc_dec_model, args.run),
"z_data[data={}].npz".format(data_proportion))
with np.load(z_data_file, "r") as f:
all_z_mean = f['all_z_mean']
all_z_stddev = f['all_z_stddev']
print_("")
print_("all_z_mean.shape: {}".format(all_z_mean.shape))
print_("all_z_stddev.shape: {}".format(all_z_stddev.shape))
# ================================= #
# Compute the mutual information
# ================================= #
mi_file = join(
args.informativeness_metrics_dir,
"{}_{}".format(args.enc_dec_model, args.run),
'results[bins={},bin_limits={},data={}].npz'.format(
num_bins, bin_limits, data_proportion))
with np.load(mi_file, "r") as f:
sorted_MI_z_x = f['sorted_MI_z_x']
sorted_z_ids = f['sorted_z_comps']
H_z = f['H_z']
if top_k > 0:
top_MI = sorted_MI_z_x[:top_k]
top_z_ids = sorted_z_ids[:top_k]
bot_MI = sorted_MI_z_x[-top_k:]
bot_z_ids = sorted_z_ids[-top_k:]
top_bot_MI = np.concatenate([top_MI, bot_MI], axis=0)
top_bot_z_ids = np.concatenate([top_z_ids, bot_z_ids], axis=0)
print_("top MI: {}".format(top_MI))
print_("top_z_ids: {}".format(top_z_ids))
print_("bot MI: {}".format(bot_MI))
print_("bot_z_ids: {}".format(bot_z_ids))
else:
top_bot_MI = sorted_MI_z_x
top_bot_z_ids = sorted_z_ids
# ================================= #
H_z1z2_mean_mat = np.full(
[len(top_bot_z_ids), len(top_bot_z_ids)], -1, dtype=np.float32)
MI_z1z2_mean_mat = np.full(
[len(top_bot_z_ids), len(top_bot_z_ids)], -1, dtype=np.float32)
H_z1z2_mean = []
MI_z1z2_mean = []
z1z2_ids = []
# Compute the mutual information
# ================================= #
for i in range(len(top_bot_z_ids)):
z_idx1 = top_bot_z_ids[i]
H_s1 = H_z[z_idx1]
for j in range(i + 1, len(top_bot_z_ids)):
z_idx2 = top_bot_z_ids[j]
H_s2 = H_z[z_idx2]
print_("")
print_("Compute MI(z{}_mean, z{}_mean)!".format(z_idx1, z_idx2))
s1s2_mean_counter = np.zeros([num_bins, num_bins], dtype=np.int32)
for batch_ids in iterate_data(len(all_z_mean),
100,
shuffle=False,
include_remaining=True):
s1 = at_bin(all_z_mean[batch_ids, z_idx1], bins, one_hot=False)
s2 = at_bin(all_z_mean[batch_ids, z_idx2], bins, one_hot=False)
for s1_, s2_ in zip(s1, s2):
s1s2_mean_counter[s1_, s2_] += 1
# I(s1, s2) = Q(s1, s2) * (log Q(s1, s2) - log Q(s1) log Q(s2))
# ---------------------------------- #
Q_s1s2_mean = (s1s2_mean_counter *
1.0) / np.sum(s1s2_mean_counter).astype(np.float32)
log_Q_s1s2_mean = np.log(np.maximum(Q_s1s2_mean, eps))
H_s1s2_mean = -np.sum(Q_s1s2_mean * log_Q_s1s2_mean)
MI_s1s2_mean = H_s1 + H_s2 - H_s1s2_mean
print_("H_s1: {}".format(H_s1))
print_("H_s2: {}".format(H_s2))
print_("H_s1s2_mean: {}".format(H_s1s2_mean))
print_("MI_s1s2_mean: {}".format(MI_s1s2_mean))
H_z1z2_mean.append(H_s1s2_mean)
MI_z1z2_mean.append(MI_s1s2_mean)
z1z2_ids.append((z_idx1, z_idx2))
H_z1z2_mean_mat[i, j] = H_s1s2_mean
H_z1z2_mean_mat[j, i] = H_s1s2_mean
MI_z1z2_mean_mat[i, j] = MI_s1s2_mean
MI_z1z2_mean_mat[j, i] = MI_s1s2_mean
H_z1z2_mean = np.asarray(H_z1z2_mean, dtype=np.float32)
MI_z1z2_mean = np.asarray(MI_z1z2_mean, dtype=np.float32)
z1z2_ids = np.asarray(z1z2_ids, dtype=np.int32)
result_file = join(
save_dir, "results[bins={},bin_limits={},data={},k={}].npz".format(
num_bins, bin_limits, data_proportion, top_k))
results = {
'H_z1z2_mean': H_z1z2_mean,
'MI_z1z2_mean': MI_z1z2_mean,
'H_z1z2_mean_mat': H_z1z2_mean_mat,
'MI_z1z2_mean_mat': MI_z1z2_mean_mat,
'z1z2_ids': z1z2_ids,
}
np.savez_compressed(result_file, **results)
# ================================= #
f.close()