def test_CVAE_latent_space():
x = train_set.Ys_index_label
idxs_labels = []
for i in range(10):
idxs_label = np.where(x == i)
idxs_labels += [(full_Xs[idxs_label], full_Ys[idxs_label])]
plot = PlotTools()
model = CVAE
params = {
'loss_type': 'VAE',
'learning_rate': 0.01,
'latent_code_size': 2,
'encoder_net_shapes': (512, 256, 128, 64, 32),
'encoder_kwargs': {
'tail_bn': False,
'tail_activation': 'none',
'linear_stack_bn': False,
'linear_stack_activation': 'lrelu',
},
'decoder_net_shapes': (32, 64, 128, 256, 512),
'decoder_kwargs': {
'tail_bn': True,
'tail_activation': 'sigmoid',
'linear_stack_bn': True,
'linear_stack_activation': 'relu'
},
'batch_size': 256,
# 'KL_D_rate': 0.01
}
ae = model(**params)
n_iter = 50
for i in range(n_iter):
ae.train(full_Xs, full_Ys, epoch=1)
metric = ae.metric(full_Xs, full_Ys)
# ae.save(save_path)
if np.isnan(metric):
print(f'metric is {metric}')
break
print(metric)
codes = []
for x, y in idxs_labels:
code = ae.code(x, y)
codes += [code]
plot.scatter_2d(*codes, title=f'CVAE_latent_space_epoch_{i}.png')
for label, code in enumerate(codes):
plot.scatter_2d(code, title=f'vae_latent_space_epoch_{i}_label+{label}.png')
del ae
def test_VAE_latent_space(n_iter=100):
dataset_path = """C:\\Users\\demetoir_desktop\\PycharmProjects\\MLtools\\data\\MNIST"""
dataset_pack = MNIST().load(dataset_path)
dataset_pack.shuffle()
train_set = dataset_pack['train']
full_Xs, full_Ys = train_set.full_batch()
x = train_set.Ys_index_label
idxs_labels = []
for i in range(10):
idxs_label = np.where(x == i)
idxs_labels += [(full_Xs[idxs_label], full_Ys[idxs_label])]
plot = PlotTools()
model = VAE
params = {
'loss_type': 'VAE',
'learning_rate': 0.01,
'latent_code_size': 2,
'encoder_net_shapes': (512, 256, 128, 64, 32),
'encoder_kwargs': {
'tail_bn': False,
'tail_activation': 'lrelu',
'linear_stack_bn': False,
'linear_stack_activation': 'lrelu',
},
'decoder_net_shapes': (32, 64, 128, 256, 512),
'decoder_kwargs': {
'tail_bn': True,
'tail_activation': 'sigmoid',
'linear_stack_bn': True,
'linear_stack_activation': 'relu'
},
'batch_size': 256,
# 'KL_D_rate': 0.01
}
ae = model(**params)
for i in range(n_iter):
ae.train(full_Xs, epoch=1)
# ae.save(save_path)
metric = ae.metric(full_Xs)
if np.isnan(metric):
print(f'metric is {metric}')
break
print(metric)
codes = [ae.code(x) for x, y in idxs_labels]
plot.scatter_2d(*codes, title=f'vae_latent_space_epoch_{i}.png')
del ae
def test_VAE_latent_space_grid_search(n_iter=6):
dataset_path = """C:\\Users\\demetoir_desktop\\PycharmProjects\\MLtools\\data\\MNIST"""
dataset_pack = MNIST().load(dataset_path)
dataset_pack.shuffle()
train_set = dataset_pack['train']
full_Xs, full_Ys = train_set.full_batch()
x = train_set.Ys_index_label
idxs_labels = []
for i in range(10):
idxs_label = np.where(x == i)
idxs_labels += [(full_Xs[idxs_label], full_Ys[idxs_label])]
plot = PlotTools()
model = VAE
save_path = './CVAE.pkl'
BOOLs = [True, False]
linear_stack_activation = ['sigmoid', 'tanh', 'relu', 'lrelu', 'elu']
param_grid = params_to_dict(
latent_code_size=[2],
learning_rate=[0.005],
encoder_net_shapes=[(512, 256, 128, 64, 32)],
decoder_net_shapes=[(32, 64, 128, 256, 512)],
batch_size=[256],
# KL_D_rate=[0.01],
loss_type=['VAE'],
encoder_kwargs=params_to_dict(
linear_stack_bn=[False],
linear_stack_activation=['relu', 'lrelu', 'elu'],
tail_bn=[False],
tail_activation=['lrelu', 'none'],
),
decoder_kwargs=params_to_dict(
linear_stack_bn=[True],
linear_stack_activation=['relu'],
tail_bn=[False],
tail_activation=['sigmoid'],
))
param_grid = param_grid_full(param_grid)
df = DF({'params': list(param_grid)})
df.to_csv('./params.csv')
for param_idx, params in enumerate(param_grid):
pprint(param_idx, params)
ae = model(**params)
for i in range(50):
ae.train(full_Xs, epoch=1)
# ae.save(save_path)
metric = ae.metric(full_Xs)
if np.isnan(metric):
print(f'param_idx:{param_idx}, metric is {metric}')
break
print(metric)
codes = []
for x, y in idxs_labels:
code = ae.code(x)
codes += [code]
plot.scatter_2d(
*codes,
title=f'param_idx_{param_idx}_vae_latent_space_epoch_{i}.png')
del ae
def test_AAE_latent_space():
x = train_set.Ys_index_label
idxs_labels = []
for i in range(10):
idxs_label = np.where(x == i)
idxs_labels += [(full_Xs[idxs_label], full_Ys[idxs_label])]
plot = PlotTools()
model = AAE
save_path = './CVAE.pkl'
BOOLs = [True, False]
linear_stack_activation = ['sigmoid', 'tanh', 'relu', 'lrelu', 'elu']
param_grid = params_to_dict(
latent_code_size=[2],
learning_rate=[0.01],
encoder_net_shapes=[(512, 256)],
decoder_net_shapes=[(256, 512)],
batch_size=[100],
# KL_D_rate=[0.01],
encoder_kwargs=params_to_dict(
linear_stack_bn=[False],
linear_stack_activation=['relu'],
tail_bn=[False],
tail_activation=['none'],
),
decoder_kwargs=params_to_dict(
linear_stack_bn=[False],
linear_stack_activation=['relu'],
tail_bn=[False],
tail_activation=['sigmoid'],
))
param_grid = param_grid_full(param_grid)
df = DF({'params': list(param_grid)})
df.to_csv('./params.csv')
for param_idx, params in enumerate(param_grid):
pprint(param_idx, params)
ae = model(**params)
for i in range(100):
ae.train(full_Xs, full_Ys, epoch=1)
# ae.save(save_path)
metric = ae.metric(full_Xs, full_Ys)
if any([np.isnan(val) for val in metric.values()]):
print(f'param_idx:{param_idx}, metric is {metric}')
break
print(metric)
codes = []
for x, y in idxs_labels:
code = ae.code(x)
codes += [code]
plot.scatter_2d(
*codes,
title=f'param_idx_{param_idx}/aae_latent_space_epoch_{i}.png')
# for label, code in enumerate(codes):
# plot.scatter_2d(code, title=f'param_idx_{param_idx}/aae_latent_space_epoch_{i}_label+{label}.png')
recon = ae.recon(sample_Xs, sample_Ys)
gen = ae.generate(full_Ys[:30])
code_walk = np.concatenate(
[ae.augmentation(sample_Xs, sample_Ys) for _ in range(5)],
axis=0)
recon_sharpen = ae.recon_sharpen(sample_Xs, sample_Ys)
np_img = np.concatenate(
[sample_Xs, recon, recon_sharpen, gen, code_walk])
np_img = np_img_float32_to_uint8(np_img)
file_name = f'./matplot/param_idx_{param_idx}/aae_img_epoch_{i}.png'
plot.plot_image_tile(np_img, file_name, column=5)
# sample_imgs = Xs_gen
del ae
def test_CVAE_latent_space_grid_search():
x = train_set.Ys_index_label
idxs_labels = []
for i in range(10):
idxs_label = np.where(x == i)
idxs_labels += [(full_Xs[idxs_label], full_Ys[idxs_label])]
plot = PlotTools()
model = CVAE
save_path = './CVAE.pkl'
BOOLs = [True, False]
linear_stack_activation = ['sigmoid', 'tanh', 'relu', 'lrelu', 'elu']
param_grid = params_to_dict(
latent_code_size=[2],
learning_rate=[0.005],
encoder_net_shapes=[(512, 256, 128, 64, 32)],
decoder_net_shapes=[(32, 64, 128, 256, 512)],
batch_size=[256],
# KL_D_rate=[0.01],
loss_type=['VAE'],
encoder_kwargs=params_to_dict(
linear_stack_bn=[False],
linear_stack_activation=['relu', 'lrelu', 'elu'],
tail_bn=[False],
tail_activation=['lrelu', 'none'],
),
decoder_kwargs=params_to_dict(
linear_stack_bn=[True],
linear_stack_activation=['relu'],
tail_bn=[False],
tail_activation=['sigmoid'],
)
)
param_grid = param_grid_full(param_grid)
df = DF({
'params': list(param_grid)
})
df.to_csv('./params.csv')
for param_idx, params in enumerate(param_grid):
pprint(param_idx, params)
ae = model(**params)
for i in range(10):
ae.train(full_Xs, full_Ys, epoch=1)
# ae.save(save_path)
metric = ae.metric(full_Xs, full_Ys)
if np.isnan(metric):
print(f'param_idx:{param_idx}, metric is {metric}')
break
print(metric)
codes = []
for x, y in idxs_labels:
code = ae.code(x, y)
codes += [code]
plot.scatter_2d(*codes, title=f'param_idx_{param_idx}/vae_latent_space_epoch_{i}.png')
# for label, code in enumerate(codes):
# plot.scatter_2d(code, title=f'param_idx_{param_idx}/vae_latent_space_epoch_{i}_label+{label}.png')
recon = ae.recon(sample_Xs, sample_Ys)
gen = ae.generate(sample_Ys)
np_img = np.concatenate([sample_Xs, recon, gen])
np_img = np_img_float32_to_uint8(np_img)
# sample_imgs = Xs_gen
file_name = f'./matplot/param_idx_{param_idx}/vae_img_epoch_{i}.png'
tile = np_img_to_tile(np_img, column_size=5)
np_image_save(tile, file_name)
del ae