def __readfile(self, names, c_size, m_size, labels, shuffle_ind, flag):
N = len(names)
data_temp = np.load(names[0])
GRID_IN = data_temp[()][labels[0]][...]
Y_LABEL = data_temp[()][labels[1]][..., None]
Y2 = np.copy(-1 * GRID_IN[..., 1][..., None])
rnd_ind = du.shuffle_ind(200)
GRID_IN[..., shuffle_ind] = GRID_IN[rnd_ind, ..., shuffle_ind]
return [GRID_IN], [Y_LABEL, Y2]
train_size = 100
record = 999
for i in range(epochs):
print('epoch = {}'.format(i))
start_time = time.time()
shuffle(trainfile64)
shuffle(trainfile96)
trainfiles = trainfile64[:160] + trainfile96[:160]
# shuffling at epoch begin
shuffle(trainfiles)
# loop over batches
for j, name in enumerate(trainfiles):
# ----- import batch data subset ----- #
inds = du.shuffle_ind(batch_size)[:train_size]
temp_batch = np.load(name, allow_pickle=True)[()]
X = temp_batch['batch'][inds, ...]
# ------------------------------------ #
# ----- D training ----- #
# Latent space sampling
Wf = np.random.normal(0.0, 1.0, size=[train_size, latent_size])
# soft labels
dummy_bad = np.ones(train_size) * 0.1 + np.random.uniform(
-0.02, 0.02, train_size)
dummy_good = np.ones(train_size) * 0.9 + np.random.uniform(
-0.02, 0.02, train_size)
# get G_output (channel last)
g_in = [Wf, X[..., input_flag]]
g_out = G_style.predict(g_in) # <-- np.array
tol = 0
train_size = 100
for i in range(epochs):
print('epoch = {}'.format(i))
if i == 0:
record = G.evaluate_generator(gen_valid, verbose=1)
print('Initial validation loss: {}'.format(record))
start_time = time.time()
# learning rate schedule
# shuffling at epoch begin
shuffle(trainfiles)
# loop over batches
for j, name in enumerate(trainfiles):
inds = du.shuffle_ind(batch_size)
inds = inds[:train_size]
# dynamic soft labels
y_bad = np.ones(train_size) * 0.1 + np.random.uniform(
-0.02, 0.02, train_size)
y_good = np.ones(train_size) - y_bad
dummy_bad = keras.utils.to_categorical(y_bad)
dummy_good = keras.utils.to_categorical(y_good)
# import batch data
temp_batch = np.load(name, allow_pickle=True)[()]
X = temp_batch['batch'][inds, ...]
# get G_output
g_in = X[..., input_flag]
g_out = G.predict([g_in]) # <-- np.array
# import batch data
temp_batch = np.load(name, allow_pickle=True)[()]
X = temp_batch['batch']
# D training
D.trainable = True
g_in = X[..., input_flag]
g_out = G.predict([g_in]) # <-- np.array
d_in_fake = np.concatenate((g_out, g_in), axis=-1) # channel last
d_in_true = X[..., inout_flag]
if mix:
d_in = np.concatenate((d_in_fake, d_in_true),
axis=0) # batch size doubled
d_target = np.concatenate((dummy_bad, dummy_good), axis=0)
d_shuffle_ind = du.shuffle_ind(2 * batch_size)
d_loss1 = D.train_on_batch(d_in[d_shuffle_ind, ...],
d_target[d_shuffle_ind, ...])
d_loss2 = 0
else:
d_loss1 = D.train_on_batch(d_in_true, dummy_good)
d_loss2 = D.train_on_batch(d_in_fake, dummy_bad)
d_loss_sum = d_loss1 + d_loss2
D_LOSS[i * L_train + j] = d_loss_sum
LOSS = {'D_LOSS': D_LOSS}
np.save(hist_path, LOSS)
if j % 50 == 0:
print('\t{} step loss = {}'.format(j, d_loss_sum))
print('Save to {}'.format(model_path))