np.min(image_s_trains), np.max(image_s_trains)))
print("[image_t_trains] shape={}, dtype={}, min={}, max={}".format(
image_t_trains.shape, image_t_trains.dtype,
np.min(image_t_trains), np.max(image_t_trains)))
print("[image_s_valids] shape={}, dtype={}, min={}, max={}".format(
image_s_valids.shape, image_s_valids.dtype,
np.min(image_s_valids), np.max(image_s_valids)))
print("[image_t_valids] shape={}, dtype={}, min={}, max={}".format(
image_t_valids.shape, image_t_valids.dtype,
np.min(image_t_valids), np.max(image_t_valids)))
#================================
# モデルの構造を定義する。
#================================
with mirrored_strategy.scope():
model_G = TempleteNetworks(out_dim=3)
model_G(
tf.zeros([args.batch_size, args.image_height, args.image_width, 3],
dtype=tf.float32))
#================================
# loss 設定
#================================
with mirrored_strategy.scope():
loss_mse = tf.keras.losses.MeanSquaredError()
#================================
# optimizer 設定
#================================
with mirrored_strategy.scope():
optimizer_G = tf.keras.optimizers.Adam(learning_rate=args.lr,
args.dataset_dir,
datamode="test",
image_height=args.image_height,
image_width=args.image_width,
data_augument=False,
debug=args.debug)
dloader_test = torch.utils.data.DataLoader(ds_test,
batch_size=args.batch_size_test,
shuffle=False,
num_workers=args.n_workers,
pin_memory=True)
#================================
# モデルの構造を定義する。
#================================
model_G = TempleteNetworks().to(device)
if (args.debug):
print("model_G\n", model_G)
# モデルを読み込む
if not args.load_checkpoints_path == '' and os.path.exists(
args.load_checkpoints_path):
load_checkpoint(model_G, device, args.load_checkpoints_path)
#================================
# モデルの推論
#================================
print("Starting Testing Loop...")
n_print = 1
model_G.eval()
for step, inputs in enumerate(tqdm(dloader_test, desc="Samplings")):
data_augument=False)
#================================
# 変数とプレースホルダを設定
#================================
image_s_holder = tf.placeholder(
tf.float32, [None, args.image_height, args.image_width, 3],
name="image_s_holder")
image_t_holder = tf.placeholder(
tf.float32, [None, args.image_height, args.image_width, 3],
name="image_t_holder")
#================================
# モデルの構造を定義する。
#================================
model_G = TempleteNetworks(out_dim=3)
output_op = model_G(image_s_holder)
#================================
# loss 関数の設定
#================================
with tf.name_scope('loss'):
loss_op = tf.reduce_mean(tf.abs(image_t_holder - output_op))
#================================
# optimizer の設定
#================================
with tf.name_scope('optimizer'):
optimizer_op = tf.train.AdamOptimizer(learning_rate=args.lr,
beta1=args.beta1,
beta2=args.beta2)
image_height=args.image_height,
image_width=args.image_width,
n_channels=3,
batch_size=args.batch_size,
use_tfrecord=args.use_tfrecord,
seed=args.seed)
if (args.debug):
print("n_trains : ", n_trains)
print("n_valids : ", n_valids)
print("ds_train : ", ds_train)
print("ds_valid : ", ds_valid)
#================================
# モデルの構造を定義する。
#================================
model_G = TempleteNetworks(out_dim=3)
if (args.debug):
model_G(
tf.zeros([args.batch_size, args.image_height, args.image_width, 3],
dtype=tf.float32)
) # 動的作成されるネットワークなので、一度ネットワークに入力データを供給しないと summary() を出力できない
model_G.summary()
#================================
# モデルの読み込み
#================================
if (args.load_checkpoints_dir):
ckpt_state = tf.train.get_checkpoint_state(args.load_checkpoints_dir)
model_G.load_weights(ckpt_state.model_checkpoint_path)
print("load checkpoints in `{}`.".format(
ckpt_state.model_checkpoint_path))
dloader_train = torch.utils.data.DataLoader(Subset(ds_train, train_index),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.n_workers,
pin_memory=True)
dloader_valid = torch.utils.data.DataLoader(
Subset(ds_train, valid_index),
batch_size=args.batch_size_valid,
shuffle=False,
num_workers=args.n_workers,
pin_memory=True)
#================================
# モデルの構造を定義する。
#================================
model_G = TempleteNetworks().to(device)
if (args.debug):
print("model_G\n", model_G)
# モデルを読み込む
if not args.load_checkpoints_path == '' and os.path.exists(
args.load_checkpoints_path):
load_checkpoint(model_G, device, args.load_checkpoints_path)
#================================
# optimizer_G の設定
#================================
optimizer_G = optim.Adam(params=model_G.parameters(),
lr=args.lr,
betas=(args.beta1, args.beta2))