steps_etopo = L_etopo // batch_size
cut, cut_etopo = 100, 300
map_sizes = [64, 96, 128]
c_size = 4
labels = ['X', 'Y']
flag_c = [True] * c_size
save_dir = '/glade/work/ksha/data/Keras/BACKUP/'
for sea in seasons:
print(sea)
# validation data
valid_sub = glob(BATCH_dir +
'{}*BATCH*{}*SUB_{}*.npy'.format(VAR, kind, sea))
gen_valid_sub = ku.grid_grid_gen(valid_sub,
batch_size,
c_size,
m_size,
labels,
flag=flag_c)
# --------------------------------------------------------------------------------------- #
# ===== step 1: elev. tune ===== #
record_t = 120
max_tol = 5
tol = 0
for n in range(20):
print('====================')
if n == 0:
model_name = 'UNET_{}_B4_{}.hdf'.format(VAR, sea)
unet_backbone = keras.models.load_model(save_dir + model_name)
unet_tune2 = au.UNET_AE(layer_N, c_size, dropout=False)
unet_tune2 = freeze_unet2(unet_tune2, l=0, lr=1e-6)
unet_tune2.set_weights(unet_backbone.get_weights())
VAR = 'TMIN'
c_size = 4; norm = 'STD'; kind = 'std_clim'; batch_num = 200
layer_N = [56, 112, 224, 448]
lr = [5e-5, 5e-6]
# data pipeline setup
L_train, batch_size, m_size = 800, 1, batch_num
steps = L_train//batch_size
labels = ['X', 'Y']; flag_c = [True]*c_size
seasons = ['djf', 'mam', 'jja', 'son']
for sea in seasons:
print('========== {} =========='.format(sea))
# validation batches
valid_files = glob(BATCH_dir+'{}*BATCH*{}*VORI_{}*.npy'.format(VAR, kind, sea))
shuffle(valid_files)
gen_valid = ku.grid_grid_gen(valid_files, batch_size, c_size, m_size, labels, flag=flag_c)
# model import
save_dir = '/glade/work/ksha/data/Keras/BACKUP/'
model_name = 'UNET_{}_B4_{}'.format(VAR, sea)
model_name_tune = 'UNET_{}_B4_{}_tune_train'.format(VAR, sea)
model = keras.models.load_model(save_dir+model_name+'.hdf') # restart options
## opt
opt_adam = keras.optimizers.SGD(lr=5e-5, decay=0.01)
## callback
callbacks = [keras.callbacks.EarlyStopping(monitor='val_HR_temp_loss', min_delta=0.0000001, patience=3, verbose=True),
keras.callbacks.ModelCheckpoint(filepath=save_dir+model_name_tune+'.hdf', verbose=True,
monitor='val_HR_temp_loss', save_best_only=True)]
model.compile(loss=keras.losses.mean_absolute_error, optimizer=opt_adam, metrics=[keras.losses.mean_absolute_error])
train_files = glob(BATCH_dir+'{}*BATCH*{}*TORI_{}*.npy'.format(VAR, kind, sea))
shuffle(train_files)
gen_train = ku.grid_grid_gen(train_files, batch_size, c_size, m_size, labels, flag=flag_c)
for VAR in VARS:
print(VAR)
model_unet_a = keras.models.load_model(save_dir +
'UNET_{}_A_std'.format(VAR) +
'.hdf')
model_unet_c = keras.models.load_model(save_dir +
'UNET_{}_B_tune'.format(VAR) +
'.hdf')
train_files = glob(BATCH_dir + '{}*BATCH*128*TORI[0-9]*.npy'.format(VAR))
L_train, batch_size, m_size = len(train_files), 1, 200
steps = L_train // batch_size
train_gen = ku.grid_grid_gen(train_files,
batch_size,
c_size,
m_size,
labels,
flag=flag_c)
train_gen_LRT = grid_grid_gen(train_files,
batch_size,
c_size,
m_size,
labels,
0,
flag=flag_c)
train_gen_HRZ = grid_grid_gen(train_files,
batch_size,
c_size,
m_size,
labels,
1,
