From 96e97968725f52c6e147adf7ca1125514bdf9188 Mon Sep 17 00:00:00 2001
From: tomrink <rink@ssec.wisc.edu>
Date: Thu, 6 Jul 2023 13:35:02 -0500
Subject: [PATCH] snapshot...
---
.../deeplearning/cloud_opd_srcnn_abi_v2.py | 927 ++++++++++++++++++
1 file changed, 927 insertions(+)
create mode 100644 modules/deeplearning/cloud_opd_srcnn_abi_v2.py
diff --git a/modules/deeplearning/cloud_opd_srcnn_abi_v2.py b/modules/deeplearning/cloud_opd_srcnn_abi_v2.py
new file mode 100644
index 00000000..3037a221
--- /dev/null
+++ b/modules/deeplearning/cloud_opd_srcnn_abi_v2.py
@@ -0,0 +1,927 @@
+import gc
+import glob
+import tensorflow as tf
+
+from util.setup import logdir, modeldir, now, ancillary_path
+from util.util import EarlyStop, normalize, denormalize, scale, descale, get_grid_values_all, resample_2d_linear, smooth_2d
+import os, datetime
+import numpy as np
+import pickle
+import h5py
+import time
+
+LOG_DEVICE_PLACEMENT = False
+
+PROC_BATCH_SIZE = 4
+PROC_BATCH_BUFFER_SIZE = 5000
+
+NumClasses = 2
+if NumClasses == 2:
+ NumLogits = 1
+else:
+ NumLogits = NumClasses
+
+BATCH_SIZE = 128
+NUM_EPOCHS = 80
+
+TRACK_MOVING_AVERAGE = False
+EARLY_STOP = True
+
+NOISE_TRAINING = False
+NOISE_STDDEV = 0.01
+DO_AUGMENT = True
+
+DO_SMOOTH = False
+SIGMA = 1.0
+DO_ZERO_OUT = False
+
+# setup scaling parameters dictionary
+mean_std_dct = {}
+mean_std_file = ancillary_path+'mean_std_lo_hi_l2.pkl'
+f = open(mean_std_file, 'rb')
+mean_std_dct_l2 = pickle.load(f)
+f.close()
+
+mean_std_file = ancillary_path+'mean_std_lo_hi_l1b.pkl'
+f = open(mean_std_file, 'rb')
+mean_std_dct_l1b = pickle.load(f)
+f.close()
+
+mean_std_dct.update(mean_std_dct_l1b)
+mean_std_dct.update(mean_std_dct_l2)
+
+IMG_DEPTH = 1
+
+label_param = 'cld_opd_dcomp'
+
+params = ['temp_11_0um_nom', 'refl_0_65um_nom', 'refl_submin_ch01', 'refl_submax_ch01', 'refl_substddev_ch01', 'temp_stddev3x3_ch31', 'refl_stddev3x3_ch01', label_param]
+params_i = ['temp_11_0um_nom', 'refl_0_65um_nom', 'temp_stddev3x3_ch31', 'refl_stddev3x3_ch01', label_param]
+data_params_half = ['temp_11_0um_nom', 'refl_0_65um_nom']
+data_params_full = ['refl_0_65um_nom']
+sub_fields = ['refl_submin_ch01', 'refl_submax_ch01', 'refl_substddev_ch01']
+# sub_fields = ['refl_stddev3x3_ch01']
+
+label_idx_i = params_i.index(label_param)
+label_idx = params.index(label_param)
+
+print('data_params_half: ', data_params_half)
+print('data_params_full: ', data_params_full)
+print('label_param: ', label_param)
+
+KERNEL_SIZE = 3
+
+
+def build_residual_conv2d_block(conv, num_filters, block_name, activation=tf.nn.relu, padding='SAME',
+ kernel_initializer='he_uniform', scale=None, kernel_size=3,
+ do_drop_out=True, drop_rate=0.5, do_batch_norm=True):
+
+ with tf.name_scope(block_name):
+ skip = tf.keras.layers.Conv2D(num_filters, kernel_size=kernel_size, padding=padding, kernel_initializer=kernel_initializer, activation=activation)(conv)
+ skip = tf.keras.layers.Conv2D(num_filters, kernel_size=kernel_size, padding=padding, activation=None)(skip)
+
+ if scale is not None:
+ skip = tf.keras.layers.Lambda(lambda x: x * scale)(skip)
+
+ if do_drop_out:
+ skip = tf.keras.layers.Dropout(drop_rate)(skip)
+
+ if do_batch_norm:
+ skip = tf.keras.layers.BatchNormalization()(skip)
+
+ conv = conv + skip
+ print(block_name+':', conv.shape)
+
+ return conv
+
+
+def upsample_nearest(grd):
+ bsize, ylen, xlen = grd.shape
+ up = np.zeros((bsize, ylen*2, xlen*2))
+
+ up[:, 0::2, 0::2] = grd[:, 0::, 0::]
+ up[:, 1::2, 0::2] = grd[:, 0::, 0::]
+ up[:, 0::2, 1::2] = grd[:, 0::, 0::]
+ up[:, 1::2, 1::2] = grd[:, 0::, 0::]
+
+ return up
+
+
+def upsample_mean(grd):
+ bsize, ylen, xlen = grd.shape
+ up = np.zeros((bsize, ylen*2, xlen*2))
+
+ up[:, ::2, ::2] = grd[:, ::2, ::2]
+ up[:, 1::2, ::2] = grd[:, ::2, ::2]
+ up[:, ::2, 1::2] = grd[:, ::2, ::2]
+ up[:, 1::2, 1::2] = grd[:, ::2, ::2]
+
+ return up
+
+
+def get_grid_cell_mean(grd_k):
+ grd_k = np.where(np.isnan(grd_k), 0, grd_k)
+ a = grd_k[:, 0::2, 0::2]
+ b = grd_k[:, 1::2, 0::2]
+ c = grd_k[:, 0::2, 1::2]
+ d = grd_k[:, 1::2, 1::2]
+ mean = np.nanmean([a, b, c, d], axis=0)
+
+ return mean
+
+
+def get_min_max_std(grd_k):
+ grd_k = np.where(np.isnan(grd_k), 0, grd_k)
+ a = grd_k[:, 0::2, 0::2]
+ b = grd_k[:, 1::2, 0::2]
+ c = grd_k[:, 0::2, 1::2]
+ d = grd_k[:, 1::2, 1::2]
+
+ lo = np.nanmin([a, b, c, d], axis=0)
+ hi = np.nanmax([a, b, c, d], axis=0)
+ std = np.nanstd([a, b, c, d], axis=0)
+ avg = np.nanmean([a, b, c, d], axis=0)
+
+ return lo, hi, std, avg
+
+
+def upsample_static(grd, x_2, y_2, t, s, y_k, x_k):
+ grd = resample_2d_linear(x_2, y_2, grd, t, s)
+ # grd = grd[:, y_k, x_k]
+ return grd
+
+
+class SRCNN:
+
+ def __init__(self, LEN_Y=128, LEN_X=128):
+
+ self.train_data = None
+ self.train_label = None
+ self.test_data = None
+ self.test_label = None
+ self.test_data_denorm = None
+
+ self.train_dataset = None
+ self.inner_train_dataset = None
+ self.test_dataset = None
+ self.eval_dataset = None
+ self.X_img = None
+ self.inputs = []
+
+ self.logits = None
+
+ self.predict_data = None
+ self.predict_dataset = None
+ self.mean_list = None
+ self.std_list = None
+
+ self.training_op = None
+ self.correct = None
+ self.accuracy = None
+ self.loss = None
+ self.pred_class = None
+ self.variable_averages = None
+
+ self.global_step = None
+
+ self.writer_train = None
+ self.writer_valid = None
+ self.writer_train_valid_loss = None
+
+ self.model = None
+ self.optimizer = None
+ self.ema = None
+ self.train_loss = None
+ self.train_accuracy = None
+ self.test_loss = None
+ self.test_accuracy = None
+
+ self.test_labels = []
+ self.test_preds = []
+ self.test_probs = None
+
+ self.learningRateSchedule = None
+ self.num_data_samples = None
+ self.initial_learning_rate = None
+
+ self.train_data_files = None
+ self.train_label_files = None
+ self.test_data_files = None
+ self.test_label_files = None
+
+ # self.n_chans = len(data_params_half) + len(data_params_full) + 1
+ self.n_chans = 6
+
+ self.X_img = tf.keras.Input(shape=(None, None, self.n_chans))
+
+ self.inputs.append(self.X_img)
+
+ self.slc_x_m = slice(1, int(LEN_X / 2) + 4)
+ self.slc_y_m = slice(1, int(LEN_Y / 2) + 4)
+ self.slc_x = slice(3, LEN_X + 5)
+ self.slc_y = slice(3, LEN_Y + 5)
+ self.slc_x_2 = slice(2, LEN_X + 7, 2)
+ self.slc_y_2 = slice(2, LEN_Y + 7, 2)
+ self.x_2 = np.arange(int(LEN_X / 2) + 3)
+ self.y_2 = np.arange(int(LEN_Y / 2) + 3)
+ self.t = np.arange(0, int(LEN_X / 2) + 3, 0.5)
+ self.s = np.arange(0, int(LEN_Y / 2) + 3, 0.5)
+ self.x_k = slice(1, LEN_X + 3)
+ self.y_k = slice(1, LEN_Y + 3)
+ self.x_128 = slice(4, LEN_X + 4)
+ self.y_128 = slice(4, LEN_Y + 4)
+ self.LEN_X = LEN_X
+ self.LEN_Y = LEN_Y
+
+ tf.debugging.set_log_device_placement(LOG_DEVICE_PLACEMENT)
+
+ def upsample(self, grd):
+ grd = resample_2d_linear(self.x_2, self.y_2, grd, self.t, self.s)
+ grd = grd[:, self.y_k, self.x_k]
+ return grd
+
+ def get_in_mem_data_batch(self, idxs, is_training):
+ if is_training:
+ data_files = self.train_data_files
+ label_files = self.train_label_files
+ else:
+ data_files = self.test_data_files
+ label_files = self.test_label_files
+
+ data_s = []
+ label_s = []
+ for k in idxs:
+ f = data_files[k]
+ nda = np.load(f)
+ data_s.append(nda)
+
+ f = label_files[k]
+ nda = np.load(f)
+ label_s.append(nda)
+ input_data = np.concatenate(data_s)
+ input_label = np.concatenate(label_s)
+
+ data_norm = []
+ for param in data_params_half:
+ idx = params.index(param)
+ tmp = input_data[:, idx, :, :]
+ tmp = np.where(np.isnan(tmp), 0.0, tmp)
+ tmp = tmp[:, self.slc_y_m, self.slc_x_m]
+ tmp = self.upsample(tmp)
+ tmp = smooth_2d(tmp)
+ tmp = normalize(tmp, param, mean_std_dct)
+ data_norm.append(tmp)
+
+ tmp = input_label[:, label_idx_i, :, :]
+ tmp = tmp.copy()
+ tmp = np.where(np.isnan(tmp), 0.0, tmp)
+ tmp = tmp[:, self.slc_y_2, self.slc_x_2]
+ tmp = self.upsample(tmp)
+ tmp = smooth_2d(tmp)
+ tmp = normalize(tmp, label_param, mean_std_dct)
+ data_norm.append(tmp)
+
+ # for param in sub_fields:
+ # idx = params.index(param)
+ # tmp = input_data[:, idx, :, :]
+ # tmp = np.where(np.isnan(tmp), 0.0, tmp)
+ # tmp = tmp[:, self.slc_y_m, self.slc_x_m]
+ # tmp = self.upsample(tmp)
+ # # if param != 'refl_substddev_ch01':
+ # if False:
+ # tmp = normalize(tmp, 'refl_0_65um_nom', mean_std_dct)
+ # else:
+ # tmp = np.where(np.isnan(tmp), 0.0, tmp)
+ # data_norm.append(tmp)
+
+ for param in sub_fields:
+ idx = params.index(param)
+ tmp = input_data[:, idx, :, :]
+ tmp = upsample_nearest(tmp)
+ tmp = tmp[:, self.slc_y, self.slc_x]
+ if param != 'refl_substddev_ch01':
+ tmp = normalize(tmp, 'refl_0_65um_nom', mean_std_dct)
+ else:
+ tmp = np.where(np.isnan(tmp), 0, tmp)
+ data_norm.append(tmp)
+ # ---------------------------------------------------
+
+ data = np.stack(data_norm, axis=3)
+ data = data.astype(np.float32)
+
+ # -----------------------------------------------------
+ # -----------------------------------------------------
+ label = input_label[:, label_idx_i, :, :]
+ label = label.copy()
+ label = normalize(label, label_param, mean_std_dct)
+ # label = scale(label, label_param, mean_std_dct)
+ label = label[:, self.y_128, self.x_128]
+
+ label = np.where(np.isnan(label), 0.0, label)
+ label = np.expand_dims(label, axis=3)
+
+ data = data.astype(np.float32)
+ label = label.astype(np.float32)
+
+ if is_training and DO_AUGMENT:
+ data_ud = np.flip(data, axis=1)
+ label_ud = np.flip(label, axis=1)
+
+ data_lr = np.flip(data, axis=2)
+ label_lr = np.flip(label, axis=2)
+
+ data = np.concatenate([data, data_ud, data_lr])
+ label = np.concatenate([label, label_ud, label_lr])
+
+ return data, label
+
+ def get_in_mem_data_batch_train(self, idxs):
+ return self.get_in_mem_data_batch(idxs, True)
+
+ def get_in_mem_data_batch_test(self, idxs):
+ return self.get_in_mem_data_batch(idxs, False)
+
+ @tf.function(input_signature=[tf.TensorSpec(None, tf.int32)])
+ def data_function(self, indexes):
+ out = tf.numpy_function(self.get_in_mem_data_batch_train, [indexes], [tf.float32, tf.float32])
+ return out
+
+ @tf.function(input_signature=[tf.TensorSpec(None, tf.int32)])
+ def data_function_test(self, indexes):
+ out = tf.numpy_function(self.get_in_mem_data_batch_test, [indexes], [tf.float32, tf.float32])
+ return out
+
+ def get_train_dataset(self, indexes):
+ indexes = list(indexes)
+
+ dataset = tf.data.Dataset.from_tensor_slices(indexes)
+ dataset = dataset.batch(PROC_BATCH_SIZE)
+ dataset = dataset.map(self.data_function, num_parallel_calls=8)
+ dataset = dataset.cache()
+ if DO_AUGMENT:
+ dataset = dataset.shuffle(PROC_BATCH_BUFFER_SIZE)
+ dataset = dataset.prefetch(buffer_size=1)
+ self.train_dataset = dataset
+
+ def get_test_dataset(self, indexes):
+ indexes = list(indexes)
+
+ dataset = tf.data.Dataset.from_tensor_slices(indexes)
+ dataset = dataset.batch(PROC_BATCH_SIZE)
+ dataset = dataset.map(self.data_function_test, num_parallel_calls=8)
+ dataset = dataset.cache()
+ self.test_dataset = dataset
+
+ def setup_pipeline(self, train_data_files, train_label_files, test_data_files, test_label_files, num_train_samples):
+ self.train_data_files = train_data_files
+ self.train_label_files = train_label_files
+ self.test_data_files = test_data_files
+ self.test_label_files = test_label_files
+
+ trn_idxs = np.arange(len(train_data_files))
+ np.random.shuffle(trn_idxs)
+
+ tst_idxs = np.arange(len(test_data_files))
+
+ self.get_train_dataset(trn_idxs)
+ self.get_test_dataset(tst_idxs)
+
+ self.num_data_samples = num_train_samples # approximately
+
+ print('datetime: ', now)
+ print('training and test data: ')
+ print('---------------------------')
+ print('num train samples: ', self.num_data_samples)
+ print('BATCH SIZE: ', BATCH_SIZE)
+ print('num test samples: ', tst_idxs.shape[0])
+ print('setup_pipeline: Done')
+
+ def setup_test_pipeline(self, test_data_files, test_label_files):
+ self.test_data_files = test_data_files
+ self.test_label_files = test_label_files
+ tst_idxs = np.arange(len(test_data_files))
+ self.get_test_dataset(tst_idxs)
+ print('setup_test_pipeline: Done')
+
+ def build_srcnn(self, do_drop_out=False, do_batch_norm=False, drop_rate=0.5, factor=2):
+ print('build_cnn')
+ padding = "SAME"
+
+ # activation = tf.nn.relu
+ # activation = tf.nn.elu
+ activation = tf.nn.relu
+ momentum = 0.99
+
+ num_filters = 64
+
+ input_2d = self.inputs[0]
+ input_2d_conv = input_2d[:, :, :, 0:3]
+ input_2d_no_conv = input_2d[:, 1:(130-1), 1:(130-1), 3:]
+ print('input: ', input_2d.shape)
+ print('input_2d_conv: ', input_2d_conv.shape)
+ print('input_2d_no_conv: ', input_2d_no_conv.shape)
+
+ conv = conv_b = tf.keras.layers.Conv2D(num_filters, kernel_size=KERNEL_SIZE, kernel_initializer='he_uniform', activation=activation, padding='VALID')(input_2d_conv)
+ print('conv: ', conv.shape)
+
+ conv_nc = tf.keras.layers.Conv2D(num_filters, kernel_size=1, kernel_initializer='he_uniform', activation=activation, padding='SAME')(input_2d_no_conv)
+ print('conv_nc: ', conv_nc.shape)
+
+ # if NOISE_TRAINING:
+ # conv = conv_b = tf.keras.layers.GaussianNoise(stddev=NOISE_STDDEV)(conv)
+
+ scale = 0.2
+
+ conv_b = build_residual_conv2d_block(conv_b, num_filters, 'Residual_Block_1', kernel_size=KERNEL_SIZE, scale=scale)
+
+ conv_b = build_residual_conv2d_block(conv_b, num_filters, 'Residual_Block_2', kernel_size=KERNEL_SIZE, scale=scale)
+
+ conv_b = build_residual_conv2d_block(conv_b, num_filters, 'Residual_Block_3', kernel_size=KERNEL_SIZE, scale=scale)
+
+ conv_b = build_residual_conv2d_block(conv_b, num_filters, 'Residual_Block_4', kernel_size=KERNEL_SIZE, scale=scale)
+
+ # conv_b = build_residual_conv2d_block(conv_b, num_filters, 'Residual_Block_5', kernel_size=KERNEL_SIZE, scale=scale)
+
+ # conv_b = build_residual_conv2d_block(conv_b, num_filters, 'Residual_Block_6', kernel_size=KERNEL_SIZE, scale=scale)
+
+ conv_b = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, activation=activation, kernel_initializer='he_uniform', padding=padding)(conv_b)
+
+ # -------------------------------------------------------------------------------------
+ #conv_nc = build_residual_conv2d_block(conv_nc, num_filters, 'Residual_Block_nc_1', kernel_size=1, scale=scale)
+ #conv_nc = build_residual_conv2d_block(conv_nc, num_filters, 'Residual_Block_nc_2', kernel_size=1, scale=scale)
+ print('conv_nc: ', conv_nc.shape)
+
+ conv = conv + conv_b + conv_nc
+ print(conv.shape)
+
+ conv = tf.keras.layers.Conv2D(16, kernel_size=1, strides=1, padding=padding)(conv)
+ conv = tf.keras.layers.Conv2D(16, kernel_size=1, strides=1, padding=padding)(conv)
+ conv = tf.keras.layers.Conv2D(16, kernel_size=1, strides=1, padding=padding)(conv)
+ conv = tf.keras.layers.Conv2D(16, kernel_size=1, strides=1, padding=padding)(conv)
+
+ # This is effectively a Dense layer
+ self.logits = tf.keras.layers.Conv2D(1, kernel_size=1, strides=1, padding=padding, name='regression')(conv)
+ print(self.logits.shape)
+
+ def build_training(self):
+
+ self.loss = tf.keras.losses.MeanSquaredError() # Regression
+
+ # decayed_learning_rate = learning_rate * decay_rate ^ (global_step / decay_steps)
+ initial_learning_rate = 0.002
+ decay_rate = 0.95
+ steps_per_epoch = int(self.num_data_samples/BATCH_SIZE) # one epoch
+ decay_steps = int(steps_per_epoch) * 4
+ print('initial rate, decay rate, steps/epoch, decay steps: ', initial_learning_rate, decay_rate, steps_per_epoch, decay_steps)
+
+ self.learningRateSchedule = tf.keras.optimizers.schedules.ExponentialDecay(initial_learning_rate, decay_steps, decay_rate)
+
+ optimizer = tf.keras.optimizers.Adam(learning_rate=self.learningRateSchedule)
+
+ if TRACK_MOVING_AVERAGE:
+ # Not really sure this works properly (from tfa)
+ # optimizer = tfa.optimizers.MovingAverage(optimizer)
+ self.ema = tf.train.ExponentialMovingAverage(decay=0.9999)
+
+ self.optimizer = optimizer
+ self.initial_learning_rate = initial_learning_rate
+
+ def build_evaluation(self):
+ self.train_accuracy = tf.keras.metrics.MeanAbsoluteError(name='train_accuracy')
+ self.test_accuracy = tf.keras.metrics.MeanAbsoluteError(name='test_accuracy')
+ self.train_loss = tf.keras.metrics.Mean(name='train_loss')
+ self.test_loss = tf.keras.metrics.Mean(name='test_loss')
+
+ @tf.function(input_signature=[tf.TensorSpec(None, tf.float32), tf.TensorSpec(None, tf.float32)])
+ def train_step(self, inputs, labels):
+ with tf.GradientTape() as tape:
+ pred = self.model([inputs], training=True)
+ loss = self.loss(labels, pred)
+ total_loss = loss
+ if len(self.model.losses) > 0:
+ reg_loss = tf.math.add_n(self.model.losses)
+ total_loss = loss + reg_loss
+ gradients = tape.gradient(total_loss, self.model.trainable_variables)
+ self.optimizer.apply_gradients(zip(gradients, self.model.trainable_variables))
+ if TRACK_MOVING_AVERAGE:
+ self.ema.apply(self.model.trainable_variables)
+
+ self.train_loss(loss)
+ self.train_accuracy(labels, pred)
+
+ return loss
+
+ @tf.function(input_signature=[tf.TensorSpec(None, tf.float32), tf.TensorSpec(None, tf.float32)])
+ def test_step(self, inputs, labels):
+ pred = self.model([inputs], training=False)
+ t_loss = self.loss(labels, pred)
+
+ self.test_loss(t_loss)
+ self.test_accuracy(labels, pred)
+
+ # @tf.function(input_signature=[tf.TensorSpec(None, tf.float32), tf.TensorSpec(None, tf.float32)])
+ # decorator commented out because pred.numpy(): pred not evaluated yet.
+ def predict(self, inputs, labels):
+ pred = self.model([inputs], training=False)
+ # t_loss = self.loss(tf.squeeze(labels, axis=[3]), pred)
+ t_loss = self.loss(labels, pred)
+
+ self.test_labels.append(labels)
+ self.test_preds.append(pred.numpy())
+
+ self.test_loss(t_loss)
+ self.test_accuracy(labels, pred)
+
+ def reset_test_metrics(self):
+ self.test_loss.reset_states()
+ self.test_accuracy.reset_states()
+
+ def do_training(self, ckpt_dir=None):
+
+ if ckpt_dir is None:
+ if not os.path.exists(modeldir):
+ os.mkdir(modeldir)
+ ckpt = tf.train.Checkpoint(step=tf.Variable(1), model=self.model)
+ ckpt_manager = tf.train.CheckpointManager(ckpt, modeldir, max_to_keep=3)
+ else:
+ ckpt = tf.train.Checkpoint(step=tf.Variable(1), model=self.model)
+ ckpt_manager = tf.train.CheckpointManager(ckpt, ckpt_dir, max_to_keep=3)
+ ckpt.restore(ckpt_manager.latest_checkpoint)
+
+ self.writer_train = tf.summary.create_file_writer(os.path.join(logdir, 'plot_train'))
+ self.writer_valid = tf.summary.create_file_writer(os.path.join(logdir, 'plot_valid'))
+ self.writer_train_valid_loss = tf.summary.create_file_writer(os.path.join(logdir, 'plot_train_valid_loss'))
+
+ step = 0
+ total_time = 0
+ best_test_loss = np.finfo(dtype=np.float).max
+
+ if EARLY_STOP:
+ es = EarlyStop()
+
+ for epoch in range(NUM_EPOCHS):
+ self.train_loss.reset_states()
+ self.train_accuracy.reset_states()
+
+ t0 = datetime.datetime.now().timestamp()
+
+ proc_batch_cnt = 0
+ n_samples = 0
+
+ for data, label in self.train_dataset:
+ trn_ds = tf.data.Dataset.from_tensor_slices((data, label))
+ trn_ds = trn_ds.batch(BATCH_SIZE)
+ for mini_batch in trn_ds:
+ if self.learningRateSchedule is not None:
+ loss = self.train_step(mini_batch[0], mini_batch[1])
+
+ if (step % 100) == 0:
+
+ with self.writer_train.as_default():
+ tf.summary.scalar('loss_trn', loss.numpy(), step=step)
+ tf.summary.scalar('learning_rate', self.optimizer._decayed_lr('float32').numpy(), step=step)
+ tf.summary.scalar('num_train_steps', step, step=step)
+ tf.summary.scalar('num_epochs', epoch, step=step)
+
+ self.reset_test_metrics()
+ for data_tst, label_tst in self.test_dataset:
+ tst_ds = tf.data.Dataset.from_tensor_slices((data_tst, label_tst))
+ tst_ds = tst_ds.batch(BATCH_SIZE)
+ for mini_batch_test in tst_ds:
+ self.test_step(mini_batch_test[0], mini_batch_test[1])
+
+ with self.writer_valid.as_default():
+ tf.summary.scalar('loss_val', self.test_loss.result(), step=step)
+ tf.summary.scalar('acc_val', self.test_accuracy.result(), step=step)
+
+ with self.writer_train_valid_loss.as_default():
+ tf.summary.scalar('loss_trn', loss.numpy(), step=step)
+ tf.summary.scalar('loss_val', self.test_loss.result(), step=step)
+
+ print('****** test loss, acc, lr: ', self.test_loss.result().numpy(), self.test_accuracy.result().numpy(),
+ self.optimizer._decayed_lr('float32').numpy())
+
+ step += 1
+ print('train loss: ', loss.numpy())
+
+ proc_batch_cnt += 1
+ n_samples += data.shape[0]
+ print('proc_batch_cnt: ', proc_batch_cnt, n_samples)
+
+ t1 = datetime.datetime.now().timestamp()
+ print('End of Epoch: ', epoch+1, 'elapsed time: ', (t1-t0))
+ total_time += (t1-t0)
+
+ self.reset_test_metrics()
+ for data, label in self.test_dataset:
+ ds = tf.data.Dataset.from_tensor_slices((data, label))
+ ds = ds.batch(BATCH_SIZE)
+ for mini_batch in ds:
+ self.test_step(mini_batch[0], mini_batch[1])
+
+ print('loss, acc: ', self.test_loss.result().numpy(), self.test_accuracy.result().numpy())
+ print('------------------------------------------------------')
+
+ tst_loss = self.test_loss.result().numpy()
+ if tst_loss < best_test_loss:
+ best_test_loss = tst_loss
+ ckpt_manager.save()
+
+ if EARLY_STOP and es.check_stop(tst_loss):
+ break
+
+ print('total time: ', total_time)
+ self.writer_train.close()
+ self.writer_valid.close()
+ self.writer_train_valid_loss.close()
+
+ def build_model(self):
+ self.build_srcnn()
+ self.model = tf.keras.Model(self.inputs, self.logits)
+
+ def restore(self, ckpt_dir):
+
+ ckpt = tf.train.Checkpoint(step=tf.Variable(1), model=self.model)
+ ckpt_manager = tf.train.CheckpointManager(ckpt, ckpt_dir, max_to_keep=3)
+ ckpt.restore(ckpt_manager.latest_checkpoint)
+
+ self.reset_test_metrics()
+
+ for data, label in self.test_dataset:
+ ds = tf.data.Dataset.from_tensor_slices((data, label))
+ ds = ds.batch(BATCH_SIZE)
+ for mini_batch_test in ds:
+ self.predict(mini_batch_test[0], mini_batch_test[1])
+
+ print('loss, acc: ', self.test_loss.result().numpy(), self.test_accuracy.result().numpy())
+
+ labels = np.concatenate(self.test_labels)
+ preds = np.concatenate(self.test_preds)
+ print(labels.shape, preds.shape)
+
+ labels_denorm = denormalize(labels, label_param, mean_std_dct)
+ preds_denorm = denormalize(preds, label_param, mean_std_dct)
+ # labels_denorm = descale(labels, label_param, mean_std_dct)
+ # preds_denorm = descale(preds, label_param, mean_std_dct)
+
+ return labels_denorm, preds_denorm
+
+ def do_evaluate(self, inputs, ckpt_dir):
+
+ ckpt = tf.train.Checkpoint(step=tf.Variable(1), model=self.model)
+ ckpt_manager = tf.train.CheckpointManager(ckpt, ckpt_dir, max_to_keep=3)
+ ckpt.restore(ckpt_manager.latest_checkpoint)
+
+ self.reset_test_metrics()
+
+ pred = self.model([inputs], training=False)
+ self.test_probs = pred
+ pred = pred.numpy()
+
+ return pred
+
+ def run(self, directory, ckpt_dir=None, num_data_samples=50000):
+ train_data_files = glob.glob(directory+'train*mres*.npy')
+ valid_data_files = glob.glob(directory+'valid*mres*.npy')
+ train_label_files = [f.replace('mres', 'ires') for f in train_data_files]
+ valid_label_files = [f.replace('mres', 'ires') for f in valid_data_files]
+ self.setup_pipeline(train_data_files, train_label_files, valid_data_files, valid_label_files, num_data_samples)
+
+ self.build_model()
+ self.build_training()
+ self.build_evaluation()
+ self.do_training(ckpt_dir=ckpt_dir)
+
+ def run_restore(self, directory, ckpt_dir):
+ self.num_data_samples = 1000
+
+ valid_data_files = glob.glob(directory + 'valid*mres*.npy')
+ valid_label_files = [f.replace('mres', 'ires') for f in valid_data_files]
+ self.setup_test_pipeline(valid_data_files, valid_label_files)
+
+ self.build_model()
+ self.build_training()
+ self.build_evaluation()
+ return self.restore(ckpt_dir)
+
+ def run_evaluate(self, data, ckpt_dir):
+ # data = tf.convert_to_tensor(data, dtype=tf.float32)
+ self.num_data_samples = 80000
+ self.build_model()
+ self.build_training()
+ self.build_evaluation()
+ return self.do_evaluate(data, ckpt_dir)
+
+ def setup_inference(self, ckpt_dir):
+ self.num_data_samples = 80000
+ self.build_model()
+ self.build_training()
+ self.build_evaluation()
+
+ ckpt = tf.train.Checkpoint(step=tf.Variable(1), model=self.model)
+ ckpt_manager = tf.train.CheckpointManager(ckpt, ckpt_dir, max_to_keep=3)
+ ckpt.restore(ckpt_manager.latest_checkpoint)
+
+ def do_inference(self, inputs):
+ self.reset_test_metrics()
+
+ pred = self.model([inputs], training=False)
+ self.test_probs = pred
+ pred = pred.numpy()
+
+ return pred
+
+ def run_inference(self, in_file, out_file):
+ gc.collect()
+ t0 = time.time()
+
+ h5f = h5py.File(in_file, 'r')
+
+ refl = get_grid_values_all(h5f, 'refl_0_65um_nom')
+ LEN_Y, LEN_X = refl.shape
+ bt = get_grid_values_all(h5f, 'temp_11_0um_nom')
+ cld_opd = get_grid_values_all(h5f, 'cld_opd_dcomp')
+ refl_sub_lo = get_grid_values_all(h5f, 'refl_0_65um_nom_min_sub')
+ refl_sub_hi = get_grid_values_all(h5f, 'refl_0_65um_nom_max_sub')
+ refl_sub_std = get_grid_values_all(h5f, 'refl_0_65um_nom_stddev_sub')
+ t1 = time.time()
+ print('read data time: ', (t1 - t0))
+
+ cld_opd_sres = self.run_inference_(bt, refl, refl_sub_lo, refl_sub_hi, refl_sub_std, cld_opd, 2*LEN_Y, 2*LEN_X)
+
+ cld_opd_sres_out = np.zeros((LEN_Y, LEN_X), dtype=np.int8)
+ border = int((KERNEL_SIZE - 1) / 2)
+ cld_opd_sres_out[border:LEN_Y - border, border:LEN_X - border] = cld_opd_sres[0, :, :]
+
+ h5f.close()
+
+ if out_file is not None:
+ np.save(out_file, (cld_opd_sres, bt, refl, cld_opd))
+ else:
+ return cld_opd_sres
+
+ def run_inference_(self, bt, refl, refl_sub_lo, refl_sub_hi, refl_sub_std, cld_opd, LEN_Y, LEN_X):
+
+ self.slc_x_m = slice(1, int(LEN_X / 2) + 4)
+ self.slc_y_m = slice(1, int(LEN_Y / 2) + 4)
+ self.slc_x = slice(3, LEN_X + 5)
+ self.slc_y = slice(3, LEN_Y + 5)
+ self.slc_x_2 = slice(2, LEN_X + 7, 2)
+ self.slc_y_2 = slice(2, LEN_Y + 7, 2)
+ self.x_2 = np.arange(int(LEN_X / 2) + 3)
+ self.y_2 = np.arange(int(LEN_Y / 2) + 3)
+ self.t = np.arange(0, int(LEN_X / 2) + 3, 0.5)
+ self.s = np.arange(0, int(LEN_Y / 2) + 3, 0.5)
+ self.x_k = slice(1, LEN_X + 3)
+ self.y_k = slice(1, LEN_Y + 3)
+
+ t0 = time.time()
+ bt = np.where(np.isnan(bt), 0, bt)
+ bt = bt[self.slc_y_m, self.slc_x_m]
+ bt = np.expand_dims(bt, axis=0)
+ # bt_us = upsample_static(bt, x_2, y_2, t, s, None, None)
+ bt_us = self.upsample(bt)
+ bt_us = smooth_2d(bt_us)
+ bt_us = normalize(bt_us, 'temp_11_0um_nom', mean_std_dct)
+
+ refl = np.where(np.isnan(refl), 0, refl)
+ refl = refl[self.slc_y_m, self.slc_x_m]
+ refl = np.expand_dims(refl, axis=0)
+ refl_us = self.upsample(refl)
+ refl_us = smooth_2d(refl_us)
+ refl_us = normalize(refl_us, 'refl_0_65um_nom', mean_std_dct)
+
+ cld_opd = np.where(np.isnan(cld_opd), 0, cld_opd)
+ cld_opd = cld_opd[self.slc_y_m, self.slc_x_m]
+ cld_opd = np.expand_dims(cld_opd, axis=0)
+ # cld_opd_us = upsample_static(cld_opd, x_2, y_2, t, s, None, None)
+ cld_opd_us = self.upsample(cld_opd)
+ cld_opd_us = smooth_2d(cld_opd_us)
+ cld_opd_us = normalize(cld_opd_us, label_param, mean_std_dct)
+
+ refl_sub_lo = np.expand_dims(refl_sub_lo, axis=0)
+ refl_sub_lo = upsample_nearest(refl_sub_lo)
+ refl_sub_lo = refl_sub_lo[self.slc_y, self.slc_x]
+ refl_sub_lo = normalize(refl_sub_lo, 'refl_0_65um_nom', mean_std_dct)
+
+ refl_sub_hi = np.expand_dims(refl_sub_hi, axis=0)
+ refl_sub_hi = upsample_nearest(refl_sub_hi)
+ refl_sub_hi = refl_sub_hi[self.slc_y, self.slc_x]
+ refl_sub_hi = normalize(refl_sub_hi, 'refl_0_65um_nom', mean_std_dct)
+
+ refl_sub_std = np.expand_dims(refl_sub_std, axis=0)
+ refl_sub_std = upsample_nearest(refl_sub_std)
+ refl_sub_std = refl_sub_std[self.slc_y, self.slc_x]
+
+ t1 = time.time()
+ print('upsample/normalize time: ', (t1 - t0))
+
+ data = np.stack([bt_us, refl_us, refl_sub_lo, refl_sub_hi, refl_sub_std, cld_opd_us], axis=3)
+
+ cld_opd_sres = self.do_inference(data)
+ cld_opd_sres = denormalize(cld_opd_sres, label_param, mean_std_dct)
+
+ return cld_opd_sres
+
+
+def run_restore_static(directory, ckpt_dir, out_file=None):
+ nn = SRCNN()
+ labels_denorm, preds_denorm = nn.run_restore(directory, ckpt_dir)
+ if out_file is not None:
+ np.save(out_file, [labels_denorm, preds_denorm])
+
+
+def run_evaluate_static(in_file, out_file, ckpt_dir):
+
+ h5f = h5py.File(in_file, 'r')
+
+ refl = get_grid_values_all(h5f, 'refl_0_65um_nom')
+ LEN_Y, LEN_X = refl.shape
+ print(LEN_Y, LEN_X)
+
+ bt = get_grid_values_all(h5f, 'temp_11_0um_nom')
+
+ cld_opd = get_grid_values_all(h5f, 'cld_opd_dcomp_1')
+
+ refl_sub_lo = get_grid_values_all(h5f, 'refl_0_65um_nom_min_sub')
+ refl_sub_hi = get_grid_values_all(h5f, 'refl_0_65um_nom_max_sub')
+ refl_sub_std = get_grid_values_all(h5f, 'refl_0_65um_nom_stddev_sub')
+
+ nn = SRCNN()
+
+ slc_x = slice(0, (LEN_X - 16) + 4)
+ slc_y = slice(0, (LEN_Y - 16) + 4)
+ x_2 = np.arange((LEN_X - 16) + 4)
+ y_2 = np.arange((LEN_Y - 16) + 4)
+ t = np.arange(0, (LEN_X - 16) + 4, 0.5)
+ s = np.arange(0, (LEN_Y - 16) + 4, 0.5)
+
+ refl = np.where(np.isnan(refl), 0, bt)
+ refl = refl[slc_y, slc_x]
+ refl = np.expand_dims(refl, axis=0)
+ refl_us = upsample_static(refl, x_2, y_2, t, s, None, None)
+ print(refl_us.shape)
+ refl_us = normalize(refl_us, 'refl_0_65um_nom', mean_std_dct)
+ print('REFL done')
+
+ bt = np.where(np.isnan(bt), 0, bt)
+ bt = bt[slc_y, slc_x]
+ bt = np.expand_dims(bt, axis=0)
+ bt_us = upsample_static(bt, x_2, y_2, t, s, None, None)
+ bt_us = normalize(bt_us, 'temp_11_0um_nom', mean_std_dct)
+ print('BT done')
+
+ refl_sub_lo = refl_sub_lo[slc_y, slc_x]
+ refl_sub_lo = np.expand_dims(refl_sub_lo, axis=0)
+ refl_sub_lo = upsample_nearest(refl_sub_lo)
+ refl_sub_lo = normalize(refl_sub_lo, 'refl_0_65um_nom', mean_std_dct)
+
+ refl_sub_hi = refl_sub_hi[slc_y, slc_x]
+ refl_sub_hi = np.expand_dims(refl_sub_hi, axis=0)
+ refl_sub_hi = upsample_nearest(refl_sub_hi)
+ refl_sub_hi = normalize(refl_sub_hi, 'refl_0_65um_nom', mean_std_dct)
+
+ refl_sub_std = refl_sub_std[slc_y, slc_x]
+ refl_sub_std = np.expand_dims(refl_sub_std, axis=0)
+ refl_sub_std = upsample_nearest(refl_sub_std)
+
+ cld_opd = np.where(np.isnan(cld_opd), 0, cld_opd)
+ cld_opd = cld_opd[slc_y, slc_x]
+ cld_opd = np.expand_dims(cld_opd, axis=0)
+ cld_opd_us = upsample_static(cld_opd, x_2, y_2, t, s, None, None)
+ cld_opd_us = normalize(cld_opd_us, label_param, mean_std_dct)
+ print('OPD done')
+
+ # data = np.stack([bt_us, refl_us, refl_sub_lo, refl_sub_hi, refl_sub_std, cld_opd_us], axis=3)
+ data = np.stack([bt_us, refl_us, cld_opd_us], axis=3)
+ print('INPUT: ', data.shape)
+
+ cld_opd_sres = nn.run_evaluate(data, ckpt_dir)
+ # cld_opd_sres = descale(cld_opd_sres, label_param, mean_std_dct)
+ cld_opd_sres = denormalize(cld_opd_sres, label_param, mean_std_dct)
+ _, ylen, xlen, _ = cld_opd_sres.shape
+ print('OUT: ', ylen, xlen)
+
+ cld_opd_sres_out = np.zeros((2*LEN_Y, 2*LEN_X), dtype=np.float32)
+ refl_out = np.zeros((LEN_Y, LEN_X), dtype=np.float32)
+ cld_opd_out = np.zeros((LEN_Y, LEN_X), dtype=np.float32)
+
+ border = int((KERNEL_SIZE - 1) / 2)
+ cld_opd_sres_out[border:(border+ylen), border:(border+xlen)] = cld_opd_sres[0, :, :, 0]
+ # refl_out[0:(ylen+2*border), 0:(xlen+2*border)] = refl[0, :, :]
+ # cld_opd_out[0:(ylen+2*border), 0:(xlen+2*border)] = cld_opd[0, :, :]
+
+ # refl_out = denormalize(refl_out, 'refl_0_65um_nom', mean_std_dct)
+ # cld_opd_out = denormalize(cld_opd_out, label_param, mean_std_dct)
+
+ h5f.close()
+
+ if out_file is not None:
+ # np.save(out_file, (cld_opd_sres_out, refl_out, cld_opd_out, cld_opd_hres))
+ np.save(out_file, cld_opd_sres_out)
+ else:
+ return cld_opd_sres_out, bt, refl
+
+
+if __name__ == "__main__":
+ nn = SRCNN()
+ nn.run('matchup_filename')
--
GitLab