Skip to content
Snippets Groups Projects
srcnn.py 23.36 KiB
import glob
import tensorflow as tf
from util.setup import logdir, modeldir, cachepath, now, ancillary_path
from util.util import EarlyStop, normalize, denormalize, resample, resample_one, resample_2d_linear, resample_2d_linear_one, get_grid_values_all
import os, datetime
import numpy as np
import pickle
import h5py

# L1B M/I-bands: /apollo/cloud/scratch/cwhite/VIIRS_HRES/2019/2019_01_01/
# CLAVRx: /apollo/cloud/scratch/Satellite_Output/VIIRS_HRES/2019/2019_01_01/
# /apollo/cloud/scratch/Satellite_Output/andi/NEW/VIIRS_HRES/2019

LOG_DEVICE_PLACEMENT = False

PROC_BATCH_SIZE = 2
PROC_BATCH_BUFFER_SIZE = 50000

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 = 0.10
DO_AUGMENT = True

# 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)

# emis_params = ['temp_10_4um_nom', 'temp_11_0um_nom', 'temp_12_0um_nom', 'temp_13_3um_nom', 'temp_3_75um_nom',
#                'temp_6_7um_nom', 'temp_6_2um_nom', 'temp_7_3um_nom', 'temp_8_5um_nom', 'temp_9_7um_nom']

data_params = ['temp_11_0um_nom', 'temp_12_0um_nom', 'cloud_fraction']
label_params = ['temp_11_0um_nom', 'temp_12_0um_nom', 'cloud_fraction']


DO_ZERO_OUT = False

data_idx, label_idx = 0, 0
data_param = data_params[data_idx]
label_param = label_params[label_idx]
print(data_param+', '+label_param)

x_134 = np.arange(134)
y_134 = np.arange(134)
x_64 = np.arange(64)
y_64 = np.arange(64)
x_134_2 = x_134[3:131:2]
y_134_2 = y_134[3:131:2]
#x_134_2 = x_134[2:133:2]
#y_134_2 = y_134[2:133:2]
t = np.arange(0, 64, 0.5)
s = np.arange(0, 64, 0.5)


def build_residual_conv2d_block(conv, num_filters, block_name, activation=tf.nn.relu, padding='SAME', kernel_initializer='he_uniform', scale=None):

    with tf.name_scope(block_name):
        skip = tf.keras.layers.Conv2D(num_filters, kernel_size=3, padding=padding, kernel_initializer=kernel_initializer, activation=activation)(conv)
        skip = tf.keras.layers.Conv2D(num_filters, kernel_size=3, padding=padding, activation=None)(skip)
        if scale is not None:
            skip = tf.keras.layers.Lambda(lambda x: x * scale)(skip)
        conv = conv + skip
        print(block_name+':', conv.shape)

    return conv


class SRCNN:
    
    def __init__(self):

        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.X_prof = None
        self.X_u = None
        self.X_v = None
        self.X_sfc = None
        self.inputs = []
        self.y = None
        self.handle = None
        self.inner_handle = None
        self.in_mem_batch = None

        self.h5f_l1b_trn = None
        self.h5f_l1b_tst = None
        self.h5f_l2_trn = None
        self.h5f_l2_tst = None

        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.OUT_OF_RANGE = False

        self.abi = None
        self.temp = None
        self.wv = None
        self.lbfp = None
        self.sfc = None

        self.in_mem_data_cache = {}
        self.in_mem_data_cache_test = {}

        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_auc = None
        self.test_recall = None
        self.test_precision = None
        self.test_confusion_matrix = None
        self.test_true_pos = None
        self.test_true_neg = None
        self.test_false_pos = None
        self.test_false_neg = 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.data_dct = None
        self.train_data_files = None
        self.train_label_files = None
        self.test_data_files = None
        self.test_label_files = None

        self.train_data_nda = None
        self.train_label_nda = None
        self.test_data_nda = None
        self.test_label_nda = None

        self.n_chans = 1

        self.X_img = tf.keras.Input(shape=(None, None, self.n_chans))
        # self.X_img = tf.keras.Input(shape=(36, 36, self.n_chans))
        # self.X_img = tf.keras.Input(shape=(34, 34, self.n_chans))
        # self.X_img = tf.keras.Input(shape=(66, 66, self.n_chans))

        self.inputs.append(self.X_img)

        tf.debugging.set_log_device_placement(LOG_DEVICE_PLACEMENT)

    def get_in_mem_data_batch(self, idxs, is_training):
        if is_training:
            files = self.train_data_files
        else:
            files = self.test_data_files

        label_s = []
        for k in idxs:
            f = files[k]
            nda = np.load(f)
            label_s.append(nda)

        data = np.concatenate(label_s)
        label = data.copy()

        data = data[:, data_idx, 3:131:2, 3:131:2]
        data = resample(y_64, x_64, data, s, t)
        data = np.expand_dims(data, axis=3)

        # label = label[:, label_idx, 3:131:2, 3:131:2]
        label = label[:, label_idx, 3:131, 3:131]
        label = np.expand_dims(label, axis=3)

        data = data.astype(np.float32)
        label = label.astype(np.float32)

        if data_param != 'cloud_fraction':
            data = normalize(data, data_param, mean_std_dct, add_noise=True, noise_scale=0.005)
        if label_param != 'cloud_fraction':
            label = normalize(label, label_param, mean_std_dct)

        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)

    def get_in_mem_data_batch_eval(self, idxs):
        data = []
        for param in self.train_params:
            nda = self.data_dct[param]
            nda = normalize(nda, param, mean_std_dct)
            data.append(nda)
        data = np.stack(data)
        data = data.astype(np.float32)
        data = np.transpose(data, axes=(1, 2, 0))
        data = np.expand_dims(data, axis=0)

        return data

    @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

    @tf.function(input_signature=[tf.TensorSpec(None, tf.int32)])
    def data_function_evaluate(self, indexes):
        # TODO: modify for user specified altitude
        out = tf.numpy_function(self.get_in_mem_data_batch_eval, [indexes], [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 get_evaluate_dataset(self, indexes):
        indexes = list(indexes)

        dataset = tf.data.Dataset.from_tensor_slices(indexes)
        dataset = dataset.map(self.data_function_evaluate, num_parallel_calls=8)
        self.eval_dataset = dataset

    def setup_pipeline(self, train_data_files, test_data_files, num_train_samples):

        self.train_data_files = train_data_files
        self.test_data_files = test_data_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):
        self.test_data_files = test_data_files
        tst_idxs = np.arange(len(test_data_files))
        self.get_test_dataset(tst_idxs)
        print('setup_test_pipeline: Done')

    def setup_eval_pipeline(self, filename):
        idxs = [0]
        self.num_data_samples = idxs.shape[0]

        self.get_evaluate_dataset(idxs)

    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]
        print('input: ', input_2d.shape)
        # conv = tf.keras.layers.Conv2D(num_filters, kernel_size=5, strides=1, padding='VALID', activation=None)(input_2d)
        conv = input_2d
        print('input: ', conv.shape)

        # conv = conv_b = tf.keras.layers.Conv2D(num_filters, kernel_size=3, padding=padding)(input_2d)
        conv = conv_b = tf.keras.layers.Conv2D(num_filters, kernel_size=3, kernel_initializer='he_uniform', activation=activation, padding='SAME')(input_2d)
        print(conv.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', scale=scale)

        conv_b = build_residual_conv2d_block(conv_b, num_filters, 'Residual_Block_2', scale=scale)

        conv_b = build_residual_conv2d_block(conv_b, num_filters, 'Residual_Block_3', scale=scale)

        conv_b = build_residual_conv2d_block(conv_b, num_filters, 'Residual_Block_4', scale=scale)

        conv_b = build_residual_conv2d_block(conv_b, num_filters, 'Residual_Block_5', scale=scale)

        conv_b = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, kernel_initializer='he_uniform', padding=padding)(conv_b)

        conv = conv + conv_b
        print(conv.shape)

        self.logits = tf.keras.layers.Conv2D(1, kernel_size=3, strides=1, padding=padding, name='regression')(conv)

        print(self.logits.shape)

    def build_training(self):
        # if NumClasses == 2:
        #     self.loss = tf.keras.losses.BinaryCrossentropy(from_logits=False)  # for two-class only
        # else:
        #     self.loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)  # For multi-class
        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 / 2)
        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
    def train_step(self, mini_batch):
        inputs = [mini_batch[0]]
        labels = mini_batch[1]
        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
    def test_step(self, mini_batch):
        inputs = [mini_batch[0]]
        labels = mini_batch[1]
        pred = self.model(inputs, training=False)
        t_loss = self.loss(labels, pred)

        self.test_loss(t_loss)
        self.test_accuracy(labels, pred)

    def predict(self, mini_batch):
        inputs = [mini_batch[0]]
        labels = mini_batch[1]
        pred = self.model(inputs, training=False)
        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 get_metrics(self):
        recall = self.test_recall.result()
        precsn = self.test_precision.result()
        f1 = 2 * (precsn * recall) / (precsn + recall)

        tn = self.test_true_neg.result()
        tp = self.test_true_pos.result()
        fn = self.test_false_neg.result()
        fp = self.test_false_pos.result()

        mcc = ((tp * tn) - (fp * fn)) / np.sqrt((tp + fp) * (tp + fn) * (tn + fp) * (tn + fn))
        return f1, mcc

    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)

        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)

                    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)

                        # if NumClasses == 2:
                        #     f1, mcc = self.get_metrics()

                        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)

            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()

        # f = open(home_dir+'/best_stats_'+now+'.pkl', 'wb')
        # pickle.dump((best_test_loss, best_test_acc, best_test_recall, best_test_precision, best_test_auc, best_test_f1, best_test_mcc), f)
        # f.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)

        print('loss, acc: ', self.test_loss.result().numpy(), self.test_accuracy.result().numpy())

    def do_evaluate(self, nda_lr, param, 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)

        data = normalize(nda_lr, param, mean_std_dct)
        data = np.expand_dims(data, axis=0)
        data = np.expand_dims(data, axis=3)

        self.reset_test_metrics()

        pred = self.model([data], training=False)
        self.test_probs = pred
        pred = pred.numpy()

        return denormalize(pred, param, mean_std_dct)

    def run(self, directory):
        train_data_files = glob.glob(directory+'data_train*.npy')
        valid_data_files = glob.glob(directory+'data_valid*.npy')

        self.setup_pipeline(train_data_files, valid_data_files, 50000)
        self.build_model()
        self.build_training()
        self.build_evaluation()
        self.do_training()

    def run_restore(self, directory, ckpt_dir):
        valid_data_files = glob.glob(directory + 'data_valid*.npy')
        self.num_data_samples = 1000
        self.setup_test_pipeline(valid_data_files)
        self.build_model()
        self.build_training()
        self.build_evaluation()
        self.restore(ckpt_dir)

    def run_evaluate(self, nda_lr, param, ckpt_dir):
        self.num_data_samples = 80000
        self.build_model()
        self.build_training()
        self.build_evaluation()
        return self.do_evaluate(nda_lr, param, ckpt_dir)


def run_evaluate_static(in_file, out_file, ckpt_dir):
    nda = np.load(in_file)

    nda = nda[:, data_idx, 3:131:2, 3:131:2]
    nda = resample(y_64, x_64, nda, s, t)
    nda = np.expand_dims(nda, axis=3)

    nn = SRCNN()
    out_sr = nn.run_evaluate(nda, data_param, ckpt_dir)
    if out_file is not None:
        np.save(out_file, out_sr)
    else:
        return out_sr


def run_evaluate_static_new(in_file, out_file, ckpt_dir):
    h5f = h5py.File(in_file, 'r')
    grd = get_grid_values_all(h5f, data_param)
    grd = grd[::2, ::2]
    leny, lenx = grd.shape
    x = np.arange(lenx)
    y = np.arange(leny)
    x_up = np.arange(0, lenx, 0.5)
    y_up = np.arange(0, leny, 0.5)

    grd = resample_one(y, x, grd, y_up, x_up)

    nn = SRCNN()
    out_sr = nn.run_evaluate(grd, data_param, ckpt_dir)
    if out_file is not None:
        np.save(out_file, out_sr)
    else:
        return out_sr


if __name__ == "__main__":
    nn = SRCNN()
    nn.run('matchup_filename')