unet.py 40.08 KiB
import glob
import tensorflow as tf
from util.setup import logdir, modeldir, cachepath, now, ancillary_path, home_dir
from util.util import EarlyStop, normalize, make_for_full_domain_predict
import matplotlib.pyplot as plt
import os, datetime
import numpy as np
import pickle
import h5py
LOG_DEVICE_PLACEMENT = False
PROC_BATCH_SIZE = 1024
PROC_BATCH_BUFFER_SIZE = 50000
NumClasses = 2
if NumClasses == 2:
NumLogits = 1
else:
NumLogits = NumClasses
BATCH_SIZE = 128
NUM_EPOCHS = 100
TRACK_MOVING_AVERAGE = False
EARLY_STOP = True
TRIPLET = False
CONV3D = False
NOISE_TRAINING = True
NOISE_STDDEV = 0.10
DO_AUGMENT = True
img_width = 16
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)
# -- NIGHT L2 -----------------------------
train_params_l2_night = ['cld_height_acha', 'cld_geo_thick', 'cld_temp_acha', 'cld_press_acha', 'supercooled_cloud_fraction',
'cld_emiss_acha', 'conv_cloud_fraction', 'cld_reff_acha', 'cld_opd_acha']
# -- DAY L2 --------------------------------
train_params_l2_day = ['cld_height_acha', 'cld_geo_thick', 'cld_temp_acha', 'cld_press_acha', 'supercooled_cloud_fraction',
'cld_emiss_acha', 'conv_cloud_fraction', 'cld_reff_dcomp', 'cld_opd_dcomp', 'iwc_dcomp', 'lwc_dcomp']
# -- DAY L1B --------------------------------
train_params_l1b_day = ['temp_10_4um_nom', 'temp_11_0um_nom', 'temp_12_0um_nom', 'temp_13_3um_nom', 'temp_3_75um_nom',
'temp_6_2um_nom', 'temp_6_7um_nom', 'temp_7_3um_nom', 'temp_8_5um_nom', 'temp_9_7um_nom',
'refl_0_47um_nom', 'refl_0_65um_nom', 'refl_0_86um_nom', 'refl_1_38um_nom', 'refl_1_60um_nom']
# -- NIGHT L1B -------------------------------
train_params_l1b_night = ['temp_10_4um_nom', 'temp_11_0um_nom', 'temp_12_0um_nom', 'temp_13_3um_nom', 'temp_3_75um_nom',
'temp_6_2um_nom', 'temp_6_7um_nom', 'temp_7_3um_nom', 'temp_8_5um_nom', 'temp_9_7um_nom']
# -- DAY LUNAR ---------------------------------
# train_params_l1b = ['cld_height_acha', 'cld_geo_thick', 'cld_temp_acha', 'cld_press_acha', 'supercooled_cloud_fraction',
# 'cld_emiss_acha', 'conv_cloud_fraction', 'cld_reff_dcomp', 'cld_opd_dcomp', 'iwc_dcomp', 'lwc_dcomp']
# ---------------------------------------------
train_params = train_params_l1b_day + train_params_l2_day
# -- Zero out params (Experimentation Only) ------------
zero_out_params = ['cld_reff_dcomp', 'cld_opd_dcomp', 'iwc_dcomp', 'lwc_dcomp']
DO_ZERO_OUT = False
def build_conv2d_block(conv, num_filters, activation, block_name, padding='SAME'):
with tf.name_scope(block_name):
skip = conv
conv = tf.keras.layers.Conv2D(num_filters, kernel_size=5, strides=1, padding=padding, activation=activation)(conv)
conv = tf.keras.layers.MaxPool2D(padding=padding)(conv)
conv = tf.keras.layers.BatchNormalization()(conv)
print(conv.shape)
skip = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, padding=padding, activation=None)(skip)
skip = tf.keras.layers.MaxPool2D(padding=padding)(skip)
skip = tf.keras.layers.BatchNormalization()(skip)
conv = conv + skip
conv = tf.keras.layers.LeakyReLU()(conv)
print(conv.shape)
return conv
class UNET:
def __init__(self, day_night='DAY', l1b_or_l2='both', use_flight_altitude=False, gpu_device=0, datapath=None):
if day_night == 'DAY':
self.train_params_l1b = train_params_l1b_day
self.train_params_l2 = train_params_l2_day
if l1b_or_l2 == 'both':
self.train_params = train_params_l1b_day + train_params_l2_day
elif l1b_or_l2 == 'l1b':
self.train_params = train_params_l1b_day
elif l1b_or_l2 == 'l2':
self.train_params = train_params_l2_day
else:
self.train_params_l1b = train_params_l1b_night
self.train_params_l2 = train_params_l2_night
if l1b_or_l2 == 'both':
self.train_params = train_params_l1b_night + train_params_l2_night
elif l1b_or_l2 == 'l1b':
self.train_params = train_params_l1b_night
elif l1b_or_l2 == 'l2':
self.train_params = train_params_l2_night
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.gpu_device = gpu_device
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
# n_chans = len(self.train_params)
n_chans = 3
if TRIPLET:
n_chans *= 3
self.X_img = tf.keras.Input(shape=(None, None, n_chans))
self.inputs.append(self.X_img)
#self.inputs.append(tf.keras.Input(shape=(None, None, 5)))
self.inputs.append(tf.keras.Input(shape=(None, None, 3)))
self.flight_level = 0
self.DISK_CACHE = False
self.USE_FLIGHT_ALTITUDE = use_flight_altitude
if datapath is not None:
self.DISK_CACHE = False
f = open(datapath, 'rb')
self.in_mem_data_cache = pickle.load(f)
f.close()
tf.debugging.set_log_device_placement(LOG_DEVICE_PLACEMENT)
# Doesn't seem to play well with SLURM
# gpus = tf.config.experimental.list_physical_devices('GPU')
# if gpus:
# try:
# # Currently, memory growth needs to be the same across GPUs
# for gpu in gpus:
# tf.config.experimental.set_memory_growth(gpu, True)
# logical_gpus = tf.config.experimental.list_logical_devices('GPU')
# print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")
# except RuntimeError as e:
# # Memory growth must be set before GPUs have been initialized
# print(e)
# def get_in_mem_data_batch(self, idxs, is_training):
#
# # sort these to use as numpy indexing arrays
# nd_idxs = np.array(idxs)
# nd_idxs = np.sort(nd_idxs)
#
# data = []
# for param in self.train_params:
# nda = self.get_parameter_data(param, nd_idxs, is_training)
# nda = normalize(nda, param, mean_std_dct)
# if DO_ZERO_OUT and is_training:
# try:
# zero_out_params.index(param)
# nda[:,] = 0.0
# except ValueError:
# pass
# data.append(nda)
# data = np.stack(data)
# data = data.astype(np.float32)
# data = np.transpose(data, axes=(1, 2, 3, 0))
#
# data_alt = self.get_scalar_data(nd_idxs, is_training)
#
# label = self.get_label_data(nd_idxs, is_training)
# label = np.where(label == -1, 0, label)
#
# # binary, two class
# if NumClasses == 2:
# label = np.where(label != 0, 1, label)
# label = label.reshape((label.shape[0], 1))
# elif NumClasses == 3:
# label = np.where(np.logical_or(label == 1, label == 2), 1, label)
# label = np.where(np.invert(np.logical_or(label == 0, label == 1)), 2, label)
# label = label.reshape((label.shape[0], 1))
#
# if is_training and DO_AUGMENT:
# data_ud = np.flip(data, axis=1)
# data_alt_ud = np.copy(data_alt)
# label_ud = np.copy(label)
#
# data_lr = np.flip(data, axis=2)
# data_alt_lr = np.copy(data_alt)
# label_lr = np.copy(label)
#
# data = np.concatenate([data, data_ud, data_lr])
# data_alt = np.concatenate([data_alt, data_alt_ud, data_alt_lr])
# label = np.concatenate([label, label_ud, label_lr])
#
# return data, data_alt, label
def get_in_mem_data_batch(self, idxs, is_training):
data = []
for k in idxs:
if is_training:
nda = plt.imread(self.train_data_files[k])
else:
nda = plt.imread(self.test_data_files[k])
data.append(nda[0:64, 0:64])
data = np.stack(data)
data = data.astype(np.float32)
label = []
for k in idxs:
if is_training:
nda = plt.imread(self.train_label_files[k])
else:
nda = plt.imread(self.test_label_files[k])
label.append(nda[0:64, 0:64])
label = np.stack(label)
label = label.astype(np.int32)
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, data, label
# def get_parameter_data(self, param, nd_idxs, is_training):
# if is_training:
# if param in self.train_params_l1b:
# h5f = self.h5f_l1b_trn
# else:
# h5f = self.h5f_l2_trn
# else:
# if param in self.train_params_l1b:
# h5f = self.h5f_l1b_tst
# else:
# h5f = self.h5f_l2_tst
#
# nda = h5f[param][nd_idxs,]
# return nda
#
# def get_label_data(self, nd_idxs, is_training):
# # Note: labels will be same for nd_idxs across both L1B and L2
# if is_training:
# if self.h5f_l1b_trn is not None:
# h5f = self.h5f_l1b_trn
# else:
# h5f = self.h5f_l2_trn
# else:
# if self.h5f_l1b_tst is not None:
# h5f = self.h5f_l1b_tst
# else:
# h5f = self.h5f_l2_tst
#
# label = h5f['icing_intensity'][nd_idxs]
# label = label.astype(np.int32)
# return 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)
nda = np.zeros([1])
nda[0] = self.flight_level
nda = tf.one_hot(nda, 5).numpy()
nda = np.expand_dims(nda, axis=0)
nda = np.expand_dims(nda, axis=0)
return data, nda
@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, tf.int32])
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, tf.int32])
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, 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, filename_l1b_trn, filename_l1b_tst, filename_l2_trn, filename_l2_tst, trn_idxs=None, tst_idxs=None, seed=None):
if filename_l1b_trn is not None:
self.h5f_l1b_trn = h5py.File(filename_l1b_trn, 'r')
if filename_l1b_tst is not None:
self.h5f_l1b_tst = h5py.File(filename_l1b_tst, 'r')
if filename_l2_trn is not None:
self.h5f_l2_trn = h5py.File(filename_l2_trn, 'r')
if filename_l2_tst is not None:
self.h5f_l2_tst = h5py.File(filename_l2_tst, 'r')
if trn_idxs is None:
# Note: time is same across both L1B and L2 for idxs
if self.h5f_l1b_trn is not None:
h5f = self.h5f_l1b_trn
else:
h5f = self.h5f_l2_trn
time = h5f['time']
trn_idxs = np.arange(time.shape[0])
if seed is not None:
np.random.seed(seed)
np.random.shuffle(trn_idxs)
if self.h5f_l1b_tst is not None:
h5f = self.h5f_l1b_tst
else:
h5f = self.h5f_l2_tst
time = h5f['time']
tst_idxs = np.arange(time.shape[0])
if seed is not None:
np.random.seed(seed)
np.random.shuffle(tst_idxs)
self.num_data_samples = trn_idxs.shape[0]
self.get_train_dataset(trn_idxs)
self.get_test_dataset(tst_idxs)
print('datetime: ', now)
print('training and test data: ')
print(filename_l1b_trn)
print(filename_l1b_tst)
print(filename_l2_trn)
print(filename_l2_tst)
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, filename_l1b, filename_l2, seed=None, shuffle=False):
if filename_l1b is not None:
self.h5f_l1b_tst = h5py.File(filename_l1b, 'r')
if filename_l2 is not None:
self.h5f_l2_tst = h5py.File(filename_l2, 'r')
if self.h5f_l1b_tst is not None:
h5f = self.h5f_l1b_tst
else:
h5f = self.h5f_l2_tst
time = h5f['time']
tst_idxs = np.arange(time.shape[0])
self.num_data_samples = len(tst_idxs)
if seed is not None:
np.random.seed(seed)
if shuffle:
np.random.shuffle(tst_idxs)
self.get_test_dataset(tst_idxs)
print('num test samples: ', tst_idxs.shape[0])
print('setup_test_pipeline: Done')
def setup_eval_pipeline(self, data_dct, num_tiles=1):
self.data_dct = data_dct
idxs = np.arange(num_tiles)
self.num_data_samples = idxs.shape[0]
self.get_evaluate_dataset(idxs)
def setup_salt_pipeline(self, data_path, label_path, perc=0.15):
data_files = np.array(glob.glob(data_path+'*.png'))
label_files = np.array(glob.glob(label_path+'*.png'))
num_data_samples = len(data_files)
idxs = np.arange(num_data_samples)
np.random.shuffle(idxs)
num_train = int(num_data_samples*(1-perc))
self.num_data_samples = num_train
trn_idxs = idxs[0:num_train]
np.sort(trn_idxs)
tst_idxs = idxs[num_train:]
np.sort(tst_idxs)
self.train_data_files = data_files[trn_idxs]
self.train_label_files = label_files[trn_idxs]
self.test_data_files = data_files[tst_idxs]
self.test_label_files = label_files[tst_idxs]
trn_idxs = np.arange(len(trn_idxs))
tst_idxs = np.arange(len(tst_idxs))
np.random.shuffle(trn_idxs)
self.get_train_dataset(trn_idxs)
self.get_test_dataset(tst_idxs)
f = open(home_dir+'/salt_test_files.pkl', 'wb')
pickle.dump((self.test_data_files, self.test_label_files), f)
f.close()
def setup_salt_restore(self, test_files='/Users/tomrink/salt_test_files.pkl'):
tup = pickle.load(open(test_files, 'rb'))
self.test_data_files = tup[0]
self.test_label_files = tup[1]
self.num_data_samples = len(self.test_data_files)
tst_idxs = np.arange(self.num_data_samples)
self.get_test_dataset(tst_idxs)
def build_unet(self):
print('build_cnn')
# padding = "VALID"
padding = "SAME"
# activation = tf.nn.relu
# activation = tf.nn.elu
activation = tf.nn.leaky_relu
momentum = 0.99
# num_filters = len(self.train_params) * 4
num_filters = 3 * 4
input_2d = self.inputs[0]
conv = tf.keras.layers.Conv2D(num_filters, kernel_size=5, strides=1, padding=padding, activation=None)(input_2d)
print('Contracting Branch')
print('input: ', conv.shape)
skip = conv
if NOISE_TRAINING:
conv = tf.keras.layers.GaussianNoise(stddev=NOISE_STDDEV)(conv)
# Contracting (Encoding) ------------------------------------------------------------------------------------
conv_1 = conv
conv = tf.keras.layers.Conv2D(num_filters, kernel_size=5, strides=1, padding=padding, activation=activation)(conv)
conv = tf.keras.layers.MaxPool2D(padding=padding)(conv)
conv = tf.keras.layers.BatchNormalization()(conv)
skip = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, padding=padding, activation=None)(skip)
skip = tf.keras.layers.MaxPool2D(padding=padding)(skip)
skip = tf.keras.layers.BatchNormalization()(skip)
conv = conv + skip
conv = tf.keras.layers.LeakyReLU()(conv)
print('1d: ', conv.shape)
# -----------------------------------------------------------------------------------------------------------
conv_2 = conv
skip = conv
num_filters *= 2
conv = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, padding=padding, activation=activation)(conv)
conv = tf.keras.layers.MaxPool2D(padding=padding)(conv)
conv = tf.keras.layers.BatchNormalization()(conv)
skip = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, padding=padding, activation=None)(skip)
skip = tf.keras.layers.MaxPool2D(padding=padding)(skip)
skip = tf.keras.layers.BatchNormalization()(skip)
conv = conv + skip
conv = tf.keras.layers.LeakyReLU()(conv)
print('2d: ', conv.shape)
# ----------------------------------------------------------------------------------------------------------
conv_3 = conv
skip = conv
num_filters *= 2
conv = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, padding=padding, activation=activation)(conv)
conv = tf.keras.layers.MaxPool2D(padding=padding)(conv)
conv = tf.keras.layers.BatchNormalization()(conv)
skip = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, padding=padding, activation=None)(skip)
skip = tf.keras.layers.MaxPool2D(padding=padding)(skip)
skip = tf.keras.layers.BatchNormalization()(skip)
conv = conv + skip
conv = tf.keras.layers.LeakyReLU()(conv)
print('3d: ', conv.shape)
# -----------------------------------------------------------------------------------------------------------
conv_4 = conv
skip = conv
num_filters *= 2
conv = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, padding=padding, activation=activation)(conv)
conv = tf.keras.layers.MaxPool2D(padding=padding)(conv)
conv = tf.keras.layers.BatchNormalization()(conv)
skip = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, padding=padding, activation=None)(skip)
skip = tf.keras.layers.MaxPool2D(padding=padding)(skip)
skip = tf.keras.layers.BatchNormalization()(skip)
conv = conv + skip
conv = tf.keras.layers.LeakyReLU()(conv)
print('4d: ', conv.shape)
# Expanding (Decoding) branch -------------------------------------------------------------------------------
print('expanding branch')
num_filters /= 2
conv = tf.keras.layers.Conv2DTranspose(num_filters, kernel_size=3, strides=2, padding=padding)(conv)
conv = tf.keras.layers.concatenate([conv, conv_4])
conv = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, padding=padding, activation=activation)(conv)
conv = tf.keras.layers.BatchNormalization()(conv)
print('5: ', conv.shape)
num_filters /= 2
conv = tf.keras.layers.Conv2DTranspose(num_filters, kernel_size=3, strides=2, padding=padding)(conv)
conv = tf.keras.layers.concatenate([conv, conv_3])
conv = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, padding=padding, activation=activation)(conv)
conv = tf.keras.layers.BatchNormalization()(conv)
print('6: ', conv.shape)
num_filters /= 2
conv = tf.keras.layers.Conv2DTranspose(num_filters, kernel_size=3, strides=2, padding=padding)(conv)
conv = tf.keras.layers.concatenate([conv, conv_2])
conv = tf.keras.layers.Conv2D(num_filters, kernel_size=3, strides=1, padding=padding, activation=activation)(conv)
conv = tf.keras.layers.BatchNormalization()(conv)
print('7: ', conv.shape)
num_filters /= 2
conv = tf.keras.layers.Conv2DTranspose(num_filters, kernel_size=3, strides=2, padding=padding)(conv)
print('8: ', conv.shape)
if NumClasses == 2:
activation = tf.nn.sigmoid # For binary
else:
activation = tf.nn.softmax # For multi-class
# Called logits, but these are actually probabilities, see activation
self.logits = tf.keras.layers.Conv2D(1, kernel_size=1, strides=1, padding=padding, name='probability', activation=activation)(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
# decayed_learning_rate = learning_rate * decay_rate ^ (global_step / decay_steps)
initial_learning_rate = 0.006
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_loss = tf.keras.metrics.Mean(name='train_loss')
self.test_loss = tf.keras.metrics.Mean(name='test_loss')
if NumClasses == 2:
self.train_accuracy = tf.keras.metrics.BinaryAccuracy(name='train_accuracy')
self.test_accuracy = tf.keras.metrics.BinaryAccuracy(name='test_accuracy')
self.test_auc = tf.keras.metrics.AUC(name='test_auc')
self.test_recall = tf.keras.metrics.Recall(name='test_recall')
self.test_precision = tf.keras.metrics.Precision(name='test_precision')
self.test_true_neg = tf.keras.metrics.TrueNegatives(name='test_true_neg')
self.test_true_pos = tf.keras.metrics.TruePositives(name='test_true_pos')
self.test_false_neg = tf.keras.metrics.FalseNegatives(name='test_false_neg')
self.test_false_pos = tf.keras.metrics.FalsePositives(name='test_false_pos')
else:
self.train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')
self.test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='test_accuracy')
@tf.function
def train_step(self, mini_batch):
inputs = [mini_batch[0], mini_batch[1]]
labels = mini_batch[2]
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], mini_batch[1]]
labels = mini_batch[2]
pred = self.model(inputs, training=False)
t_loss = self.loss(labels, pred)
self.test_loss(t_loss)
self.test_accuracy(labels, pred)
if NumClasses == 2:
self.test_auc(labels, pred)
self.test_recall(labels, pred)
self.test_precision(labels, pred)
self.test_true_neg(labels, pred)
self.test_true_pos(labels, pred)
self.test_false_neg(labels, pred)
self.test_false_pos(labels, pred)
def predict(self, mini_batch):
inputs = [mini_batch[0], mini_batch[1]]
labels = mini_batch[2]
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)
if NumClasses == 2:
self.test_auc(labels, pred)
self.test_recall(labels, pred)
self.test_precision(labels, pred)
self.test_true_neg(labels, pred)
self.test_true_pos(labels, pred)
self.test_false_neg(labels, pred)
self.test_false_pos(labels, pred)
def reset_test_metrics(self):
self.test_loss.reset_states()
self.test_accuracy.reset_states()
if NumClasses == 2:
self.test_auc.reset_states()
self.test_recall.reset_states()
self.test_precision.reset_states()
self.test_true_neg.reset_states()
self.test_true_pos.reset_states()
self.test_false_neg.reset_states()
self.test_false_pos.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
best_test_acc = 0
best_test_recall = 0
best_test_precision = 0
best_test_auc = 0
best_test_f1 = 0
best_test_mcc = 0
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 data0, data1, label in self.train_dataset:
trn_ds = tf.data.Dataset.from_tensor_slices((data0, data1, 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 data0_tst, data1_tst, label_tst in self.test_dataset:
tst_ds = tf.data.Dataset.from_tensor_slices((data0_tst, data1_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)
if NumClasses == 2:
tf.summary.scalar('auc_val', self.test_auc.result(), step=step)
tf.summary.scalar('recall_val', self.test_recall.result(), step=step)
tf.summary.scalar('prec_val', self.test_precision.result(), step=step)
tf.summary.scalar('f1_val', f1, step=step)
tf.summary.scalar('mcc_val', mcc, step=step)
tf.summary.scalar('num_train_steps', step, step=step)
tf.summary.scalar('num_epochs', epoch, 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 += data0.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 data0, data1, label in self.test_dataset:
ds = tf.data.Dataset.from_tensor_slices((data0, data1, label))
ds = ds.batch(BATCH_SIZE)
for mini_batch in ds:
self.test_step(mini_batch)
if NumClasses == 2:
f1, mcc = self.get_metrics()
print('loss, acc, recall, precision, auc, f1, mcc: ', self.test_loss.result().numpy(), self.test_accuracy.result().numpy(),
self.test_recall.result().numpy(), self.test_precision.result().numpy(), self.test_auc.result().numpy(), f1.numpy(), mcc.numpy())
else:
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
if NumClasses == 2:
best_test_acc = self.test_accuracy.result().numpy()
best_test_recall = self.test_recall.result().numpy()
best_test_precision = self.test_precision.result().numpy()
best_test_auc = self.test_auc.result().numpy()
best_test_f1 = f1.numpy()
best_test_mcc = mcc.numpy()
ckpt_manager.save()
if self.DISK_CACHE and epoch == 0:
f = open(cachepath, 'wb')
pickle.dump(self.in_mem_data_cache, f)
f.close()
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()
if self.h5f_l1b_trn is not None:
self.h5f_l1b_trn.close()
if self.h5f_l1b_tst is not None:
self.h5f_l1b_tst.close()
if self.h5f_l2_trn is not None:
self.h5f_l2_trn.close()
if self.h5f_l2_tst is not None:
self.h5f_l2_tst.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_unet()
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 data0, data1, label in self.test_dataset:
ds = tf.data.Dataset.from_tensor_slices((data0, data1, label))
ds = ds.batch(BATCH_SIZE)
for mini_batch_test in ds:
self.predict(mini_batch_test)
f1, mcc = self.get_metrics()
print('loss, acc: ', self.test_loss.result().numpy(), self.test_accuracy.result().numpy(), self.test_recall.result().numpy(),
self.test_precision.result().numpy(), self.test_auc.result().numpy(), f1.numpy(), mcc.numpy())
labels = np.concatenate(self.test_labels)
self.test_labels = labels
preds = np.concatenate(self.test_preds)
self.test_probs = preds
if NumClasses == 2:
preds = np.where(preds > 0.5, 1, 0)
else:
preds = np.argmax(preds, axis=1)
self.test_preds = preds
def do_evaluate(self, prob_thresh=0.5):
self.reset_test_metrics()
pred_s = []
for data in self.eval_dataset:
print(data[0].shape, data[1].shape)
pred = self.model([data])
print(pred.shape, np.histogram(pred.numpy()))
preds = np.concatenate(pred_s)
preds = preds[:,0]
self.test_probs = preds
if NumClasses == 2:
preds = np.where(preds > prob_thresh, 1, 0)
else:
preds = np.argmax(preds, axis=1)
self.test_preds = preds
def run(self, filename_l1b_trn, filename_l1b_tst, filename_l2_trn, filename_l2_tst):
self.setup_pipeline(filename_l1b_trn, filename_l1b_tst, filename_l2_trn, filename_l2_tst)
self.build_model()
self.build_training()
self.build_evaluation()
self.do_training()
def run_test(self, data_path='/Users/tomrink/data/salt/train/images/', label_path='/Users/tomrink/data/salt/train/masks/'):
self.setup_salt_pipeline(data_path, label_path)
self.build_model()
self.build_training()
self.build_evaluation()
self.do_training()
def run_restore_test(self, data_path='/Users/tomrink/data/salt/train/images/', label_path='/Users/tomrink/data/salt/train/masks/', ckpt_dir=None):
#self.setup_salt_pipeline(data_path, label_path)
self.setup_salt_restore()
self.build_model()
self.build_training()
self.build_evaluation()
self.restore(ckpt_dir)
def run_restore(self, filename_l1b, filename_l2, ckpt_dir):
self.setup_test_pipeline(filename_l1b, filename_l2)
self.build_model()
self.build_training()
self.build_evaluation()
self.restore(ckpt_dir)
if self.h5f_l1b_tst is not None:
self.h5f_l1b_tst.close()
if self.h5f_l2_tst is not None:
self.h5f_l2_tst.close()
def run_evaluate(self, filename, ckpt_dir):
data_dct, ll, cc = make_for_full_domain_predict(filename, name_list=self.train_params)
self.setup_eval_pipeline(data_dct, len(ll))
self.build_model()
self.build_training()
self.build_evaluation()
self.do_evaluate(ckpt_dir)
if __name__ == "__main__":
nn = UNET()
nn.run('matchup_filename')