def test(DATASET="Texas", CONFIG=None):
"""
1. Fetch data (x, y, change_map)
2. Compute/estimate A_x and A_y (for patches)
3. Compute change_prior
4. Define dataset with (x, A_x, y, A_y, p). Choose patch size compatible
with affinity computations.
5. Train CrossCyclicImageTransformer unsupervised
a. Evaluate the image transformations in some way?
6. Evaluate the change detection scheme
a. change_map = threshold [(x - f_y(y))/2 + (y - f_x(x))/2]
"""
if CONFIG is None:
CONFIG = get_config_kACE(DATASET)
print(f"Loading {DATASET} data")
x_im, y_im, EVALUATE, (C_X, C_Y) = datasets.fetch(DATASET, **CONFIG)
if tf.test.is_gpu_available() and not CONFIG["debug"]:
C_CODE = 3
print("here")
TRANSLATION_SPEC = {
"enc_X": {
"input_chs": C_X,
"filter_spec": [50, 50, C_CODE]
},
"enc_Y": {
"input_chs": C_Y,
"filter_spec": [50, 50, C_CODE]
},
"dec_X": {
"input_chs": C_CODE,
"filter_spec": [50, 50, C_X]
},
"dec_Y": {
"input_chs": C_CODE,
"filter_spec": [50, 50, C_Y]
},
}
else:
print("why here?")
C_CODE = 1
TRANSLATION_SPEC = {
"enc_X": {
"input_chs": C_X,
"filter_spec": [C_CODE]
},
"enc_Y": {
"input_chs": C_Y,
"filter_spec": [C_CODE]
},
"dec_X": {
"input_chs": C_CODE,
"filter_spec": [C_X]
},
"dec_Y": {
"input_chs": C_CODE,
"filter_spec": [C_Y]
},
}
print("Change Detector Init")
cd = Kern_AceNet(TRANSLATION_SPEC, **CONFIG)
print("Training")
training_time = 0
cross_loss_weight = tf.expand_dims(
tf.zeros(x_im.shape[:-1], dtype=tf.float32), -1)
for epochs in CONFIG["list_epochs"]:
CONFIG.update(epochs=epochs)
tr_gen, dtypes, shapes = datasets._training_data_generator(
x_im[0], y_im[0], cross_loss_weight[0], CONFIG["patch_size"])
TRAIN = tf.data.Dataset.from_generator(tr_gen, dtypes, shapes)
TRAIN = TRAIN.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
tr_time, _ = cd.train(TRAIN, evaluation_dataset=EVALUATE, **CONFIG)
for x, y, _ in EVALUATE.batch(1):
alpha = cd([x, y])
cross_loss_weight = 1.0 - alpha
training_time += tr_time
cd.save_all_weights()
cd.final_evaluate(EVALUATE, **CONFIG)
# Alternate evaluation
cd.alt_evaluate(EVALUATE, **CONFIG)
final_kappa = cd.metrics_history["cohens kappa"][-1]
final_acc = cd.metrics_history["ACC"][-1]
performance = (final_kappa, final_acc)
timestamp = cd.timestamp
epoch = cd.epoch.numpy()
speed = (epoch, training_time, timestamp)
# Save config parameters
with open(os.path.join(CONFIG["logdir"], "config.txt"), "w+") as f:
print(CONFIG, file=f)
# Save full translated images
train_flag_tmp = False
x_code, y_code = cd._enc_x(x, train_flag_tmp), cd._enc_y(y, train_flag_tmp)
x_hat, y_hat = cd._dec_x(y_code, train_flag_tmp), cd._dec_y(
x_code, train_flag_tmp)
im_dir = CONFIG["logdir"]
#print("SHAPE OF X_HAT: ", x_hat.shape)
#print("SHAPE OF Y_HAT: ", y_hat.shape)
x_hat_out = np.squeeze(np.array(x_hat))
y_hat_out = np.squeeze(np.array(y_hat))
#print("SHAPE OF X_HAT_OUT: ", x_hat_out.shape)
#print("SHAPE OF Y_HAT_OUT: ", y_hat_out.shape)
tifffile.imsave(os.path.join(im_dir, "x_hat_full.tif"),
x_hat_out,
planarconfig="contig")
tifffile.imsave(os.path.join(im_dir, "y_hat_full.tif"),
y_hat_out,
planarconfig="contig")
del x_hat, x_code, y_hat, y_code
del cd
gc.collect()
return performance, speed
def test(DATASET="Texas", CONFIG=None):
"""
1. Fetch data (x, y, change_map)
2. Compute/estimate A_x and A_y (for patches)
3. Compute change_prior
4. Define dataset with (x, A_x, y, A_y, p). Choose patch size compatible
with affinity computations.
5. Train CrossCyclicImageTransformer unsupervised
a. Evaluate the image transformations in some way?
6. Evaluate the change detection scheme
a. change_map = threshold [(x - f_y(y))/2 + (y - f_x(x))/2]
"""
if CONFIG is None:
CONFIG = get_config_kACE(DATASET)
print(f"Loading {DATASET} data")
x_im, y_im, EVALUATE, (C_X, C_Y) = datasets.fetch(DATASET, **CONFIG)
if tf.test.is_gpu_available() and not CONFIG["debug"]:
C_CODE = 3
TRANSLATION_SPEC = {
"enc_X": {
"input_chs": C_X,
"filter_spec": [50, 50, C_CODE]
},
"enc_Y": {
"input_chs": C_Y,
"filter_spec": [50, 50, C_CODE]
},
"dec_X": {
"input_chs": C_CODE,
"filter_spec": [50, 50, C_X]
},
"dec_Y": {
"input_chs": C_CODE,
"filter_spec": [50, 50, C_Y]
},
}
else:
C_CODE = 1
TRANSLATION_SPEC = {
"enc_X": {
"input_chs": C_X,
"filter_spec": [C_CODE]
},
"enc_Y": {
"input_chs": C_Y,
"filter_spec": [C_CODE]
},
"dec_X": {
"input_chs": C_CODE,
"filter_spec": [C_X]
},
"dec_Y": {
"input_chs": C_CODE,
"filter_spec": [C_Y]
},
}
print("Change Detector Init")
cd = Kern_AceNet(TRANSLATION_SPEC, **CONFIG)
print("Training")
training_time = 0
cross_loss_weight = tf.expand_dims(
tf.zeros(x_im.shape[:-1], dtype=tf.float32), -1)
for epochs in CONFIG["list_epochs"]:
CONFIG.update(epochs=epochs)
tr_gen, dtypes, shapes = datasets._training_data_generator(
x_im[0], y_im[0], cross_loss_weight[0], CONFIG["patch_size"])
TRAIN = tf.data.Dataset.from_generator(tr_gen, dtypes, shapes)
TRAIN = TRAIN.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
tr_time, _ = cd.train(TRAIN, evaluation_dataset=EVALUATE, **CONFIG)
for x, y, _ in EVALUATE.batch(1):
alpha = cd([x, y])
cross_loss_weight = 1.0 - alpha
training_time += tr_time
cd.final_evaluate(EVALUATE, **CONFIG)
final_kappa = cd.metrics_history["cohens kappa"][-1]
final_acc = cd.metrics_history["ACC"][-1]
performance = (final_kappa, final_acc)
timestamp = cd.timestamp
epoch = cd.epoch.numpy()
speed = (epoch, training_time, timestamp)
del cd
gc.collect()
return performance, speed
#process_list = polmak_list
#process_list = ["Polmak-Air05-Air10-align_sub0", "Polmak-Air10-Air15-align_sub0"] #"Polmak-Air05-Air15-align_sub0"] #,
#process_list = ["Polmak-PAL-A2-align", "Polmak-LS5-S2", "Polmak-LS2004-RLee_C-align", "Texas", "California", "Polmak-LS5-PGNLM_A", "Polmak-LS5-PGNLM_C"]
# ["Polmak-Air05-Air10-align_sub0", "Polmak-Air05-Air15-align_sub0"]
# process_list = ["Polmak-LS5-S2", "Polmak-A2-S2", "Polmak-A2-S2-collocate", "Polmak-LS5-PGNLM_A", "Polmak-LS5-PGNLM_C"]
for DATASET in process_list:
if DATASET in [
"Polmak-Pal-RS2_010817-collocate", "Polmak-LS5-S2-NDVI",
"Polmak-Air05-S2-align_sub0", "Polmak-Air15-S2-align_sub0"
]: # "Polmak-Pal-RS2_010817-collocate",
print("Skipping dataset: " + DATASET)
continue
else:
print(DATASET)
CONFIG = get_config_kACE(DATASET)
# Basis for difference image
CONFIG["difference_basis"] = "original" # "translated" #
# Bandwidth for domain difference images
CONFIG["domain_diff_bw_x"] = tf.constant(
3.0, dtype=tf.float32) # tf.constant(3, dtype=tf.float32)
CONFIG["domain_diff_bw_y"] = tf.constant(
3.0, dtype=tf.float32) # tf.constant(3, dtype=tf.float32)
#CONFIG["krnl_width_x"] = 1.0
#CONFIG["krnl_width_y"] = 1.0
# Suffix to add to log output name
suffix = "" #"_NOTILDE" # "_sigma25pct" # "_domdiffBW3_kwx0p50_kwy0p50" # "_NOTILDE_kwx1p0_kwy1p0" #
suffix += "_diff-" + CONFIG["difference_basis"]
print(suffix)
def test(DATASET="Texas", CONFIG=None):
"""
1. Fetch data (x, y, change_map)
2. Compute/estimate A_x and A_y (for patches)
3. Compute change_prior
4. Define dataset with (x, A_x, y, A_y, p). Choose patch size compatible
with affinity computations.
5. Train CrossCyclicImageTransformer unsupervised
a. Evaluate the image transformations in some way?
6. Evaluate the change detection scheme
a. change_map = threshold [(x - f_y(y))/2 + (y - f_x(x))/2]
"""
if CONFIG is None:
CONFIG = get_config_kACE(DATASET)
print(f"Loading {DATASET} data")
x_im, y_im, EVALUATE, (C_X, C_Y) = datasets.fetch(DATASET, **CONFIG)
if tf.config.list_physical_devices("GPU") and not CONFIG["debug"]:
C_CODE = 3
print("here")
TRANSLATION_SPEC = {
"enc_X": {"input_chs": C_X, "filter_spec": [50, 50, C_CODE]},
"enc_Y": {"input_chs": C_Y, "filter_spec": [50, 50, C_CODE]},
"dec_X": {"input_chs": C_CODE, "filter_spec": [50, 50, C_X]},
"dec_Y": {"input_chs": C_CODE, "filter_spec": [50, 50, C_Y]},
}
else:
print("why here?")
C_CODE = 1
TRANSLATION_SPEC = {
"enc_X": {"input_chs": C_X, "filter_spec": [C_CODE]},
"enc_Y": {"input_chs": C_Y, "filter_spec": [C_CODE]},
"dec_X": {"input_chs": C_CODE, "filter_spec": [C_X]},
"dec_Y": {"input_chs": C_CODE, "filter_spec": [C_Y]},
}
print("Change Detector Init")
cd = Kern_AceNet(TRANSLATION_SPEC, **CONFIG)
print("Training")
training_time = 0
cross_loss_weight = tf.expand_dims(tf.zeros(x_im.shape[:-1], dtype=tf.float32), -1)
for epochs in CONFIG["list_epochs"]:
CONFIG.update(epochs=epochs)
tr_gen, dtypes, shapes = datasets._training_data_generator(
x_im[0], y_im[0], cross_loss_weight[0], CONFIG["patch_size"]
)
TRAIN = tf.data.Dataset.from_generator(tr_gen, dtypes, shapes)
TRAIN = TRAIN.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
tr_time, _ = cd.train(TRAIN, evaluation_dataset=EVALUATE, **CONFIG)
for x, y, _ in EVALUATE.batch(1):
alpha = cd([x, y])
cross_loss_weight = 1.0 - alpha
training_time += tr_time
cd.load_all_weights(cd.log_path)
cd.final_evaluate(EVALUATE, **CONFIG)
metrics = {}
for key in list(cd.difference_img_metrics.keys()) + list(
cd.change_map_metrics.keys()
):
metrics[key] = cd.metrics_history[key][-1]
metrics["F1"] = metrics["TP"] / (
metrics["TP"] + 0.5 * (metrics["FP"] + metrics["FN"])
)
timestamp = cd.timestamp
epoch = cd.epoch.numpy()
speed = (epoch, training_time, timestamp)
del cd
gc.collect()
return metrics, speed
