def main():
"""
:return:
"""
### Settings
mod = 5
w_patch = 16 * 2
"""
Data (all important modalities)
"""
# folder_windows = r'C:\Users\Laurens_laptop_w\OneDrive - UGent\data\10lamb'
train_data = get_10lamb_old(mod)
img_x, img_y_tr, _, _ = get_training_data(train_data)
# Normalise the input!
img_x = rescale0to1(img_x)
"""
Train segmentation
1) reuse everything
2) fix encoder
"""
if 1:
if 1:
b_encoder_fixed = False
info_enc_fixed = '_enc_fixed' if b_encoder_fixed else ''
get_info = lambda: f'10lamb_kfold_pretrained{info_enc_fixed}/unet_enc_k{k}_ifold{i_fold}'
n_epochs = 40
k = 10
if k == 10:
epoch_w = 100
else:
raise NotImplementedError()
### Settings you don't have to change:
w_patch = 50
w_ext_in = 28
b_double = False
padding = 'valid'
# TODO flag for converting encoder to dilated conv
def get_unet_pretrained_encoder():
model_encoder = get_model_encoder()
encoder_inputs = model_encoder.input
decoder_outputs = decoder(model_encoder, f_out=2)
model_pretrained_unet = Model(encoder_inputs, decoder_outputs)
from methods.examples import compile_segm
compile_segm(model_pretrained_unet, lr=1e-4)
model_pretrained_unet.summary()
return model_pretrained_unet
"""
Train
"""
k_fold_train_data = get_10lamb_6patches(5)
for i_fold in range(6):
"""
Get a new network (not trained yet for segmentation)
"""
model_pretrained_unet = get_unet_pretrained_encoder()
n_pretrained_unet = NeuralNet(model_pretrained_unet)
"""
The data
"""
train_data_i = k_fold_train_data.k_split_i(i_fold)
info = get_info()
img_y_tr = train_data_i.get_y_train()
img_y_te = train_data_i.get_y_test()
flow_tr = get_flow(
img_x,
img_y_tr,
w_patch=w_patch, # Comes from 10
w_ext_in=w_ext_in)
flow_te = get_flow(
img_x,
img_y_te,
w_patch=w_patch, # Comes from 10
w_ext_in=w_ext_in)
n_pretrained_unet.train(flow_tr,
flow_te,
epochs=n_epochs,
verbose=1,
info=info)
"""
Prediction
"""
n_pretrained_unet.w_ext = w_ext_in
y_pred = n_pretrained_unet.predict(img_x)
concurrent([y_pred[..., 1]])
"""
Classification
"""
if 1:
im_clean = img_x[..., :3]
k = 8
i_fold = 3
epoch_last = 40
from methods.examples import kappa_loss, weighted_categorical_crossentropy
from performance.metrics import accuracy_with0, jaccard_with0
loss = weighted_categorical_crossentropy((1, 1))
list_y_pred = []
### K fold validation
k_fold_train_data = get_10lamb_6patches(5)
train_data_i = k_fold_train_data.k_split_i(i_fold)
img_y_tr = train_data_i.get_y_train()
img_y_te = train_data_i.get_y_test()
for epoch in np.arange(31, epoch_last + 1):
filepath_model = f'/scratch/lameeus/data/ghent_altar/net_weight/10lamb_kfold/ti_unet_k{k}_kfold{i_fold}/w_{epoch}.h5'
model = load_model(filepath_model,
custom_objects={
'loss': loss,
'accuracy_with0': accuracy_with0,
'jaccard_with0': jaccard_with0,
'kappa_loss': kappa_loss
})
n = NeuralNet(model, w_ext=10)
y_pred = n.predict(img_x)
list_y_pred.append(y_pred)
y_pred_mean = np.mean(list_y_pred, axis=0)
q1 = y_pred_mean[..., 1]
concurrent([q1, q1.round(), im_clean])
"""
Optimal threshold (making conf matrix symmetric, not based on maximising kappa)
"""
y_gt = np.any([img_y_tr, img_y_te], axis=0)
from performance.testing import _get_scores, filter_non_zero
def foo_performance(y_true, y_pred, thresh):
# is basically argmax
y_pred_thresh_arg = np.greater_equal(y_pred[..., 1], thresh)
y_true_flat, y_pred_thresh_arg_flat = filter_non_zero(
y_true, y_pred_thresh_arg)
y_te_argmax = np.argmax(y_true_flat, axis=-1)
# Kappa
return _get_scores(y_te_argmax, y_pred_thresh_arg_flat)[-1]
"""
1. BEST? PERFORMANCE based on test set
"""
print('1. Test distribution optimization')
thresh = optimal_test_thresh_equal_distribution(img_y_te, y_pred_mean)
q1_thresh = np.greater_equal(q1, thresh)
concurrent([q1, q1_thresh, im_clean])
print(f'thresh: {thresh}')
# Test, train, both
print('Kappa performance:')
print('\ttrain:', foo_performance(img_y_tr, y_pred_mean, thresh))
print('\ttestset:', foo_performance(img_y_te, y_pred_mean, thresh))
print('\tboth:', foo_performance(y_gt, y_pred_mean, thresh))
print('\nIncremental optimization on test set')
test_thresh2 = test_thresh_incremental(y_pred_mean,
img_y_tr,
img_y_te,
n=5,
verbose=0)
print('Kappa performance:')
print('\ttrain:', foo_performance(img_y_tr, y_pred_mean, test_thresh2))
print('\ttestset:', foo_performance(img_y_te, y_pred_mean,
test_thresh2))
print('\tboth:', foo_performance(y_gt, y_pred_mean, test_thresh2))
"""
2. based on train
"""
print('\n2. Training distribution optimization')
thresh = optimal_test_thresh_equal_distribution(img_y_tr, y_pred_mean)
q1_thresh = np.greater_equal(q1, thresh)
concurrent([q1, q1_thresh, im_clean])
print(f'thresh: {thresh}')
# Test, train, both
print('Kappa performance:')
print('\ttrain:', foo_performance(img_y_tr, y_pred_mean, thresh))
print('\ttestset:', foo_performance(img_y_te, y_pred_mean, thresh))
print('\tboth:', foo_performance(y_gt, y_pred_mean, thresh))
"""
3. CONSISTENT: based on train+set
"""
print('\n3. all GT distribution optimization')
thresh = optimal_test_thresh_equal_distribution(y_gt, y_pred_mean)
q1_thresh = np.greater_equal(q1, thresh)
concurrent([q1, q1_thresh, im_clean])
print(f'thresh: {thresh}')
# Test, train, both
print('Kappa performance:')
print('\ttrain:', foo_performance(img_y_tr, y_pred_mean, thresh))
print('\ttestset:', foo_performance(img_y_te, y_pred_mean, thresh))
print('\tboth:', foo_performance(y_gt, y_pred_mean, thresh))
if 0:
"""
4. DUMB/Not needed: Based on prediction of whole panel
"""
thresh = optimal_test_thresh_equal_distribution(y_gt,
y_pred_mean,
mask_true=False)
q1_thresh = np.greater_equal(q1, thresh)
concurrent([q1, q1_thresh, im_clean])
print('Done')
def train_segm(self):
from figures_paper.overlay import semi_transparant
if self.fixed_enc == -2:
def get_n_model_regular(i_fold=None, k=None, epoch=None):
n_in = 9
model_tiunet = ti_unet(n_in,
filters=k,
w=self.w_patch,
ext_in=10 // 2,
batch_norm=True,
wrong_batch_norm=True)
compile_segm(model_tiunet, 1e-4)
""" TODO wrong tiunet """
n = NeuralNet(model_tiunet, w_ext=10)
info = f'10lamb_kfold/ti_unet_k{k}_kfold{i_fold}'
folder = os.path.join(
'/scratch/lameeus/data/ghent_altar/net_weight', info)
n.load(folder=folder, epoch=epoch)
return n
n_segm = get_n_model_regular(i_fold=self.i_fold,
k=self.k,
epoch=40)
elif self.fixed_enc == -1:
# No init
model_segm = self.get_tiunet_preenc(k=self.k, lr=self.lr_opt)
n_segm = NeuralNet(model_segm, w_ext=self.w_ext_in_ti)
elif self.fixed_enc in [0, 1, 2]:
# Load model
model_segm = self.get_tiunet_preenc(k=self.k, lr=self.lr_opt)
n_segm = NeuralNet(model_segm, w_ext=self.w_ext_in_ti)
# Train on set
folder_weights = '/scratch/lameeus/data/ghent_altar/net_weight'
if self.fixed_enc == 0:
folder1 = '10lamb_kfold_pretrained_batchnorm'
elif self.fixed_enc == 1:
folder1 = '10lamb_kfold_pretrained_encfixed_batchnorm'
elif self.fixed_enc == 2:
folder1 = '10lamb_kfold_pretrained_prefixed_batchnorm'
folder2 = f'{"tiunet"}_d{self.depth}_k{self.k}_ifold{self.i_fold}'
n_segm.load(os.path.join(folder_weights, folder1, folder2), 100)
elif self.fixed_enc == 3:
model_segm = self.get_tiunet_preenc(k=self.k,
lr=self.lr_opt,
set_info=f'_{self.set_nr}',
epoch_start=100)
n_segm = NeuralNet(model_segm, w_ext=self.w_ext_in_ti)
else:
NotImplementedError()
def foo(n_segm, b=0):
y_pred = n_segm.predict(self.img_x)
thresh_single = optimal_test_thresh_equal_distribution(
self.img_y_te, y_pred)
# data_single_i = {'k': self.k,
# 'i_fold': i_fold,
# 'epoch': epoch}
print(foo_performance(self.img_y_te, y_pred, thresh_single))
img_clean = self.img_x[..., :3]
concurrent([
img_clean, y_pred[..., 1], y_pred[..., 1] >= thresh_single,
semi_transparant(img_clean, y_pred[..., 1] >= thresh_single)
])
if b:
from data.datatools import imsave
folder = '/home/lameeus/data/ghent_altar/output/hierarchy/'
info_epoch = f'_epoch{n_segm.epoch}' if n_segm.epoch > 0 else ''
filename = folder + f'13_small/pred_transfer_kfoldenc{self.fixed_enc}_ifold{self.i_fold}_avg{info_epoch}.png'
imsave(filename, y_pred[..., 1])
def set_encoder_state(model, trainable=False):
assert len(model.layers) == 14
for layer in model.layers[:7]:
layer.trainable = trainable
compile_segm(model)
n_segm.epoch = 0
# Without pretraining
foo(n_segm)
# epochs
set_encoder_state(n_segm.model, trainable=False)
for _ in range(10):
n_segm.train(self.flow_tr_10, epochs=1, verbose=2)
foo(n_segm, 0)
if 0:
set_encoder_state(n_segm.model, trainable=True)
for _ in range(1):
n_segm.train(self.flow_tr_set_10, epochs=10, verbose=2)
foo(n_segm, 0)
set_encoder_state(n_segm.model, trainable=True)
for _ in range(10):
n_segm.train(self.flow_tr_set, epochs=1, verbose=2)
foo(n_segm, 0)
class Main(object):
def __init__(self, k=None, seed=None):
if k is not None:
self.k = k
self.seed = seed
self.model()
self.train()
def model(self, b_optimal_lr=False):
features_out = 3 if data_name == '19botrightcrack3' else 2
if model_name == 'ti-unet':
model = ti_unet(
9,
filters=self.k,
w=w_patch,
ext_in=w_ext_in // 2,
batch_norm=True,
max_depth=d,
features_out=features_out,
)
elif model_name == 'unet':
# model = ti_unet(9, filters=self.k, w=w_patch, ext_in=w_ext_in // 2, batch_norm=True,
# max_depth=d)
print('NO BATCH NORM? (not implemented)')
model = unet(9,
filters=self.k,
w=w_patch,
ext_in=w_ext_in // 2,
max_depth=d,
n_per_block=1)
else:
raise ValueError(model_name)
model.summary()
compile_segm(model, lr=lr)
if b_optimal_lr:
from neuralNetwork.optimization import find_learning_rate
global flow_tr
find_learning_rate(model, flow_tr)
self.neural_net = NeuralNet(model, w_ext=w_ext_in)
if data_name[:5] == '1319_': # pre Load model!
# TODO which epoch to start from, I guess 10 should have an impact
epoch_start = 50 # probably better (learned something)
self.neural_net.load(
f'C:/Users/admin/Data/ghent_altar/net_weight/1319/{model_name}_d{d}_k{self.k}',
epoch=epoch_start)
def train(self, epochs=epochs, steps_per_epoch=100):
global flow_tr
info = f'{data_name}/{model_name}_d{d}_k{self.k}'
try:
if n_per_class is not None:
info = '_'.join([info, f'n{n_per_class}'])
except NameError:
pass # If n_per_class does not exist, don't add it
if self.seed is not None:
info += f'_s{self.seed}'
self.neural_net.train(
flow_tr,
# validation=flow_va,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
info=info)
# Solve Memory problems?
del self.neural_net
def eval(self):
pass
def train_segm(self):
folder_save = '/home/lameeus/data/ghent_altar/dataframes'
info_batchnorm = '_batchnorm' if self.batch_norm else ''
info_fixed = '_encfixed' if self.fixed_enc == 1 else '_prefixed' if self.fixed_enc == 2 else ''
info_model = 'tiunet' if self.ti else 'unet'
filename_single = f'pretrained/{info_model}_10lamb_kfold{info_fixed}{info_batchnorm}/d{self.depth}_single'
path_single = os.path.join(folder_save, filename_single + '.csv')
get_info = lambda: f'10lamb_kfold_pretrained{info_fixed}{info_batchnorm}/{info_model}_d{self.depth}_k{self.k}_ifold{i_fold}'
img_y_all = self.k_fold_train_data.get_train_data_all().get_y_train()
def get_model():
if self.ti:
model = self.get_tiunet_preenc(k=self.k, lr=self.lr_opt)
else:
model = self.get_unet_preenc(k=self.k, lr=self.lr_opt)
if self.fixed_enc == 2:
n_temp = NeuralNet(model)
folder_weights = '/scratch/lameeus/data/ghent_altar/net_weight'
folder1 = f'10lamb_kfold_pretrained{"_encfixed"}{info_batchnorm}'
folder2 = f'{info_model}_d{self.depth}_k{self.k}_ifold{i_fold}'
n_temp.load(os.path.join(folder_weights, folder1, folder2),
100)
del (n_temp)
return model
w_ext = self.w_ext_in_ti if self.ti else self.w_ext_in_ae
if not self.lr_opt:
model_segm = get_model()
find_learning_rate(model_segm, self.flow_segm, lr1=1e0)
for i_fold in range(6):
print(f'i_fold = {i_fold}')
model_segm = get_model()
n_segm = NeuralNet(model_segm, w_ext=w_ext)
train_data_i = self.k_fold_train_data.k_split_i(i_fold)
img_y_tr = train_data_i.get_y_train()
img_y_te = train_data_i.get_y_test()
flow_tr = get_flow(self.img_x,
img_y_tr,
w_patch=self.w_patch,
w_ext_in=w_ext)
flow_te = get_flow(self.img_x,
img_y_te,
w_patch=self.w_patch,
w_ext_in=w_ext)
info = get_info()
for epoch in range(self.epochs):
n_segm.train(flow_tr, flow_te, epochs=1, verbose=2, info=info)
y_pred = n_segm.predict(self.img_x)
thresh_single = optimal_test_thresh_equal_distribution(
img_y_all, y_pred)
data_single_i = {'k': self.k, 'i_fold': i_fold, 'epoch': epoch}
data_single_i.update(
foo_performance(img_y_te, y_pred, thresh_single))
lst_data_single = [data_single_i]
df_single = pd.DataFrame(lst_data_single)
pandas_save(path_single, df_single, append=True)
return
class Main(object):
k = 8
n_per_class = 80
d = 1
epochs = 100
if d == 1:
w_ext_in = 10
elif d == 2:
w_ext_in = 26
w_patch = 10
if 0:
lr = 1e-3
steps_per_epoch = 100
else:
lr = 1e-4
steps_per_epoch = 1000
def __init__(self, k=None, n_per_class=None):
if k is not None:
self.k = k
if n_per_class is not None:
self.n_per_class = n_per_class
# Get net
self.model_train = self.main_net(set_net)
from methods.basic import NeuralNet
self.neural_net = NeuralNet(self.model_train, w_ext=self.w_ext_in)
# Get data
train_data = self.main_data(set_data)
n_val_datas = len(self.val_datas)
lst_data = [[] for _ in range(n_val_datas + 1)]
for _ in range(self.epochs):
# Train
self.main_train(train_data)
# Evaluate
data_lst = self.main_eval(train_data)
for i, data_i in enumerate(data_lst):
data_i.update({'epoch': self.neural_net.epoch})
lst_data[i].append(data_i)
for i in range(n_val_datas + 1):
df = pd.DataFrame(lst_data[i])
print(df)
if i == 0:
data_name = 'val'
else:
data_name = self.val_datas[i - 1]['name']
model_name = f'{set_net["name"]}_data{data_name}_d{self.d}_k{self.k}_n{self.n_per_class}'
pandas_save(
f'C:/Users/admin/OneDrive - ugentbe/data/dataframes/{model_name}.csv',
df,
append=True)
if 0:
if 0:
self.neural_net.load(
'C:/Users/admin/Data/ghent_altar/net_weight/tiunet_d1_k10_n80',
4)
self.main_eval(train_data, b_plot=True)
print("Finished init")
def main_net(self, set_n):
from methods.examples import compile_segm
from neuralNetwork.architectures import ti_unet, convNet, unet
n_name = set_n['name'].lower()
if n_name == 'ti-unet':
model = ti_unet(9,
filters=self.k,
w=self.w_patch,
ext_in=self.w_ext_in // 2,
batch_norm=True,
max_depth=self.d)
elif n_name == 'simple':
model = convNet(9,
self.k,
w_in=self.w_patch + self.w_ext_in,
n_convs=5,
batch_norm=False,
padding='valid')
assert model.output_shape[-3:] == (self.w_patch, self.w_patch, 2)
elif n_name == 'unet':
print('NO BATCH NORM? (not implemented)')
model = unet(9,
filters=self.k,
w=self.w_patch,
ext_in=self.w_ext_in // 2,
max_depth=self.d,
n_per_block=1)
else:
raise ValueError(n_name)
# raise NotImplementedError('Unet is not well implemented: * Double, batchnorm? f per layer etc?')
model.summary()
compile_segm(
model, lr=self.lr) # instead of 10e-3, 10e-4 is probs more stable.
return model
def main_data(self, set_data):
if set_data['name'] == 'zach_sh':
from datasets.default_trainingsets import get_13botleftshuang
train_data = get_13botleftshuang(mod, n_per_class=self.n_per_class)
else:
raise NotImplementedError()
from data.preprocessing import rescale0to1
train_data.x = rescale0to1(train_data.x)
from datasets.default_trainingsets import xy_from_df, panel13withoutRightBot
from datasets.examples import get_13zach
_, img_y = xy_from_df(get_13zach(), mod)
img_y_top2, _ = panel13withoutRightBot(img_y)
img_y_test = np.logical_or(img_y_top2, train_data.get_y_test())
self.val_datas = [{
'name': '13_top2',
'y': img_y_top2
}, {
'name': '13_test',
'y': img_y_test
}]
return train_data
def main_train(self, train_data, steps_per_epoch=None):
if steps_per_epoch is None:
steps_per_epoch = self.steps_per_epoch
from main_general import get_training_data
from preprocessing.image import get_flow
# TODO train
x_train, y_train, x_val, y_val = get_training_data(train_data)
# Generator
flow_tr = get_flow(x_train,
y_train,
w_patch=self.w_patch,
w_ext_in=self.w_ext_in)
flow_va = get_flow(x_val,
y_val,
w_patch=self.w_patch,
w_ext_in=self.w_ext_in)
epochs = 1
self.neural_net.train(
flow_tr,
validation=flow_va,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
info=f'{set_net["name"]}_d{self.d}_k{self.k}_n{self.n_per_class}')
def main_eval(
self,
train_data,
b_plot=False,
):
x = train_data.get_x_test()
y_pred = self.neural_net.predict(x)
if b_plot:
concurrent([x[..., :3], y_pred[..., 1]], ['input', 'prediction'])
val_datas = [{'y': train_data.get_y_test()}] + self.val_datas
return _eval_func(y_pred, val_datas, b_plot=b_plot)