for n_blocks in range(3, 6):
max_context = 5 - n_blocks
for context_blocks in range(max_context + 1):
experiments.append([n_blocks, context_blocks])
if LOAD_MODEL==True:
X_train = np.load('X_train.npy',encoding='bytes')
Y_train = np.load('Y_train.npy',encoding='bytes')
X_oracle = np.load('X_oracle.npy',encoding='bytes')
Y_oracle = np.load('Y_oracle.npy',encoding='bytes')
X_val = np.load('X_val.npy',encoding='bytes')
Y_val = np.load('Y_val.npy',encoding='bytes')
X_working = np.load('X_working.npy',encoding='bytes')
Y_working = np.load('Y_working.npy',encoding='bytes')
else:
X_train,Y_train,X_oracle,Y_oracle,X_val,Y_val = train.get_data_iterators(batch_size=1,data_dir='mess_cb',target_size=(512, 512),rescale=1/255,fill_mode='constant',load_train_data=True,color_mode='rgb')
X_train = np.reshape(X_train,(200,512,512,3))
X_oracle = np.reshape(X_oracle,(568,512,512,3))
X_val = np.reshape(X_val,(192,3,512,512))
X_oracle = np.concatenate((X_train,X_oracle),axis=0)
Y_oracle = np.concatenate((Y_train,Y_oracle),axis=0)
X_oracle,Y_oracle = shuffle(X_oracle,Y_oracle,random_state=42)
X_oracle_descriptor = calculate_descriptors(X_oracle)
X_train,Y_train = create_trainset()
X_train = np.reshape(X_train,(No_of_images_to_start_with,3,512,512))
def train_and_evaluate(batch_size=8, img_dir='mess_cb', lesion_dir='eoptha_cb',
out_size=32, f_size=31, horizontal_flip=True, vertical_flip=True,
width_shift_range=0.03, height_shift_range=0.03, rotation_range=360,
zoom_range=0.03, exp_name='exp0.hdf5', n_blocks=4, context_blocks=1,
patience=75):
train_it, val_it, test_it = train.get_data_iterators(batch_size=batch_size, horizontal_flip=horizontal_flip,
vertical_flip=vertical_flip, width_shift_range=width_shift_range,
height_shift_range=height_shift_range, rotation_range=rotation_range,
zoom_range=zoom_range, data_dir=img_dir, target_size=(512, 512),
rescale=1/255., fill_mode='constant', load_train_data=True, color_mode='rgb')
train_it_e, val_it_e, test_it_e = train.get_lesion_iterators(out_size=out_size, f_size=f_size, batch_size=batch_size,
horizontal_flip=horizontal_flip, vertical_flip=vertical_flip,
width_shift_range=width_shift_range, height_shift_range=height_shift_range,
rotation_range=rotation_range, zoom_range=zoom_range,
base_dir=lesion_dir)
model, ld, _, _ = models.lesion_detector_parameterized(64, l2=.0, input_size=512, n_blocks=n_blocks,
drop_p=.0, context_blocks=context_blocks, lr=2e-4)
checkpointer = ModelCheckpoint(filepath=exp_name, verbose=1, save_best_only=True, save_weights_only=False)
callbacks = [checkpointer]
epochs = 4000
cur_patience = patience
best_loss = 10
for e in range(epochs):
print 'EPOCH {0}/{1}'.format(e + 1, epochs)
start = time.time()
r_vals = np.random.choice(2, 191)
for r in r_vals:
if r == 0:
x, y = next(train_it_e)
ld.train_on_batch(x, y)
else:
x, y = next(train_it)
model.train_on_batch(x, y)
ld.fit_generator(train_it_e, batch_size, 1, verbose=2)
hist = model.fit_generator(train_it, batch_size, 1, validation_data=val_it, nb_val_samples=192, verbose=2, callbacks=callbacks)
print '{0}s'.format(int(time.time() - start))
print ''
val_loss = hist.history['val_loss'][0]
if val_loss < best_loss:
best_loss = val_loss
cur_patience = patience
else:
cur_patience -= 1
if cur_patience == 0:
break
print 'Evaluating model {0}:'.format(exp_name)
model.load_weights(exp_name)
print 'Train:', model.evaluate_generator(train_it, 768)
print 'Validation:', model.evaluate_generator(val_it, 192)
print 'Test:', model.evaluate_generator(test_it, 240)
print 'Train AUC = {0}'.format(get_auc_score(train_it, 768))
print 'Validation AUC = {0}'.format(get_auc_score(val_it, 192))
print 'Test AUC = {0}'.format(get_auc_score(test_it, 240))
aug_params = {
'horizontal_flip': args.horizontal_flip,
'vertical_flip': args.vertical_flip,
'width_shift_range': args.width_shift_range,
'height_shift_range': args.height_shift_range,
'rotation_range': args.rotation_range,
'zoom_range': args.zoom_range
}
"""
Load data
"""
train_it, val_it, test_it = train.get_data_iterators(
batch_size=args.batch_size,
data_dir=args.weak_dir,
target_size=(args.input_size, args.input_size),
rescale=1 / 255.,
fill_mode='constant',
load_train_data=args.load_data,
color_mode='rgb',
**aug_params)
"""
Train
"""
eyequal, heatmap = models.quality_assessment(
args.nf,
input_size=args.input_size,
n_blocks=args.n_blocks,
lr=args.lr,
pooling_wreg=args.pooling_wreg,
pooling_breg=args.pooling_breg)
eyequal.fit_generator(train_it,