def __init__(self, current_price, pretrained=False):
# agent config
self.state_size = 5 # normalized previous days
self.cash_in_hand = 6000
self.total_share = 20
self.action_size = 3 # [sit, buy, sell]
self.inventory = []
self.memory = deque(maxlen=10000)
self.first_iter = True
self.initial_price = current_price
self.writer = SummaryWriter()
if pretrained:
self.model = load()
else:
self.model = model()
for i in range(20):
self.inventory.append(self.initial_price)
self.gamma = 0.95 # affinity for long term reward
self.epsilon = 1.0
self.epsilon_min = 0.01
self.epsilon_decay = 0.995
self.learning_rate = 0.001
self.n_iter = 1
self.reset_every = 1000
self.target_model = clone_model(self.model)
self.target_model.set_weights(self.model.get_weights())
def main(my_model):
path_results = 'test'
path_test = 'database/test'
create_output('output')
"""
Return the model trained. If don't have already a model trained, will
going to train the model.
"""
_model = model(shape=(32, 32, 3),
num_classes=4,
is_plot=True,
path_database='database/{}',
epochs=8,
my_model=my_model)
"""
Path 'test/' is create exclusively to save model corrected images.
After zip the directory 'test/' it is removed.
"""
create_path(path_results)
imgs = []
rotate_img = []
predicts = []
labels = {
'0': 'rotated_left',
'1': 'rotated_right',
'2': 'upright',
'3': 'upside_down',
}
for pred, img in zip(_model.predictions, os.listdir(path_test)):
pred = str(pred)
imgs.append(img)
predicts.append(labels.get(pred))
rotate_img.append(
rotate(path=path_test,
image=img,
orientation=pred,
save_path=path_results))
"""
Dump in 'output/': numpy array with corrected orientation faces,
csv with the labels resulting from the model and the zip files with
images after corrected orientation process.
"""
dump_results(imgs, predicts, path_results, rotate_img)
logging.info('The result is dump in "output/"!')
remove_path(path_results)
def grad(model, img, detector_mask, matching_true_boxes, class_one_hot, true_boxes, training=True): with tf.GradientTape() as tape: y_pred = model(img, training) loss, sub_loss = yolov2_loss(detector_mask, matching_true_boxes, class_one_hot, true_boxes, y_pred) return loss, sub_loss, tape.gradient(loss, model.trainable_variables)
IR = IR_to_uint8(get_data(IR_))
FLAIR = to_uint8(get_data(FLAIR_))
y.append(np.array(get_data(seg_)==label).astype(np.uint8)[None,...])
X_T1.append(T1[None,...])
X_IR.append(IR[None,...])
X_FLAIR.append(FLAIR[None,...])
X_T1 = np.array(X_T1)
X_FLAIR = np.array(X_FLAIR)
X_IR = np.array(X_IR)
y = np.array(y)
X_T1_val = histeq(to_uint8(get_data_with_skull_scraping(T1_val)))[None,None,...]
X_FLAIR_val = to_uint8(get_data(FLAIR_val))[None,None,...]
X_IR_val = IR_to_uint8(get_data(IR_val))[None,None,...]
y_val = np.array(get_data(segm_val)==CLASS).astype(np.uint8)[None,...]
print("STARTING TRAINING...")
model_ = model()
model_.compile('adam',dice_loss,[dice_coefficient])
model_.summary()
history = model_.fit([X_T1,X_FLAIR,X_IR],y=y,validation_data=([X_T1_val,X_FLAIR_val,X_IR_val],y_val),epochs=EPOCHS,callbacks=[keras.callbacks.ModelCheckpoint(
'weights/label'+str(label)+'/Model.val_dice_coefficient={val_dice_coefficient:.5f}.h5',
monitor='val_dice_coefficient',
verbose=0,
save_best_only=True,
save_weights_only=False,
mode='max',
period=1
)])
print("FINISHED TRAINING...")
print("Saving training history")
with open('history/trainHistoryDict'+str(label)+'.pickle', 'wb') as file_pi:
pickle.dump(history.history, file_pi)
help='Number of epochs (10 by default)')
args = parser.parse_args()
img_dir_path = args.imgs
img_dir_name = os.path.basename(os.path.normpath(img_dir_path))
label_dir_path = args.labels
label_dir_name = os.path.basename(os.path.normpath(label_dir_path))
epochs = args.epochs
imgs = glob.glob(img_dir_path + '/' + img_dir_name + '/*')
labels = glob.glob(label_dir_path + '/' + label_dir_name + '/*')
if len(imgs) != len(labels):
raise ValueError('Error: Different number of images and labels')
if epochs is None:
epochs = 10
img_path = imgs[0]
label_path = labels[0]
fcn_model = model()
train_generator = utils.image_generator(img_path, label_path, img_dir_path,
label_dir_path)
utils.train_model(fcn_model, train_generator, len(imgs),
epochs).save('model.h5')