def train_and_test():
# load network
model = vgg11(365, 1)
model.compile(
loss='categorical_crossentropy',
# Learning rate will be set by lr_finder
optimizer=SGD(lr=0.0, momentum=0.9),
metrics=['accuracy', top_5])
# load data
img_size = (224, 224)
color_mode = 'grayscale'
batch_size = 64
train_dir = '/usr/local/data/gabriel/places365_line_drawings/train'
test_dir = '/usr/local/data/gabriel/places365_line_drawings/val'
# fixed for places365
nb_train_samples = 1803460.
nb_test_samples = 36500.
train_datagen = ImageDataGenerator(rescale=1. / 255,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True)
# no test data for now
# test_datagen = ImageDataGenerator(rescale=1. / 255)
train_gen = train_datagen.flow_from_directory(train_dir,
target_size=img_size,
batch_size=batch_size,
class_mode='categorical',
color_mode=color_mode)
# no test data for now
# test_gen = test_datagen.flow_from_directory(
# test_dir,
# target_size = img_size,
# batch_size = batch_size,
# class_mode = 'categorical',
# color_mode = color_mode
# )
# find best learning rate
lr_finder = LRFinder(min_lr=1e-5,
max_lr=1e-2,
steps_per_epoch=np.ceil(nb_train_samples /
batch_size),
epochs=4)
model.fit_generator(train_gen,
steps_per_epoch=np.ceil(nb_train_samples / batch_size),
epochs=4,
callbacks=[lr_finder])
# save loss and learning rate plots to files
lr_finder.plot_loss('loss.png')
lr_finder.plot_lr('lr.png')
def _main():
annotation_path = 'train.txt'
classes_path = 'model_data/openimgs_classes.txt'
anchors_path = 'model_data/yolo_anchors.txt'
class_names = get_classes(classes_path)
num_classes = len(class_names)
anchors = get_anchors(anchors_path)
input_shape = (416, 416) # multiple of 32, hw
# use darknet53 weights
#model = create_model(input_shape, anchors, num_classes,
# freeze_body=2, weights_path='model_data/darknet53_weights.h5')
model = create_model(input_shape,
anchors,
num_classes,
freeze_body=0,
weights_path='logs/001/trained_weights_stage_2.h5')
val_split = 0.1
with open(annotation_path) as f:
lines = f.readlines()
#np.random.seed(10101)
np.random.shuffle(lines)
#np.random.seed(None)
num_val = int(len(lines) * val_split)
num_val = 10000 if num_val > 10000 else num_val
num_train = len(lines) - num_val
if True:
batch_size = 6
lr_finder = LRFinder(min_lr=1e-10,
max_lr=2e-2,
steps_per_epoch=TOTAL_ITERATIONS,
epochs=1)
for i in range(len(model.layers)):
model.layers[i].trainable = True
model.compile(optimizer=SGD(lr=1e-8),
loss={
'yolo_loss': lambda y_true, y_pred: y_pred
})
print('train on {} samples, val on {} samples, with batch size {}.'.
format(num_train, num_val, batch_size))
model.fit_generator(data_generator_wrapper(lines[:num_train],
batch_size, input_shape,
anchors, num_classes),
steps_per_epoch=TOTAL_ITERATIONS,
validation_data=data_generator_wrapper(
lines[num_train:], batch_size, input_shape,
anchors, num_classes),
validation_steps=1,
epochs=1,
initial_epoch=0,
callbacks=[lr_finder])
lr_finder.save_history('lr_finder_loss.csv')
lr_finder.plot_loss('lr_finder_loss.png')
min_lr=0.000001,
cooldown=3,
verbose=1)
stop_on_nan = keras.callbacks.TerminateOnNaN()
# LR finder
if opt.find_lr:
# pre-train to avoid model being too far away from interesting range
history = model.fit_generator(gen_x_train,
epochs=2,
verbose=1,
callbacks=[clr])
lr_finder = LRFinder(model)
lr_finder.find_generator(gen_x_train, 0.00001, 1.0, 5)
lr_finder.plot_loss()
import pdb
pdb.set_trace()
# Run training
if not opt.notrain:
# Train classifier
history = model.fit_generator(
gen_x_train,
epochs=epochs,
verbose=1, # switch to 1 for more verbosity
callbacks=[early_stopping, clr,
stop_on_nan], #, reduce_lr], #, lr, reduce_lr],
# callbacks=[early_stopping, reduce_lr], #, lr, reduce_lr],
use_multiprocessing=True,
workers=8,
def main(batch_size: int = 24,
epochs: int = 384,
train_path: str = 'train',
val_path: str = 'val',
multi_gpu_weights=None,
weights=None,
workers: int = 8,
find_lr: bool = False):
keras_model = MobileDetectNetModel.complete_model()
keras_model.summary()
if weights is not None:
keras_model.load_weights(weights, by_name=True)
train_seq = MobileDetectNetSequence(train_path,
stage="train",
batch_size=batch_size)
val_seq = MobileDetectNetSequence(val_path,
stage="val",
batch_size=batch_size)
keras_model = keras.utils.multi_gpu_model(keras_model,
gpus=[0, 1],
cpu_merge=True,
cpu_relocation=False)
if multi_gpu_weights is not None:
keras_model.load_weights(multi_gpu_weights, by_name=True)
callbacks = []
def region_loss(classes):
def loss_fn(y_true, y_pred):
# Don't penalize bounding box errors when there is no object present
return 10 * classes * K.abs(y_pred - y_true)
return loss_fn
keras_model.compile(optimizer=SGD(),
loss=[
'mean_absolute_error',
region_loss(
keras_model.get_layer('classes').output),
'binary_crossentropy'
])
if find_lr:
from lr_finder import LRFinder
lr_finder = LRFinder(keras_model)
lr_finder.find_generator(train_seq,
start_lr=0.000001,
end_lr=1,
epochs=5)
lr_finder.plot_loss()
return
filepath = "weights-{epoch:02d}-{val_loss:.4f}-multi-gpu.hdf5"
checkpoint = ModelCheckpoint(filepath,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min')
callbacks.append(checkpoint)
sgdr_sched = SGDRScheduler(0.00001,
0.01,
steps_per_epoch=np.ceil(
len(train_seq) / batch_size),
mult_factor=1.5)
callbacks.append(sgdr_sched)
keras_model.fit_generator(
train_seq,
validation_data=val_seq,
epochs=epochs,
steps_per_epoch=np.ceil(len(train_seq) / batch_size),
validation_steps=np.ceil(len(val_seq) / batch_size),
callbacks=callbacks,
use_multiprocessing=True,
workers=workers,
shuffle=True)