2martens/masterthesis
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Made ssd_train function compatible with clean code standards

Browse Source 
Signed-off-by: Jim Martens <github@2martens.de>
This commit is contained in:
Jim Martens
2019-07-10 15:25:18 +02:00
parent ed3fec54af
commit 21cef4a09e
1 changed files with 219 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

236
src/twomartens/masterthesis/cli.py
View File

@ -26,9 +26,12 @@ Functions:
prepare(...): prepares the SceneNet ground truth data
"""
import argparse
from typing import Callable, Union, Tuple, Sequence, Optional, Generator
import math
import tensorflow as tf
from twomartens.masterthesis import config as conf
@ -169,32 +172,149 @@ def _train_execute_action(args: argparse.Namespace, on_ssd: callable, on_auto_en
def _ssd_train(args: argparse.Namespace) -> None:
import os
import pickle
import tensorflow as tf
from twomartens.masterthesis import data
from twomartens.masterthesis import debug
from twomartens.masterthesis import ssd
from twomartens.masterthesis.ssd_keras.models import keras_ssd300
from twomartens.masterthesis.ssd_keras.models import keras_ssd300_dropout
_init_eager_mode()
batch_size, image_size, learning_rate, steps_per_val_epoch, nr_classes, \
iou_treshold, dropout_rate, top_k, nr_trajectories, \
coco_path, summary_path, weights_path, train_gt_path, val_gt_path, \
save_train_images, save_summaries = _ssd_train_get_config_values(conf.get_property)
use_dropout = _ssd_is_dropout(args)
summary_path, weights_path, \
pre_trained_weights_file = _ssd_train_prepare_paths(args, summary_path, weights_path)
file_names_train, instances_train, \
file_names_val, instances_val = _ssd_train_load_gt(train_gt_path, val_gt_path)
ssd_model, predictor_sizes = ssd.get_model(use_dropout,
keras_ssd300_dropout.ssd_300_dropout,
keras_ssd300.ssd_300,
image_size,
nr_classes,
"training",
iou_treshold,
dropout_rate,
top_k,
pre_trained_weights_file)
train_generator, train_length, val_generator, val_length = _ssd_train_get_generators(
data.load_scenenet_data,
file_names_train,
instances_train,
file_names_val,
instances_val,
coco_path,
batch_size,
image_size,
nr_trajectories,
predictor_sizes
)
train_length = _ssd_debug_save_images(args, save_train_images, debug.save_ssd_train_images,
summary_path, batch_size, train_generator, train_length)
nr_batches_train = _get_nr_batches(train_length, batch_size)
tensorboard_callback = _ssd_get_tensorboard_callback(args, save_summaries, summary_path)
history = _ssd_train_call(
args,
ssd.train_keras,
train_generator,
nr_batches_train,
val_generator,
steps_per_val_epoch,
ssd_model,
weights_path,
learning_rate,
tensorboard_callback
)
_ssd_save_history(summary_path, history)
def _init_eager_mode() -> None:
tf.enable_eager_execution()
batch_size = conf.get_property("Parameters.batch_size")
image_size = conf.get_property("Parameters.ssd_image_size")
use_dropout = False if args.network == "ssd" else True
summary_path = conf.get_property("Paths.summaries")
def _ssd_train_get_config_values(config_get: Callable[[str], Union[str, float, int, bool]]
) -> Tuple[int, int, int, int, int, float, float, int, int,
str, str, str, str, str,
bool, bool]:
batch_size = config_get("Parameters.batch_size")
image_size = config_get("Parameters.ssd_image_size")
learning_rate = config_get("Parameters.learning_rate")
steps_per_val_epoch = config_get("Parameters.steps_per_val_epoch")
nr_classes = config_get("Parameters.nr_classes")
iou_threshold = config_get("Parameters.ssd_iou_threshold")
dropout_rate = config_get("Parameters.ssd_dropout_rate")
top_k = config_get("Parameters.ssd_top_k")
nr_trajectories = config_get("Parameters.nr_trajectories")
coco_path = config_get("Paths.coco")
summary_path = config_get("Paths.summaries")
weights_path = config_get("Paths.weights")
train_gt_path = config_get('Paths.scenenet_gt_train')
val_gt_path = config_get('Paths.scenenet_gt_val')
save_train_images = config_get("Debug.train_images")
save_summaries = config_get("Debug.summaries")
return (
batch_size,
image_size,
learning_rate,
steps_per_val_epoch,
nr_classes,
iou_threshold,
dropout_rate,
top_k,
nr_trajectories,
#
coco_path,
summary_path,
weights_path,
train_gt_path,
val_gt_path,
#
save_train_images,
save_summaries
)
def _ssd_is_dropout(args: argparse.Namespace) -> bool:
return False if args.network == "ssd" else True
def _ssd_train_prepare_paths(args: argparse.Namespace,
summary_path: str, weights_path: str) -> Tuple[str, str, str]:
import os
summary_path = f"{summary_path}/{args.network}/train/{args.iteration}"
os.makedirs(summary_path, exist_ok=True)
weights_path = conf.get_property("Paths.weights")
coco_path = conf.get_property("Paths.coco")
pre_trained_weights_file = f"{weights_path}/{args.network}/VGG_coco_SSD_300x300_iter_400000.h5"
weights_path = f"{weights_path}/{args.network}/train/"
os.makedirs(summary_path, exist_ok=True)
os.makedirs(weights_path, exist_ok=True)
# load prepared ground truth
train_gt_path = conf.get_property('Paths.scenenet_gt_train')
val_gt_path = conf.get_property('Paths.scenenet_gt_val')
return summary_path, weights_path, pre_trained_weights_file
def _ssd_train_load_gt(train_gt_path: str, val_gt_path: str
) -> Tuple[Sequence[Sequence[str]],
Sequence[Sequence[Sequence[dict]]],
Sequence[Sequence[str]],
Sequence[Sequence[Sequence[dict]]]]:
import pickle
with open(f"{train_gt_path}/photo_paths.bin", "rb") as file:
file_names_train = pickle.load(file)
with open(f"{train_gt_path}/instances.bin", "rb") as file:
@ -204,62 +324,100 @@ def _ssd_train(args: argparse.Namespace) -> None:
with open(f"{val_gt_path}/instances.bin", "rb") as file:
instances_val = pickle.load(file)
# model
if use_dropout:
ssd_model = ssd.DropoutSSD(mode='training', weights_path=pre_trained_weights_file)
else:
ssd_model = ssd.SSD(mode='training', weights_path=pre_trained_weights_file)
return file_names_train, instances_train, file_names_val, instances_val
def _ssd_train_get_generators(load_data: callable,
file_names_train: Sequence[Sequence[str]],
instances_train: Sequence[Sequence[Sequence[dict]]],
file_names_val: Sequence[Sequence[str]],
instances_val: Sequence[Sequence[Sequence[dict]]],
coco_path: str,
batch_size: int,
image_size: int,
nr_trajectories: int,
predictor_sizes: Sequence[Sequence[int]]) -> Tuple[Generator, int, Generator, int]:
if nr_trajectories == -1:
nr_trajectories = None
train_generator, train_length = \
data.load_scenenet_data(file_names_train, instances_train, coco_path,
predictor_sizes=ssd_model.predictor_sizes,
load_data(file_names_train, instances_train, coco_path,
predictor_sizes=predictor_sizes,
batch_size=batch_size,
resized_shape=(image_size, image_size),
image_size=image_size,
training=True, evaluation=False, augment=False,
nr_trajectories=1)
nr_trajectories=nr_trajectories)
val_generator, val_length = \
data.load_scenenet_data(file_names_val, instances_val, coco_path,
predictor_sizes=ssd_model.predictor_sizes,
load_data(file_names_val, instances_val, coco_path,
predictor_sizes=predictor_sizes,
batch_size=batch_size,
resized_shape=(image_size, image_size),
image_size=image_size,
training=False, evaluation=False, augment=False,
nr_trajectories=1)
del file_names_train, instances_train, file_names_val, instances_val
nr_trajectories=nr_trajectories)
if args.debug and conf.get_property("Debug.train_images"):
from twomartens.masterthesis import debug
return train_generator, train_length, val_generator, val_length
def _ssd_debug_save_images(args: argparse.Namespace, save_images_on_debug: bool, save_images: callable,
summary_path: str, batch_size: int,
train_generator: Generator, train_length: int) -> int:
if args.debug and save_images_on_debug:
train_data = next(train_generator)
train_length -= batch_size
train_images = train_data[0]
train_labels = train_data[1]
debug.save_ssd_train_images(train_images, train_labels, summary_path)
save_images(train_images, train_labels, summary_path)
nr_batches_train = int(math.floor(train_length / batch_size))
nr_batches_val = int(math.floor(val_length / batch_size))
return train_length
if args.debug and conf.get_property("Debug.summaries"):
def _ssd_get_tensorboard_callback(args: argparse.Namespace, save_summaries_on_debug: bool,
summary_path: str) -> Union[None, tf.keras.callbacks.TensorBoard]:
if args.debug and save_summaries_on_debug:
tensorboard_callback = tf.keras.callbacks.TensorBoard(
log_dir=summary_path
)
else:
tensorboard_callback = None
history = ssd.train_keras(
return tensorboard_callback
def _get_nr_batches(data_length: int, batch_size: int) -> int:
return int(math.floor(data_length / batch_size))
def _ssd_train_call(args: argparse.Namespace, train_function: callable,
train_generator: Generator, nr_batches_train: int,
val_generator: Generator, nr_batches_val: int,
model: tf.keras.models.Model,
weights_path: str, learning_rate: float,
tensorboard_callback: Optional[tf.keras.callbacks.TensorBoard]) -> tf.keras.callbacks.History:
history = train_function(
train_generator,
nr_batches_train,
val_generator,
conf.get_property("Parameters.steps_per_val_epoch"),
ssd_model,
nr_batches_val,
model,
weights_path,
args.iteration,
initial_epoch=0,
nr_epochs=args.num_epochs,
lr=conf.get_property("Parameters.learning_rate"),
lr=learning_rate,
tensorboard_callback=tensorboard_callback
)
return history
def _ssd_save_history(summary_path: str, history: tf.keras.callbacks.History) -> None:
import pickle
with open(f"{summary_path}/history", "wb") as file:
pickle.dump(history.history, file)