# Set some seeds.
use_cuda = use_cuda and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {"num_workers": multiprocessing.cpu_count(), "pin_memory": True} if use_cuda else {}
# Set up GPU training and some training parameters.
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
# Set up the dataset and the DataLoader.
transform = transforms.Compose([transforms.ToTensor()])
train_dataset = ShapesDataset(transform=transform, csv=train_csv)
test_dataset = ShapesDataset(transform=transform, csv=test_csv)
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers)
# Instantiate the model and optimizer.
model = BetaVAE().to(device)
optimizer = optim.Adam(model.parameters(), lr=lr)
# Execute model training and testing.
if __name__ == "__main__":
start_epoch = model.load_last_model(ckpt_dir) + 1
train_losses, test_losses = utils.read_log(log_dir, ([], []))
np.random.seed(10101)
np.random.shuffle(imglist)
train_imglist = imglist[:int(count * 0.9)]
val_imglist = imglist[int(count * 0.9):]
MODEL_DIR = "logs"
COCO_MODEL_PATH = "model_data/mask_rcnn_coco.h5"
config = ShapesConfig()
# 计算训练集和验证集长度
config.STEPS_PER_EPOCH = len(train_imglist)
config.VALIDATION_STEPS = len(val_imglist)
config.display()
# 训练数据集准备
dataset_train = ShapesDataset()
dataset_train.load_shapes(len(train_imglist), img_floder, mask_floder,
train_imglist, yaml_floder)
dataset_train.prepare()
# 验证数据集准备
dataset_val = ShapesDataset()
dataset_val.load_shapes(len(val_imglist), img_floder, mask_floder,
val_imglist, yaml_floder)
dataset_val.prepare()
# 获得训练模型
model = get_train_model(config)
model.summary()
model.load_weights(COCO_MODEL_PATH, by_name=True, skip_mismatch=True)
def run(is_distributed, logs_dir):
from dataset import ShapesDataset
from mrcnn.config import Config
######################
class ShapesConfig(Config):
NAME = "shapes"
GPU_COUNT = 2
IMAGES_PER_GPU = 2
NUM_CLASSES = 1 + 3
IMAGE_MIN_DIM = 128
IMAGE_MAX_DIM = 128
RPN_ANCHOR_SCALES = (8, 16, 32, 64, 128)
TRAIN_ROIS_PER_IMAGE = 32
STEPS_PER_EPOCH = 10
VALIDATION_STEPS = 5
config = ShapesConfig()
config.display()
# Training dataset
dataset_train = ShapesDataset()
dataset_train.load_shapes(500000, config.IMAGE_SHAPE[0],
config.IMAGE_SHAPE[1])
dataset_train.prepare()
# Validation dataset
dataset_val = ShapesDataset()
dataset_val.load_shapes(5000, config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1])
dataset_val.prepare()
if is_distributed:
import mrcnn.distributed_model as modellib
else:
import mrcnn.model as modellib
from mrcnn import utils
# Local path to trained weights file
COCO_MODEL_PATH = os.path.join(ROOT_DIR, "mask_rcnn_coco.h5")
# Download COCO trained weights from Releases if needed
if not os.path.exists(COCO_MODEL_PATH):
utils.download_trained_weights(COCO_MODEL_PATH)
# number of found devices by TF
from tensorflow.python.client import device_lib
device_lib.list_local_devices()
# Create model in training mode
model = modellib.MaskRCNN("training", config, logs_dir)
# Load weights trained on MS COCO, but skip layers that
# are different due to the different number of classes
# See README @ https://github.com/matterport/Mask_RCNNfor instructions to download the COCO weights
model.load_weights(COCO_MODEL_PATH,
by_name=True,
exclude=[
"mrcnn_class_logits", "mrcnn_bbox_fc", "mrcnn_bbox",
"mrcnn_mask"
])
model.train(dataset_train,
dataset_val,
learning_rate=config.LEARNING_RATE,
epochs=1000,
layers='heads')