def setUp(self):
chainer.set_debug(True)
np.random.seed(0)
dataset = VOC('train')
img, im_info, bbox = dataset[1]
self.x = Variable(img[None, ...])
self.im_info = Variable(im_info[None, ...])
self.gt_boxes = Variable(bbox[None, ...])
def warmup(model, gpu_ids):
train_dataset = VOC('train')
iterator = iterators.MultiprocessIterator(train_dataset,
1,
shared_mem=10000000)
batch = iterator.next()
img, img_info, bbox = batch[0]
img = chainer.Variable(img[None, ...])
img_info = chainer.Variable(img_info[None, ...])
bbox = chainer.Variable(bbox[None, ...])
for gpu_id in gpu_ids:
if gpu_id >= 0:
img.to_gpu(gpu_id)
img_info.to_gpu(gpu_id)
bbox.to_gpu(gpu_id)
model.to_gpu(gpu_id)
model.rcnn_train = True
model(img, img_info, bbox)
model.rpn_train = True
model(img, img_info, bbox)
if gpu_id >= 0:
model.to_cpu()
def warmup(model, iterator, gpu_id=0):
batch = iterator.next()
img, img_info, bbox = concat_examples(batch, gpu_id)
img = chainer.Variable(img)
img_info = chainer.Variable(img_info)
bbox = chainer.Variable(bbox)
model.rcnn_train = True
model(img, img_info, bbox)
model.rpn_train = True
model(img, img_info, bbox)
if __name__ == '__main__':
batchsize = 1
train_dataset = VOC('train')
valid_dataset = VOC('val')
train_iter = iterators.SerialIterator(train_dataset, batchsize)
model = FasterRCNN()
model.to_gpu(0)
warmup(model, train_iter)
model.rcnn_train = True
serializers.load_npz('tests/train_test/snapshot_10000', model)
# optimizer = optimizers.Adam()
# optimizer.setup(model)
optimizer = optimizers.MomentumSGD(lr=0.001)
optimizer.setup(model)
devices['gpu{}'.format(gid)] = gid
# Instantiate a model
model = get_model(args.model_file, args.model_name, devices['main'],
args.rpn_in_ch, args.rpn_out_ch, args.n_anchors,
args.feat_stride, args.anchor_scales, args.num_classes,
args.spatial_scale, args.rpn_sigma, args.sigma,
args.trunk_model, True, result_dir)
# Instantiate a optimizer
optimizer = get_optimizer(model, args.opt, args.lr, args.adam_alpha,
args.adam_beta1, args.adam_beta2, args.adam_eps,
args.weight_decay)
# Setting up datasets
train = VOC('train', False)
valid = VOC('val', False)
print('train: {}, valid: {}'.format(len(train), len(valid)))
# Iterator
train_iter = iterators.MultiprocessIterator(train, args.batchsize)
valid_iter = iterators.MultiprocessIterator(valid,
args.valid_batchsize,
repeat=False,
shuffle=False)
# Updater
updater = ParallelUpdater(train_iter, optimizer, devices=devices)
trainer = training.Trainer(updater, (args.epoch, 'epoch'), out=result_dir)
# Extentions
def setUp(self):
self.dataset = VOC('train', False)
if not os.path.exists('tests/imgs'):
os.makedirs('tests/imgs')