def main():
print('Loading APC2015berkeley...')
dataset = load_APC2015berkeley()
N = len(dataset.target)
print('''
N: {0}
dataset: {1}'''.format(N, dataset))
bg_label = dataset.target_names.index('__background__')
n_labels = len(dataset.target_names)
max_batch_size = 10
for i in xrange(0, N, max_batch_size):
t_start = time.time()
fname_batch = dataset.filenames[i:i+max_batch_size]
label_batch = dataset.target[i:i+max_batch_size]
blob_batch, bbox_batch, label_batch, roi_delta_batch =\
load_batch_APC2015berkeley(fname_batch, label_batch,
bg_label, n_labels)
# show stats
elapsed_time = time.time() - t_start
head_fname_batch = ['/'.join(f.split('/')[-2:]) for f in fname_batch]
print('''
elapsed_time: {0} [s]
fnames: {1}
blob: {2}
bboxes: {3}
label: {4}
roi_delta: {5}'''.format(elapsed_time, head_fname_batch,
blob_batch.shape, bbox_batch.shape,
label_batch.shape, roi_delta_batch.shape))
def batch_loop_train(self, batch_size):
N = len(self.train_fnames)
random_index = np.random.randint(0, N, N)
for i in xrange(0, N, batch_size):
batch_index = random_index[i:i+batch_size]
batch_fnames = self.train_fnames[batch_index]
batch_labels = self.train_target[batch_index]
blobs, bboxes, t_labels, t_bboxes = load_batch_APC2015berkeley(
fnames=batch_fnames,
labels=batch_labels,
bg_label=self.dataset.background_label,
n_labels=len(self.dataset.target_names),
)
blobs, bboxes, t_labels, t_bboxes =\
map(cuda.to_gpu, [blobs, bboxes, t_labels, t_bboxes])
self.optimizer.zero_grads()
volatile = 'OFF'
blobs = Variable(blobs, volatile=volatile)
bboxes = Variable(bboxes, volatile=volatile)
t_labels = Variable(t_labels, volatile=volatile)
t_bboxes = Variable(t_bboxes, volatile=volatile)
loss = self.model(blobs, bboxes, (t_labels, t_bboxes), train=True)
loss.backward()
self.optimizer.update()
def batch_loop_test(self, batch_size):
sum_loss = 0
N = len(self.test_fnames)
random_index = np.random.randint(0, N, N)
for i in xrange(0, N, batch_size):
batch_index = random_index[i:i+batch_size]
batch_fnames = self.test_fnames[batch_index]
batch_labels = self.test_target[batch_index]
blobs, bboxes, t_labels, t_bboxes = load_batch_APC2015berkeley(
fnames=batch_fnames,
labels=batch_labels,
bg_label=self.dataset.background_label,
n_labels=len(self.dataset.target_names),
)
loss = self.model(blobs, bboxes, (t_labels, t_bboxes), train=False)
sum_loss += loss.data
print(loss.data / batch_size)
self.sum_loss.append(sum_loss / N)
