resnet_settings = [[5, 5, 20], [2, [1, 1]], [8, [2, 2]]]
total_classes = 2
output_path = 'output'
nsm_iou_threshold = 0.1
nsm_score_threshold = 0.9
nsm_max_output_size = 500
total_test_examples = 100
abox4d = genanchors(isize=ishape[:2], ssize=ssize, asizes=asizes)
abox_4dtensor = tf.constant(value=abox4d, dtype='float32')
abox_2dtensor = tf.reshape(tensor=abox_4dtensor, shape=[-1, 4])
model = build_infer_model(ishape=ishape,
resnet_settings=resnet_settings,
k=len(asizes),
total_classes=total_classes,
abox_2dtensor=abox_2dtensor,
nsm_iou_threshold=nsm_iou_threshold,
nsm_score_threshold=nsm_score_threshold,
nsm_max_output_size=nsm_max_output_size)
model.summary()
# model.load_weights('{}/weights.h5'.format(output_path), by_name=True)
gen = genx(anno_file_path=test_anno_file_path,
image_dir=test_image_dir,
ishape=ishape,
mode='test')
for batch in range(total_test_examples):
batchx_4dtensor, _, image_id = next(gen)
print('{}: 1'.format(datetime.now().time()), end='\n')
boxclz_2dtensor, valid_outputs = model.predict_on_batch(
asizes=asizes[0]) # (h1 * w1 * k1, 4)
a2box2d = genanchors(isize=ishape[:2], ssize=ssizes[1],
asizes=asizes[1]) # (h2 * w2 * k2, 4)
a3box2d = genanchors(isize=ishape[:2], ssize=ssizes[2],
asizes=asizes[2]) # (h3 * w3 * k3, 4)
abox2d = np.concatenate([a1box2d, a2box2d, a3box2d],
axis=0) # (h1*w1*k1 + h2*w2*k2 + h3*w3*k3, 4)
abox_2dtensor = tf.constant(
value=abox2d, dtype='float32') # (h1*w1*k1 + h2*w2*k2 + h3*w3*k3, 4)
model = build_infer_model(ishape=ishape,
resnet_settings=resnet_settings,
top_down_pyramid_size=top_down_pyramid_size,
k=[len(asizes[0]),
len(asizes[1]),
len(asizes[2])],
total_classes=total_classes,
abox_2dtensor=abox_2dtensor,
nsm_iou_threshold=nsm_iou_threshold,
nsm_score_threshold=nsm_score_threshold,
nsm_max_output_size=nsm_max_output_size)
# model.summary()
model.load_weights('{}/weights_best_precision.h5'.format(output_path),
by_name=True)
model.save('{}/model'.format(output_path))
converter = tf.lite.TFLiteConverter.from_saved_model(
'{}/model'.format(output_path))
# converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.experimental_new_converter = True
# converter.allow_custom_ops=True