def _do_test_3D_models(recognizer,
target_layer_name,
input_shape,
num_classes=400):
blended_imgs_target_shape, preds_target_shape = _get_target_shapes(
input_shape, num_classes=num_classes, model_type='3D')
demo_inputs = generate_gradcam_inputs(input_shape, '3D')
# parrots 3dconv is only implemented on gpu
if torch.__version__ == 'parrots':
if torch.cuda.is_available():
recognizer = recognizer.cuda()
demo_inputs['imgs'] = demo_inputs['imgs'].cuda()
demo_inputs['label'] = demo_inputs['label'].cuda()
gradcam = GradCAM(recognizer, target_layer_name)
blended_imgs, preds = gradcam(demo_inputs)
assert blended_imgs.size() == blended_imgs_target_shape
assert preds.size() == preds_target_shape
blended_imgs, preds = gradcam(demo_inputs, True)
assert blended_imgs.size() == blended_imgs_target_shape
assert preds.size() == preds_target_shape
else:
gradcam = GradCAM(recognizer, target_layer_name)
blended_imgs, preds = gradcam(demo_inputs)
assert blended_imgs.size() == blended_imgs_target_shape
assert preds.size() == preds_target_shape
blended_imgs, preds = gradcam(demo_inputs, True)
assert blended_imgs.size() == blended_imgs_target_shape
assert preds.size() == preds_target_shape
def _do_test_2D_models(recognizer,
target_layer_name,
input_shape,
num_classes=400,
device='cpu'):
demo_inputs = generate_gradcam_inputs(input_shape)
demo_inputs['imgs'] = demo_inputs['imgs'].to(device)
demo_inputs['label'] = demo_inputs['label'].to(device)
recognizer = recognizer.to(device)
gradcam = GradCAM(recognizer, target_layer_name)
blended_imgs_target_shape, preds_target_shape = _get_target_shapes(
input_shape, num_classes=num_classes, model_type='2D')
blended_imgs, preds = gradcam(demo_inputs)
assert blended_imgs.size() == blended_imgs_target_shape
assert preds.size() == preds_target_shape
blended_imgs, preds = gradcam(demo_inputs, True)
assert blended_imgs.size() == blended_imgs_target_shape
assert preds.size() == preds_target_shape