def predict_one(inputs, dir_save, threshold, threshold2, overlap):
viz = Visualizer('lib/colors.json')
i, target_name, numpy_batch_num = inputs
list_predicted_img, pred_coords = suppression(numpy_batch_num, dbox_params, threshold, threshold2, overlap)
sub_i = pd.DataFrame()
sub_i['token'] = [target_name] * len(pred_coords)
sub_i_coord = pd.DataFrame(pred_coords,
columns=['x', 'y', 'length', 'width', 'rotate', 'z', 'height', 'class', 'confidence'])
sub_i = pd.concat([sub_i, sub_i_coord], axis=1)
sub_i['name'] = [viz.names[c] for c in sub_i['class']]
if i < 50:
img_predicted = viz.draw_predicted_boxes(numpy_batch_num, dbox_params, rate=1.0, img_size=2048)
numpy2pil(img_predicted).save(dir_save / '{}_predited_{}.png'.format(i, target_name[:12]))
img_submit = viz.draw_mask_from_coords(sub_i_coord, img_size=2048)
numpy2pil(img_submit).save(dir_save / '{}_submit_{}_{}.png'.format(i, target_name[:12], len(list_predicted_img)))
return sub_i
viz = Visualizer('colors.json')
# 768 x 768
in_arr1 = np.zeros((2, 3, 768, 768), dtype=np.float32)
in_tensor1 = torch.from_numpy(in_arr1)
out_vars1 = model.forward(in_tensor1.cuda())
[print(out_var.shape) for out_var in out_vars1]
out_var_numpy1 = [tensor.cpu().data.numpy() for tensor in out_vars1]
out_var_numpy_batch1 = [[tensor[b, :, :, :] for tensor in out_var_numpy1] for b in range(2)]
img = viz.draw_predicted_boxes(out_var_numpy_batch1[0], dbox_params, img_size=in_arr1.shape[-1])
numpy2pil(img).save(dir_debug / 'sample_1-0.png')
img = viz.draw_predicted_boxes(out_var_numpy_batch1[1], dbox_params, img_size=in_arr1.shape[-1])
numpy2pil(img).save(dir_debug / 'sample_1-1.png')
# 1024 x 1024
in_arr2 = np.zeros((2, 3, 1024, 1024), dtype=np.float32)
in_tensor2 = torch.from_numpy(in_arr2)
out_vars2 = model.forward(in_tensor2.cuda())
[print(out_var.shape) for out_var in out_vars2]
out_var_numpy2 = [tensor.cpu().data.numpy() for tensor in out_vars2]
out_var_numpy_batch2 = [[tensor[b, :, :, :] for tensor in out_var_numpy2] for b in range(2)]
def validate(model, dataset, list_valid_img, target_classes, epoch):
dir_save = Path('./_valid/ep{}'.format(epoch))
dir_save.parent.mkdir(exist_ok=True)
dir_save.mkdir(exist_ok=True)
model.eval()
dataset = SampleDataset(
dir_img=f'../../input/{dataset}/train',
coord_path=f'../../input/{dataset}/coordinates.csv',
use_augmentation=epoch < 20,
list_imgs=list_valid_img,
crop_type=0,
target_classes=target_classes)
viz = Visualizer('lib/colors.json')
for target_name in list_valid_img[:16]:
img_input, found_coord, original = dataset[target_name]
writer.add_image('original/{}'.format(target_name[:12]),
np.asarray(original),
epoch,
dataformats='HWC')
original.save(dir_save / '{}_original.png'.format(target_name[:12]))
input_tensor = torch.unsqueeze(img_input, 0)
# input_tensor = augment_input(img_input)
with torch.no_grad():
net_out = model.forward(input_tensor.float().cuda())
net_out_numpy = [tensor.cpu().data.numpy() for tensor in net_out]
net_out_numpy_batch = [tensor[0, :, :, :] for tensor in net_out_numpy]
# net_out_numpy_batch = aggregate_output(net_out_numpy)
img_predicted = viz.draw_predicted_boxes(net_out_numpy_batch,
dbox_params,
rate=1.0,
img_size=original.height)
writer.add_image('predicted/{}'.format(target_name[:12]),
img_predicted, epoch)
numpy2pil(img_predicted).save(
dir_save / '{}_predited.png'.format(target_name[:12]))
mask = viz.draw_mask_from_coords(found_coord, img_size=original.height)
writer.add_image('mask/{}'.format(target_name[:12]), mask, epoch)
numpy2pil(mask).save(dir_save / '{}_mask.png'.format(target_name[:12]))
list_predicted_img, list_pred_coords = suppression(
net_out_numpy_batch, dbox_params, 0.5, 0.3, 0.1)
pred_coords_df = pd.DataFrame(list_pred_coords,
columns=[
'x', 'y', 'length', 'width',
'rotate', 'z', 'height', 'class',
'confidence'
])
img_submit = viz.draw_mask_from_coords(pred_coords_df,
img_size=original.height)
writer.add_image('submit/{}'.format(target_name[:12]), img_submit,
epoch)
numpy2pil(img_submit).save(dir_save / '{}_submit_{}.png'.format(
target_name[:12], len(list_predicted_img)))