def evaluate():
opt = get_args()
# Loading Pretarined Model
models = {}
models["encoder"] = monolayout.Encoder(18, opt.height, opt.width, True)
if opt.type == "both":
models["static_decoder"] = monolayout.Decoder(
models["encoder"].resnet_encoder.num_ch_enc)
models["dynamic_decoder"] = monolayout.Decoder(
models["encoder"].resnet_encoder.num_ch_enc)
else:
models["decoder"] = monolayout.Decoder(
models["encoder"].resnet_encoder.num_ch_enc)
for key in models.keys():
models[key].to("cuda")
models = load_model(models, opt.pretrained_path)
# Loading Validation/Testing Dataset
# Data Loaders
dataset_dict = {"3Dobject": monolayout.KITTIObject,
"odometry": monolayout.KITTIOdometry,
"argo": monolayout.Argoverse,
"raw": monolayout.KITTIRAW}
dataset = dataset_dict[opt.split]
fpath = os.path.join(
os.path.dirname(__file__),
"splits",
opt.split,
"{}_files.txt")
test_filenames = readlines(fpath.format("val"))
test_dataset = dataset(opt, test_filenames, is_train=False)
test_loader = DataLoader(
test_dataset,
1,
True,
num_workers=opt.num_workers,
pin_memory=True,
drop_last=True)
iou, mAP = np.array([0., 0.]), np.array([0., 0.])
for batch_idx, inputs in tqdm.tqdm(enumerate(test_loader)):
with torch.no_grad():
outputs = process_batch(opt, models, inputs)
pred = np.squeeze(
torch.argmax(
outputs["topview"].detach(),
1).cpu().numpy())
true = np.squeeze(inputs[opt.type + "_gt"].detach().cpu().numpy())
iou += mean_IU(pred, true)
mAP += mean_precision(pred, true)
iou /= len(test_loader)
mAP /= len(test_loader)
print("Evaluation Results: mIOU: %.4f mAP: %.4f" % (iou[1], mAP[1]))
def validation(self):
iou, mAP = np.array([0., 0.]), np.array([0., 0.])
for batch_idx, inputs in tqdm.tqdm(enumerate(self.val_loader)):
with torch.no_grad():
outputs = self.process_batch(inputs, True)
pred = np.squeeze(
torch.argmax(outputs["topview"].detach(), 1).cpu().numpy())
true = np.squeeze(inputs[self.opt.type +
"_gt"].detach().cpu().numpy())
iou += mean_IU(pred, true)
mAP += mean_precision(pred, true)
iou /= len(self.val_loader)
mAP /= len(self.val_loader)
print("Epoch: %d | Validation: mIOU: %.4f mAP: %.4f" %
(self.epoch, iou[1], mAP[1]))
feed_dict = {img: x_batch, label: y_batch}
loss, pred_logits = sess.run([log_loss, pred],
feed_dict=feed_dict)
pred_map_batch = np.argmax(pred_logits, axis=3)
# import pdb; pdb.set_trace()
for pred_map, y in zip(pred_map_batch, y_batch):
mean_iou, pixel_acc, dice = vis.add_sample(
pred_map, y_batch[0])
vis.compute_scores(suffix=it)
x_batch, y_batch = next(train_generator)
# w_map = define_map(y_batch[0])
# exclude whole-backgroud images
feed_dict = {
img: x_batch,
label: y_batch
# weight_map: w_map
}
_, loss, summary, lr, pred_logits = sess.run(
[train_step, log_loss, summary_merged, learning_rate, pred],
feed_dict=feed_dict)
global_step.assign(it).eval()
train_writer.add_summary(summary, it)
pred_map = np.argmax(pred_logits[0], axis=2)
score, _ = mean_IU(pred_map, y_batch[0])
if it % 20 == 0:
print('[iter %d, epoch %.3f]: lr=%f loss=%f, mean_IU=%f' %
(it, float(it) / opt.iter_epoch, lr, loss, score))
def evaluate():
opt = get_args()
# Loading Pretarined Model
models = {}
models["encoder"] = racklay.Encoder(18, opt.height, opt.width, True)
if opt.type == "both":
models["top_decoder"] = racklay.Decoder(
models["encoder"].resnet_encoder.num_ch_enc, 3 * opt.num_racks,
opt.occ_map_size)
models["front_decoder"] = racklay.Decoder(
models["encoder"].resnet_encoder.num_ch_enc, 3 * opt.num_racks,
opt.occ_map_size)
elif opt.type == "topview":
models["top_decoder"] = racklay.Decoder(
models["encoder"].resnet_encoder.num_ch_enc)
elif opt.type == "frontview":
models["front_decoder"] = racklay.Decoder(
models["encoder"].resnet_encoder.num_ch_enc)
for key in models.keys():
models[key].to("cuda")
models = load_model(models, opt.pretrained_path)
# Loading Validation/Testing Dataset
# Data Loaders
dataset_dict = {
"warehouse": Loader,
"3Dobject": racklay.KITTIObject,
"odometry": racklay.KITTIOdometry,
"argo": racklay.Argoverse,
"raw": racklay.KITTIRAW
}
dataset = dataset_dict[opt.split]
fpath = os.path.join(os.path.dirname(__file__), "splits", opt.split,
"{}_files.txt")
test_filenames = readlines(fpath.format("val"))
test_dataset = dataset(opt, test_filenames, is_train=False)
test_loader = DataLoader(test_dataset,
1,
True,
num_workers=opt.num_workers,
pin_memory=True,
drop_last=True)
iou_box_top, mAP_box_top = np.array([0., 0.]), np.array([0., 0.])
iou_rack_top, mAP_rack_top = np.array([0., 0.]), np.array([0., 0.])
iou_box_front, mAP_box_front = np.array([0., 0.]), np.array([0., 0.])
iou_rack_front, mAP_rack_front = np.array([0., 0.]), np.array([0., 0.])
for batch_idx, inputs in tqdm.tqdm(enumerate(test_loader)):
with torch.no_grad():
outputs = process_batch(opt, models, inputs)
#top view
if (opt.type == "both" or opt.type == "topview"):
for i in range(opt.num_racks): # For the Rack Case
input_temp = inputs["topview"][:,
i, :, :].detach().cpu().numpy()
input_onlyrack = np.zeros_like(input_temp)
input_onlyrack[input_temp == 1] = 1
input_temp = np.squeeze(input_onlyrack)
input_temp = cv2.resize(input_temp,
dsize=(opt.occ_map_size,
opt.occ_map_size),
interpolation=cv2.INTER_NEAREST)
pred = np.squeeze(
torch.argmax(
outputs["top"][:, 3 * i:3 * i + 3, :, :].detach(),
1).cpu().numpy())
pred_temp = np.zeros_like(pred)
pred_temp[pred == 1] = 1
true = np.squeeze(input_temp)
iou_rack_top += mean_IU(pred_temp, true)
mAP_rack_top += mean_precision(pred_temp, true)
for i in range(opt.num_racks):
input_temp = inputs["topview"][:,
i, :, :].detach().cpu().numpy()
input_onlybox = np.zeros_like(input_temp)
input_onlybox[input_temp == 2] = 1
input_temp = np.squeeze(input_onlybox)
input_temp = cv2.resize(input_temp,
dsize=(opt.occ_map_size,
opt.occ_map_size),
interpolation=cv2.INTER_NEAREST)
pred = np.squeeze(
torch.argmax(
outputs["top"][:, 3 * i:3 * i + 3, :, :].detach(),
1).cpu().numpy())
pred_temp = np.zeros_like(pred)
pred_temp[pred == 2] = 1
true = np.squeeze(input_temp)
iou_box_top += mean_IU(pred_temp, true)
mAP_box_top += mean_precision(pred_temp, true)
#front view
if (opt.type == "both" or opt.type == "frontview"):
for i in range(opt.num_racks): # For the Rack Case
input_temp = inputs["frontview"][:,
i, :, :].detach().cpu().numpy(
)
input_onlyrack = np.zeros_like(input_temp)
input_onlyrack[input_temp == 1] = 1
input_temp = np.squeeze(input_onlyrack)
input_temp = cv2.resize(input_temp,
dsize=(opt.occ_map_size,
opt.occ_map_size),
interpolation=cv2.INTER_NEAREST)
pred = np.squeeze(
torch.argmax(
outputs["front"][:, 3 * i:3 * i + 3, :, :].detach(),
1).cpu().numpy())
pred_temp = np.zeros_like(pred)
pred_temp[pred == 1] = 1
true = np.squeeze(input_temp)
iou_rack_front += mean_IU(pred_temp, true)
mAP_rack_front += mean_precision(pred_temp, true)
for i in range(opt.num_racks):
input_temp = inputs["frontview"][:,
i, :, :].detach().cpu().numpy(
)
input_onlybox = np.zeros_like(input_temp)
input_onlybox[input_temp == 2] = 1
input_temp = np.squeeze(input_onlybox)
input_temp = cv2.resize(input_temp,
dsize=(opt.occ_map_size,
opt.occ_map_size),
interpolation=cv2.INTER_NEAREST)
pred = np.squeeze(
torch.argmax(
outputs["front"][:, 3 * i:3 * i + 3, :, :].detach(),
1).cpu().numpy())
pred_temp = np.zeros_like(pred)
pred_temp[pred == 2] = 1
true = np.squeeze(input_temp)
iou_box_front += mean_IU(pred_temp, true)
mAP_box_front += mean_precision(pred_temp, true)
if (opt.type == "both" or opt.type == "topview"):
iou_rack_top /= (len(test_loader) * opt.num_racks)
mAP_rack_top /= (len(test_loader) * opt.num_racks)
iou_box_top /= (len(test_loader) * opt.num_racks)
mAP_box_top /= (len(test_loader) * opt.num_racks)
print("Evaluation Results for Rack Top: mIOU: %.4f mAP: %.4f" %
(iou_rack_top[1], mAP_rack_top[1]))
print("Evaluation Results for Box Top: mIOU: %.4f mAP: %.4f" %
(iou_box_top[1], mAP_box_top[1]))
if (opt.type == "both" or opt.type == "frontview"):
iou_rack_front /= (len(test_loader) * opt.num_racks)
mAP_rack_front /= (len(test_loader) * opt.num_racks)
iou_box_front /= (len(test_loader) * opt.num_racks)
mAP_box_front /= (len(test_loader) * opt.num_racks)
print("Evaluation Results for Rack Front: mIOU: %.4f mAP: %.4f" %
(iou_rack_front[1], mAP_rack_front[1]))
print("Evaluation Results for Box Front: mIOU: %.4f mAP: %.4f" %
(iou_box_front[1], mAP_box_front[1]))