def test(self, task_name=''):
if self.pred_model is not None:
self.pred_model.eval()
pos, num_sample = 0, 0
# sample actions
acts = self.enumerate_actions()
for batch_idx, (data, boxes, gt_acts, labels, _, _, _, _) in enumerate(self.val_loader):
batch_size = data.shape[0]
num_objs = gt_acts.shape[1]
pos_feat = xyxy_to_posf(boxes, data.shape)
rois = xyxy_to_rois(boxes, batch_size, data.shape[1], self.num_gpus)
gt_rois = boxes.cpu().numpy().copy()
pred_acts = np.zeros((batch_size, num_objs, 3))
conf_acts = -np.inf * np.ones((batch_size,))
# if self.random_policy:
# for i in range(batch_size):
# obj_id = np.random.randint(num_objs)
# pred_acts[i, obj_id] = acts[np.random.randint(len(acts))]
for idx, (act, obj_id) in enumerate(itertools.product(acts, range(num_objs))):
tprint(f'current batch: {idx} / {acts.shape[0] * num_objs}' + ' ' * 10)
act_array = torch.zeros((batch_size, num_objs, 3), dtype=torch.float32)
act_array[:, obj_id, :] = torch.from_numpy(act)
traj_array = self.generate_trajs(data, rois, pos_feat, act_array, boxes)
conf = self.get_act_conf(traj_array, gt_rois, obj_id)
pred_acts[conf > conf_acts] = act_array[conf > conf_acts]
conf_acts[conf > conf_acts] = conf[conf > conf_acts]
# for i in range(C.SOLVER.BATCH_SIZE):
# plot_image_idx = C.SOLVER.BATCH_SIZE * batch_idx + i
# video_idx, img_idx = self.val_loader.dataset.video_info[plot_image_idx]
# video_name = self.val_loader.dataset.video_list[video_idx]
# search_suffix = self.val_loader.dataset.search_suffix
# image_list = sorted(glob(f'{video_name}/{search_suffix}'))
# im_name = image_list[img_idx]
# video_id, image_id = im_name.split('.')[0].split('/')[-2:]
# output_name = f'{video_id}_{image_id}_{idx}'
# im_data = get_im_data(im_name, gt_rois[[i], 0:1], C.DATA_ROOT, False)
# plt.axis('off')
# plt.imshow(im_data[..., ::-1])
# _plot_bbox_traj(traj_array[i], size=160, alpha=1.0)
# x = gt_rois[i, 0, obj_id, 0] + self.ball_radius
# y = gt_rois[i, 0, obj_id, 1] + self.ball_radius
# dy, dx = act[0] * act[2], act[1] * act[2]
# plt.arrow(x, y, dx, dy, color=(0.99, 0.99, 0.99), linewidth=5)
# os.makedirs(f'{self.output_dir}/plan', exist_ok=True)
# kwargs = {'format': 'svg', 'bbox_inches': 'tight', 'pad_inches': 0}
# plt.savefig(f'{self.output_dir}/plan/pred_{output_name}.svg', **kwargs)
# plt.close()
sim_rst, debug_gt_traj_array = self.simulate_action(gt_rois, pred_acts)
pos += sim_rst.sum()
num_sample += sim_rst.shape[0]
pprint(f'{task_name} {batch_idx}/{len(self.val_loader)}: {pos / num_sample:.4f}' + ' ' * 10)
pprint(f'{task_name}: {pos / num_sample:.4f}' + ' ' * 10)
def __init__(self, data_root, split, image_ext='.jpg'):
super().__init__(data_root, split, image_ext)
self.video_list = sorted(glob(f'{self.data_root}/{self.split}/*/'))
self.anno_list = [v[:-1] + '_boxes.pkl' for v in self.video_list]
self.video_info = np.zeros((0, 2), dtype=np.int32)
for idx, video_name in enumerate(self.video_list):
tprint(f'loading progress: {idx}/{len(self.video_list)}')
num_im = len(glob(f'{video_name}/*{image_ext}'))
if self.split == 'test':
num_sw = min(1, num_im - self.seq_size + 1)
else:
num_sw = num_im - self.seq_size + 1 # number of sliding windows
if num_sw <= 0:
continue
video_info_t = np.zeros((num_sw, 2), dtype=np.int32)
video_info_t[:, 0] = idx # video index
video_info_t[:, 1] = np.arange(num_sw) # sliding window index
self.video_info = np.vstack((self.video_info, video_info_t))
def _eval_and_score_actions(cache, model, simulator, num_actions, batch_size,
observations):
actions = cache.action_array[:num_actions]
indices = np.random.RandomState(1).permutation(
len(observations))[:AUCCESS_EVAL_TASKS]
evaluator = phyre.Evaluator(
[simulator.task_ids[index] for index in indices])
for i, task_index in enumerate(indices):
tprint(f'{i}/{len(indices)} {task_index}')
scores = eval_actions(model, actions, batch_size,
observations[task_index]).tolist()
_, sorted_actions = zip(
*sorted(zip(scores, actions), key=lambda x: (-x[0], tuple(x[1]))))
for action in sorted_actions:
if (evaluator.get_attempts_for_task(i) >= phyre.MAX_TEST_ATTEMPTS):
break
status = simulator.simulate_action(task_index,
action,
need_images=False).status
evaluator.maybe_log_attempt(i, status)
return evaluator.get_aucess()
def __init__(self, data_root, split, image_ext='.jpg'):
super().__init__(data_root, split, image_ext)
if C.INPUT.PRELOAD_TO_MEMORY:
print('loading data from pickle...')
with open(f'{self.data_root}/{self.split}.pkl', 'rb') as f:
data = pickle.load(f)
self.total_img = np.transpose(data['X'], (0, 1, 4, 2, 3))
self.total_box = np.zeros((data['y'].shape[:3] + (5, )))
assert len(self.anno_list) > 0
for anno_idx, anno_name in enumerate(self.anno_list):
tprint(f'loading progress: {anno_idx}/{len(self.anno_list)}')
with open(anno_name, 'rb') as f:
boxes = pickle.load(f)
self.total_box[anno_idx] = boxes
else:
self.video_list = sorted(glob(f'{self.data_root}/{self.split}/*/'))
self.anno_list = [v[:-1] + '.pkl' for v in self.video_list]
self.video_info = np.zeros((0, 2), dtype=np.int32)
for idx, video_name in enumerate(
self.video_list if not C.INPUT.PRELOAD_TO_MEMORY else self.
total_box):
tprint(f'loading progress: {idx}/{len(self.video_list)}')
if C.INPUT.PRELOAD_TO_MEMORY:
num_sw = self.total_box[idx].shape[0] - self.seq_size + 1
else:
num_im = len(glob(f'{video_name}/*{image_ext}'))
num_sw = num_im - self.seq_size + 1 # number of sliding windows
if num_sw <= 0:
continue
video_info_t = np.zeros((num_sw, 2), dtype=np.int32)
video_info_t[:, 0] = idx # video index
video_info_t[:, 1] = np.arange(num_sw) # sliding window index
self.video_info = np.vstack((self.video_info, video_info_t))
def __init__(self, data_root, split, protocal='within', image_ext='.jpg'):
super().__init__(data_root, split, image_ext)
env_list = open(f'{data_root}/{protocal}_{split}_fold_0.txt',
'r').read().split('\n')
self.video_list = sum([
sorted(glob(f'{data_root}/images/{env.replace(":", "/")}/*.npy'))
for env in env_list
], [])
self.anno_list = [(v[:-4] + '_boxes.hkl').replace('images', 'labels')
for v in self.video_list]
# just for plot images
if plot:
self.video_list = [
k for k in self.video_list
if int(k.split('/')[-1].split('.')[0]) < 40
]
self.anno_list = [
k for k in self.anno_list
if int(k.split('/')[-1].split('_')[0]) < 40
]
assert len(self.video_list) == len(self.anno_list)
self.video_list = self.video_list[::80]
self.anno_list = self.anno_list[::80]
# video_info_name = f'for_plot.npy'
video_info_name = f'{data_root}/{protocal}_{split}_{self.input_size}_{self.pred_size}_fold_0_info.npy'
if os.path.exists(video_info_name):
print(f'loading info from: {video_info_name}')
self.video_info = np.load(video_info_name)
else:
self.video_info = np.zeros((0, 2), dtype=np.int32)
for idx, video_name in enumerate(self.video_list):
tprint(f'loading progress: {idx}/{len(self.video_list)}')
num_im = hickle.load(
video_name.replace('images', 'labels').replace(
'.npy', '_boxes.hkl')).shape[0]
num_sw = num_im - self.seq_size + 1 # number of sliding windows
if plot:
num_sw = 1
if self.input_size == 1:
num_sw = min(1, num_im - self.seq_size + 1)
if num_sw <= 0:
continue
video_info_t = np.zeros((num_sw, 2), dtype=np.int32)
video_info_t[:, 0] = idx # video index
video_info_t[:, 1] = np.arange(num_sw) # sliding window index
self.video_info = np.vstack((self.video_info, video_info_t))
np.save(video_info_name, self.video_info)
for module in C.SINGULAR_MODULES:
self.module_dict[module] = np.load(
f'{data_root}/{module}/thresh_{C.MASK_THRESH}/{protocal}_{split}_{self.input_size}_{self.pred_size}_fold_0_info.npy'
)
for module in C.DUAL_MODULES:
self.module_dict[module] = np.load(
f'{data_root}/{module}/{C.DUAL_DATATYPE}/thresh_{C.MASK_THRESH}/{protocal}_{split}_{self.input_size}_{self.pred_size}_fold_0_info.npy'
)
# load GT indicators and object info tags
self.objinfo = np.load(
f'{data_root}/thresh/{C.MASK_THRESH}/{protocal}_{split}_{self.input_size}_{self.pred_size}_fold_0_objinfo.npy'
)
self.gtindicatorinfo = np.load(
f'{data_root}/thresh/{C.MASK_THRESH}/{protocal}_{split}_{self.input_size}_{self.pred_size}_fold_0_gtindicatorinfo.npy'
)
def train_epoch(self):
for batch_idx, (data, data_pred, data_t, env_name, rois, gt_boxes,
gt_masks, valid, module_valid, g_idx, seq_l, objinfo,
gtindicator) in enumerate(self.train_loader):
self._adjust_learning_rate()
if C.RIN.ROI_MASKING or C.RIN.ROI_CROPPING:
# data should be (b x t x o x c x h x w)
data = data.permute((0, 2, 1, 3, 4, 5)) # (b, o, t, c, h, w)
data = data.reshape((data.shape[0] * data.shape[1], ) +
data.shape[2:]) # (b*o, t, c, h, w)
data, data_t = data.to(self.device), data_t.to(self.device)
pos_feat = xyxy_to_posf(rois, data.shape)
rois = xyxy_to_rois(rois,
batch=data.shape[0],
time_step=data.shape[1],
num_devices=self.num_gpus)
self.optim.zero_grad()
outputs = self.model(data,
rois,
pos_feat,
valid,
num_rollouts=self.ptrain_size,
g_idx=g_idx,
x_t=data_t,
phase='train')
labels = {
'boxes': gt_boxes.to(self.device),
'masks': gt_masks.to(self.device),
'valid': valid.to(self.device),
'module_valid': module_valid.to(self.device),
'seq_l': seq_l.to(self.device),
'gt_indicators': gtindicator.to(self.device),
}
loss = self.loss(outputs, labels, 'train')
loss.backward()
self.optim.step()
# this is an approximation for printing; the dataset size may not divide the batch size
self.iterations += self.batch_size
print_msg = ""
print_msg += f"{self.epochs:03}/{self.iterations // 1000:04}k"
print_msg += f" | "
mean_loss = np.mean(
np.array(self.box_p_step_losses[:self.ptrain_size]) /
self.loss_cnt) * 1e3
print_msg += f"{mean_loss:.3f} | "
print_msg += f" | ".join([
"{:.3f}".format(self.losses[name] * 1e3 / self.loss_cnt)
for name in self.loss_name
])
print_msg += " || {:.4f}".format(self.loss_ind)
speed = self.loss_cnt / (timer() - self.time)
eta = (self.max_iters - self.iterations) / speed / 3600
print_msg += f" | speed: {speed:.1f} | eta: {eta:.2f} h"
print_msg += (
" " * (os.get_terminal_size().columns - len(print_msg) - 10))
tprint(print_msg)
if self.iterations % self.val_interval == 0:
self.snapshot()
self.val()
self._init_loss()
self.model.train()
if self.iterations >= self.max_iters:
print('\r', end='')
print(f'{self.best_mean:.3f}')
break
def val(self):
self.model.eval()
self._init_loss()
if C.RIN.VAE:
losses = dict.fromkeys(self.loss_name, 0.0)
box_p_step_losses = [0.0 for _ in range(self.ptest_size)]
masks_step_losses = [0.0 for _ in range(self.ptest_size)]
for batch_idx, (data, data_pred, data_t, env_name, rois, gt_boxes,
gt_masks, valid, module_valid, g_idx, seq_l, objinfo,
gtindicator) in enumerate(self.val_loader):
tprint(f'eval: {batch_idx}/{len(self.val_loader)}')
with torch.no_grad():
if C.RIN.ROI_MASKING or C.RIN.ROI_CROPPING:
data = data.permute((0, 2, 1, 3, 4, 5))
data = data.reshape((data.shape[0] * data.shape[1], ) +
data.shape[2:])
data = data.to(self.device)
pos_feat = xyxy_to_posf(rois, data.shape)
rois = xyxy_to_rois(rois,
batch=data.shape[0],
time_step=data.shape[1],
num_devices=self.num_gpus)
labels = {
'boxes': gt_boxes.to(self.device),
'masks': gt_masks.to(self.device),
'valid': valid.to(self.device),
'module_valid': module_valid.to(self.device),
'seq_l': seq_l.to(self.device),
'gt_indicators': gtindicator.to(self.device),
}
outputs = self.model(data,
rois,
pos_feat,
valid,
num_rollouts=self.ptest_size,
g_idx=g_idx,
phase='test')
self.loss(outputs, labels, 'test')
# VAE multiple runs
if C.RIN.VAE:
vae_best_mean = np.mean(
np.array(self.box_p_step_losses[:self.ptest_size]) /
self.loss_cnt) * 1e3
losses_t = self.losses.copy()
box_p_step_losses_t = self.box_p_step_losses.copy()
masks_step_losses_t = self.masks_step_losses.copy()
for i in range(9):
outputs = self.model(data,
rois,
pos_feat,
valid,
num_rollouts=self.ptest_size,
g_idx=g_idx,
phase='test')
self.loss(outputs, labels, 'test')
mean_loss = np.mean(
np.array(self.box_p_step_losses[:self.ptest_size])
/ self.loss_cnt) * 1e3
if mean_loss < vae_best_mean:
losses_t = self.losses.copy()
box_p_step_losses_t = self.box_p_step_losses.copy()
masks_step_losses_t = self.masks_step_losses.copy()
vae_best_mean = mean_loss
self._init_loss()
for k, v in losses.items():
losses[k] += losses_t[k]
for i in range(len(box_p_step_losses)):
box_p_step_losses[i] += box_p_step_losses_t[i]
masks_step_losses[i] += masks_step_losses_t[i]
if C.RIN.VAE:
self.losses = losses.copy()
self.box_p_step_losses = box_p_step_losses.copy()
self.loss_cnt = len(self.val_loader)
print('\r', end='')
print_msg = ""
print_msg += f"{self.epochs:03}/{self.iterations // 1000:04}k"
print_msg += f" | "
mean_loss = np.mean(
np.array(self.box_p_step_losses[:self.ptest_size]) /
self.loss_cnt) * 1e3
print_msg += f"{mean_loss:.3f} | "
if mean_loss < self.best_mean:
self.snapshot('ckpt_best.path.tar')
self.best_mean = mean_loss
print_msg += f" | ".join([
"{:.3f}".format(self.losses[name] * 1e3 / self.loss_cnt)
for name in self.loss_name
])
print_msg += " || {:.4f}".format(self.loss_ind)
print_msg += (" " *
(os.get_terminal_size().columns - len(print_msg) - 10))
self.logger.info(print_msg)
def test(self):
self.model.eval()
if C.RIN.VAE:
losses = dict.fromkeys(self.loss_name, 0.0)
box_p_step_losses = [0.0 for _ in range(self.ptest_size)]
masks_step_losses = [0.0 for _ in range(self.ptest_size)]
for batch_idx, (data, data_pred, data_t, env_name, rois, gt_boxes,
gt_masks, valid, module_valid, g_idx,
_) in enumerate(self.val_loader):
with torch.no_grad():
# decide module_valid here for evaluation
mid = 0 # ball-only
module_valid = module_valid[:, mid, :, :]
if C.RIN.ROI_MASKING or C.RIN.ROI_CROPPING:
# data should be (b x t x o x c x h x w)
data = data.permute(
(0, 2, 1, 3, 4, 5)) # (b, o, t, c, h, w)
data = data.reshape((data.shape[0] * data.shape[1], ) +
data.shape[2:]) # (b*o, t, c, h, w)
data = data.to(self.device)
pos_feat = xyxy_to_posf(rois, data.shape)
rois = xyxy_to_rois(rois,
batch=data.shape[0],
time_step=data.shape[1],
num_devices=self.num_gpus)
labels = {
'boxes': gt_boxes.to(self.device),
'masks': gt_masks.to(self.device),
'valid': valid.to(self.device),
'module_valid': module_valid.to(self.device),
}
outputs = self.model(data,
rois,
pos_feat,
valid,
num_rollouts=self.ptest_size,
g_idx=g_idx,
phase='test')
# # *********************************************************************************
# # VISUALIZATION - generate input image and GT outputs + model outputs
# input_data = data.cpu().detach().numpy() # (128, 1, 3, 128, 128)
# gt_data = data_pred.cpu().detach().numpy() # (128, 10, 3, 128, 128)
# data_t = data_t.cpu().detach().numpy() # (128, 1, 128, 128)
# validity = valid.cpu().detach().numpy() # (128, 6)
# outputs_boxes = outputs['boxes'].cpu().detach().numpy() # (128, 10, 6, 4)
# outputs_masks = outputs['masks'].cpu().detach().numpy() # (128, 10, 6, 21, 21)
# np.save('save/'+str(batch_idx)+'input_data.npy', input_data)
# np.save('save/'+str(batch_idx)+'gt_data.npy', gt_data)
# np.save('save/'+str(batch_idx)+'data_t.npy', data_t)
# np.save('save/'+str(batch_idx)+'validity.npy', validity)
# np.save('save/'+str(batch_idx)+'outputs_boxes.npy', outputs_boxes)
# np.save('save/'+str(batch_idx)+'outputs_masks.npy', outputs_masks)
# # self.visualize_results(input_data, gt_data, outputs, data_t, validity, env_name)
# # *********************************************************************************
self.loss(outputs, labels, 'test')
# VAE multiple runs
if C.RIN.VAE:
vae_best_mean = np.mean(
np.array(self.box_p_step_losses[:self.ptest_size]) /
self.loss_cnt) * 1e3
losses_t = self.losses.copy()
box_p_step_losses_t = self.box_p_step_losses.copy()
masks_step_losses_t = self.masks_step_losses.copy()
for i in range(99):
outputs = self.model(data,
rois,
None,
num_rollouts=self.ptest_size,
g_idx=g_idx,
phase='test')
self.loss(outputs, labels, 'test')
mean_loss = np.mean(
np.array(self.box_p_step_losses[:self.ptest_size])
/ self.loss_cnt) * 1e3
if mean_loss < vae_best_mean:
losses_t = self.losses.copy()
box_p_step_losses_t = self.box_p_step_losses.copy()
masks_step_losses_t = self.masks_step_losses.copy()
vae_best_mean = mean_loss
self._init_loss()
for k, v in losses.items():
losses[k] += losses_t[k]
for i in range(len(box_p_step_losses)):
box_p_step_losses[i] += box_p_step_losses_t[i]
masks_step_losses[i] += masks_step_losses_t[i]
tprint(f'eval: {batch_idx}/{len(self.val_loader)}:' + ' ' * 20)
if self.plot_image > 0:
outputs = {
'boxes':
outputs['boxes'].cpu().numpy(),
'masks':
outputs['masks'].cpu().numpy()
if C.RIN.MASK_LOSS_WEIGHT else None,
}
outputs['boxes'][..., 0::2] *= self.input_width
outputs['boxes'][..., 1::2] *= self.input_height
outputs['boxes'] = xywh2xyxy(outputs['boxes'].reshape(
-1, 4)).reshape((data.shape[0], -1, C.RIN.NUM_OBJS, 4))
labels = {
'boxes': labels['boxes'].cpu().numpy(),
'masks': labels['masks'].cpu().numpy(),
}
labels['boxes'][..., 0::2] *= self.input_width
labels['boxes'][..., 1::2] *= self.input_height
labels['boxes'] = xywh2xyxy(labels['boxes'].reshape(
-1, 4)).reshape((data.shape[0], -1, C.RIN.NUM_OBJS, 4))
for i in range(rois.shape[0]):
batch_size = C.SOLVER.BATCH_SIZE if not C.RIN.VAE else 1
plot_image_idx = batch_size * batch_idx + i
if plot_image_idx < self.plot_image:
tprint(f'plotting: {plot_image_idx}' + ' ' * 20)
video_idx, img_idx = self.val_loader.dataset.video_info[
plot_image_idx]
video_name = self.val_loader.dataset.video_list[
video_idx]
v = valid[i].numpy().astype(np.bool)
pred_boxes_i = outputs['boxes'][i][:, v]
gt_boxes_i = labels['boxes'][i][:, v]
if 'PHYRE' in C.DATA_ROOT:
# [::-1] is to make it consistency with others where opencv is used
im_data = phyre.observations_to_float_rgb(
np.load(video_name).astype(
np.uint8))[..., ::-1]
a, b, c = video_name.split('/')[5:8]
output_name = f'{a}_{b}_{c.replace(".npy", "")}'
bg_image = np.load(video_name).astype(np.uint8)
for fg_id in [1, 2, 3, 5]:
bg_image[bg_image == fg_id] = 0
bg_image = phyre.observations_to_float_rgb(
bg_image)
# if f'{a}_{b}' not in [
# '00014_123', '00014_528', '00015_257', '00015_337', '00019_273', '00019_296'
# ]:
# continue
# if f'{a}_{b}' not in [
# '00000_069', '00001_000', '00002_185', '00003_064', '00004_823',
# '00005_111', '00006_033', '00007_090', '00008_177', '00009_930',
# '00010_508', '00011_841', '00012_071', '00013_074', '00014_214',
# '00015_016', '00016_844', '00017_129', '00018_192', '00019_244',
# '00020_010', '00021_115', '00022_537', '00023_470', '00024_048'
# ]:
# continue
else:
bg_image = None
image_list = sorted(
glob(
f'{video_name}/*{self.val_loader.dataset.image_ext}'
))
im_name = image_list[img_idx]
output_name = '_'.join(
im_name.split('.')[0].split('/')[-2:])
# deal with image data here
# plot rollout function only take care of the usage of plt
# if output_name not in ['009_015', '009_031', '009_063', '039_038', '049_011', '059_033']:
# continue
# if output_name not in ['00002_00037', '00008_00047', '00011_00048', '00013_00036',
# '00014_00033', '00020_00054', '00021_00013', '00024_00011']:
# continue
if output_name not in [
'0016_000', '0045_000', '0120_000',
'0163_000'
]:
continue
gt_boxes_i = labels['boxes'][i][:, v]
im_data = get_im_data(im_name, gt_boxes_i[None,
0:1],
C.DATA_ROOT,
self.high_resolution_plot)
if self.high_resolution_plot:
scale_w = im_data.shape[1] / self.input_width
scale_h = im_data.shape[0] / self.input_height
pred_boxes_i[..., [0, 2]] *= scale_w
pred_boxes_i[..., [1, 3]] *= scale_h
gt_boxes_i[..., [0, 2]] *= scale_w
gt_boxes_i[..., [1, 3]] *= scale_h
pred_masks_i = None
if C.RIN.MASK_LOSS_WEIGHT:
pred_masks_i = outputs['masks'][i][:, v]
plot_rollouts(im_data,
pred_boxes_i,
gt_boxes_i,
pred_masks_i,
labels['masks'][i][:, v],
output_dir=self.output_dir,
output_name=output_name,
bg_image=bg_image)
if C.RIN.VAE:
self.losses = losses.copy()
self.box_p_step_losses = box_p_step_losses.copy()
self.loss_cnt = len(self.val_loader)
print('\r', end='')
print_msg = ""
mean_loss = np.mean(
np.array(self.box_p_step_losses[:self.ptest_size]) /
self.loss_cnt) * 1e3
print_msg += f"{mean_loss:.3f} | "
print_msg += f" | ".join([
"{:.3f}".format(self.losses[name] * 1e3 / self.loss_cnt)
for name in self.loss_name
])
pprint(print_msg)