error in inference

[Keysmis]

Hi, I'm interested in your project and modified your test code to inference an image, plot 3D, but, maybe something wrong of my code, the results is wrong, can you cost a little time to have a look? very appreciate! `config = copy.deepcopy(s_config) config.network = get_default_network_config() # defined in blocks config.loss = get_default_loss_config()

config = update_config_from_file(config, s_config_file, check_necessity=True)
config = update_config_from_args(config, s_args)  # config in argument is superior to config in file
net = get_pose_net(config.network, 18).cuda()
ckpt = torch.load(s_args.model)  # or other path/to/model
ckpt_old_key = [key for key in ckpt['network'].keys()]
ckpt_new_key = [key.split('.',1)[-1] for key in ckpt['network'].keys()]
print ('ckpt_old_key:',ckpt_old_key)
print('ckpt_new_key:', ckpt_new_key)
ckpt_new = {}
for i in range(len(ckpt_old_key)):

    ckpt_new[ckpt_new_key[i]] = ckpt['network'][ckpt_old_key[i]]
net.load_state_dict(ckpt_new)
print('in valid')
net.eval()

batch_data = cv2.imread('/workspace/w1/wpc/3d_pose_estimatison/integral-human-pose/data/123.png')
batch_data = cv2.cvtColor(batch_data,cv2.COLOR_BGR2RGB)
batch_data = cv2.resize(batch_data,(288,384))
plt.figure(1)
plt.imshow(batch_data)
plt.show()

batch_data = np.array(batch_data).transpose([2,0,1])
batch_data = torch.from_numpy(np.reshape(batch_data,[1,3,batch_data.shape[1],batch_data.shape[2]])).float()
preds_in_patch_with_score = []
batch_data = batch_data.cuda()
preds = net(batch_data)

batch_data_flip = flip(batch_data, dims=3)
preds_flip = net(batch_data_flip)
patch_width = config.train.patch_width
patch_height = config.train.patch_height
flip_pair = np.array([[1, 4], [2, 5], [3, 6], [14, 11], [15, 12], [16, 13]], dtype=np.int)
pipws = get_joint_location_result(config.loss,config.train.patch_width, config.train.patch_height, preds)
pipws_flip = get_joint_location_result(config.loss,config.train.patch_width, config.train.patch_height, preds_flip)
pipws_flip[:, :, 0] = patch_width - pipws_flip[:, :, 0] - 1
for pair in flip_pair:
    tmp = pipws_flip[:, pair[0], :].copy()
    pipws_flip[:, pair[0], :] = pipws_flip[:, pair[1], :].copy()
    pipws_flip[:, pair[1], :] = tmp.copy()
preds_in_patch_with_score.append((pipws + pipws_flip) * 0.5)
_p = np.asarray(preds_in_patch_with_score)
_p = _p.reshape((_p.shape[0] * _p.shape[1], _p.shape[2], _p.shape[3]))
target_bone_length = 4502.881  # train+val
preds_in_camera_space = []
parent_ids = np.array([0, 0, 1, 2, 0, 4, 5, 0, 17, 17, 8, 17, 11, 12, 17, 14, 15, 0], dtype=np.int)
preds_in_camera_space.append(
    rescale_pose_from_patch_to_camera(_p[0],
                                      target_bone_length,
                                      parent_ids))
preds_in_camera_space = np.asarray(preds_in_camera_space)[:, :, 0:3]
fig = plt.figure(2)
ax = fig.add_subplot((111),projection='3d')
plot_3d_skeleton(ax,preds_in_camera_space[0],parent_ids,flip_pair,title='123',patch_width=patch_width,patch_height=patch_height)

`

[lck1201]

@Keysmis Hi, we have provided tools for visualization in common/utility/visualization.py. Besides, your code structure is too complex. I recommand generating prediction first, save them, and load them in second stage to plot.

[lck1201]

Close issue, re-open if you have further questions