print('img_np.HAPE', img_np.shape)
img = torch.from_numpy(img_np).float()
# Get predict heatmap
heatmaps_pred_eg = net_hg_torch(img)
heatmaps_pred_eg = heatmaps_pred_eg
heatmaps_pred_eg = heatmaps_pred_eg[-1].double() #(1,16,64,64)
# Reshape pred heatmaps
heatmaps_pred_eg_np = heatmaps_pred_eg.detach().numpy()
heatmaps_pred_eg_np = np.squeeze(heatmaps_pred_eg_np, axis=0)
heatmaps_pred_eg_np = np.swapaxes(np.swapaxes(heatmaps_pred_eg_np, 0, 2), 0,
1) #(64,64,16)
# Show heatmaps
imgutils.show_heatmaps(img_np_copy, heatmaps_pred_eg_np)
# Stack points
coord_joints = modelutils.heatmaps_to_coords(heatmaps_pred_eg_np,
resolu_out=[256, 256],
prob_threshold=0.1)
imgutils.show_stack_joints(img_np_copy, coord_joints, draw_lines=True)
#========================================================== TEST BY DATASET ==========================================================
# num_subset = 1 # 1-22
# num_instance = 277 # 0-999 256
#
# name_file_img = 'datasets/np__train_imgs_' + str(num_subset) + '.npy'
# name_file_heatmap = 'datasets/np__train_heatmaps_' + str(num_subset) + '.npy'
# name_file_pt = 'datasets/np__train_pts_' + str(num_subset) + '.npy'
#
loss = criteon(heatmaps_pred, heatmaps, target_weight)
print('LOSS:', loss)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Show pred heatmaps
heatmaps_pred_np = heatmaps_pred_copy.detach().numpy()
# heatmaps_pred_np = np.squeeze(heatmaps_pred_np, axis=0)
heatmaps_pred_np = np.swapaxes(np.swapaxes(heatmaps_pred_np, 0, 2), 0,
1) # (64,64,16)
img_np = img_copy.detach().numpy()
# img_np = np.squeeze(img_np, axis=0)
img_np = np.swapaxes(np.swapaxes(img_np, 0, 2), 0, 1) # (64,64,16)
imgutils.show_heatmaps(img_np, heatmaps_pred_np)
plt.pause(1)
plt.ioff()
# if step > 5:
# break
# ================================== Load the model ==================================
# net_hg_torch = hg_torch(num_stacks=1, num_blocks=1, num_classes=16)
# path_model_torch_load = 'models/modelparams_hg_torch_' + str(num_subset) + '.pkl'
# net_hg_torch.load_state_dict(torch.load(path_model_torch_load))
# net_hg_torch.eval()
# print('===============MODEL LOADED===============')
# ================================== Save the model ==================================
path_model_torch = 'models/modelparams_hg_torch_' + str(num_subset) + '.pkl'
# Show last layer output
heatmaps_pred_copy = heatmaps_pred_final[1]
heatmaps_copy = heatmaps[1]
img_copy = img[1]
# Show pred heatmaps
heatmaps_pred_np = heatmaps_pred_copy.detach().cpu().numpy()
heatmaps_pred_np = np.transpose(heatmaps_pred_np, (1, 2, 0))
heatmaps_np = heatmaps_copy.detach().cpu().numpy()
heatmaps_np = np.transpose(heatmaps_np, (1, 2, 0))
img_np = img_copy.detach().cpu().numpy()
img_np = np.transpose(img_np, (1, 2, 0))
# In case too many figs to print
if step % 5 == 0:
imgutils.show_heatmaps(img_np, heatmaps_pred_np, num_fig=1)
imgutils.show_heatmaps(img_np, heatmaps_np, num_fig=2)
plt.pause(0.5)
print('===================================================')
if (step % 300 == 0) and (step > 0):
suffix_epoch = i
suffix_step = step
path_ckpt_torch = 'models/ckpt_hg_torch_EPOCH' + str(
suffix_epoch) + 'STEP' + str(suffix_step) + '.tar'
torch.save(
{
'epoch': i,
'step': step,
'model_state_dict': net_hg_torch.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
# # print('img_np', img_np.shape)
# # print('heatmaps_np', heatmaps_np.shape)
# img_np = np.swapaxes(np.swapaxes(img_np, 0, 2), 0, 1)
# heatmaps_np = np.swapaxes(np.swapaxes(heatmaps_np, 0, 2), 0, 1)
# print('img_np', img_np.shape)
# print('heatmaps_np', heatmaps_np.shape)
#
# imgutils.show_heatmaps(img_np, heatmaps_np)
# joint_pts = modelutils.heatmaps_to_coords(heatmaps_np, resolu_out=[64,64], prob_threshold=0.0)
# img_np = skimage.transform.resize(img_np, [64,64])
# imgutils.show_stack_joints(img_np, joint_pts, c=[0, 0])
# pass
if __name__ == '__main__':
# ========== Test dataset ==========
ds_torch = DatasetTorch(use_flip=True, use_rand_color=True)
data_loader = datatorch.DataLoader(ds_torch, batch_size=1, shuffle=False)
for step, (img, heatmaps, pts) in enumerate(data_loader):
print('img', img)
# print('heatmaps', heatmaps.shape)
# print('pts', pts.shape)
img = np.transpose(img.squeeze().detach().numpy(), (1, 2, 0))
# img = np.fliplr(img)
heatmaps = np.transpose(heatmaps.squeeze().detach().numpy(), (1, 2, 0))
pts = pts.squeeze().detach().numpy()
print('pts', pts)
print('===========================================================')
imgutils.show_heatmaps(img, heatmaps)
# ================================== Visualize ==================================
# Load a single image
num_eg = 123
img_eg = train_imgs[num_eg, ...] #(256,256,3)
img_eg_np = img_eg.numpy()
img_eg = img_eg[np.newaxis, ...] #(1,256,256,3)
# Load heatmaps
heatmaps_eg = train_heatmaps[num_eg, ...]
heatmaps_eg = heatmaps_eg.numpy()
heatmaps_eg_pred = heatmap_pred = net_hg(img_eg)
heatmaps_eg_pred = heatmaps_eg_pred[-1]
heatmaps_eg_pred = tf.squeeze(heatmaps_eg_pred, axis=0) #(64,64,16)
heatmaps_eg_pred_np = heatmaps_eg_pred.numpy()
coord_joints_eg_pred = modelutils.heatmaps_to_coords(heatmaps_eg_pred,
resolu_out=[256, 256])
coord_joints_true = train_pts[num_eg, ...].numpy() * (256 / 64)
# Show imgs
imgutils.show_stack_joints(img_eg_np, coord_joints_true)
imgutils.show_stack_joints(img_eg_np, coord_joints_eg_pred)
# Show heatmaps
imgutils.show_heatmaps(img_eg_np, heatmaps_eg)
imgutils.show_heatmaps(img_eg_np, heatmaps_eg_pred_np)
# ================================== Visualize ==================================
print('===============================DONE===============================')