def _convert_input_data(self, data_tuple):
# Unfolding
x, nan_masks, tasks, tasks_mask, _, _, towards, _, _, _ = data_tuple
# Convert numpy to torch.tensor
tasks = torch.from_numpy(tasks).long().to(self.device)
tasks_mask = torch.from_numpy(tasks_mask * 1 + 1e-5).float().to(
self.device)
nan_masks = torch.from_numpy(nan_masks * 1 + 1e-5).float().to(
self.device)
x, towards = numpy2tensor(
self.device, x,
expand1darr(towards.astype(np.int64), 3, self.seq_dim))
# Construct tuple
input_data = (x, towards)
input_info = (x, nan_masks, tasks, tasks_mask)
return input_data, input_info
def _convert_input_data(self, data_tuple):
# Unfolding
x, nan_masks, fut_np, fut_mask_np, fut_avail_mask_np, tasks_np, tasks_mask_np, phenos_np, phenos_mask_np, towards, _, _, idpatients_np = data_tuple
# Convert numpy to torch.tensor
tasks = torch.from_numpy(tasks_np).long().to(self.device)
tasks_mask = torch.from_numpy(tasks_mask_np * 1 + 1e-5).float().to(self.device)
nan_masks = torch.from_numpy(nan_masks * 1 + 1e-5).float().to(self.device)
fut_mask = torch.from_numpy(fut_mask_np * 1 + 1e-5).float().to(self.device)
fut_avail_mask = torch.from_numpy(fut_avail_mask_np.astype(int)).to(self.device)
x, fut, towards = numpy2tensor(self.device,
x,
fut_np,
expand1darr(towards.astype(np.int64), 3, self.seq_dim)
)
# Construct tuple
input_data = (x, towards, fut_np, fut_mask_np, tasks_np, tasks_mask_np, idpatients_np, phenos_np, phenos_mask_np)
input_info = (x, nan_masks, fut, fut_mask, fut_avail_mask, tasks, tasks_mask)
return input_data, input_info
def forward_decode_only(self, motion_info, towards, num_var_dim, num_datapoints, num_kld):
motion_z, motion_mu, motion_logvar = motion_info
towards = torch.from_numpy(expand1darr(towards.astype(np.int64), 3, self.seq_dim)).float().to(self.device)
kld = torch.mean(-0.5 * (1 + motion_logvar - motion_mu.pow(2) - motion_logvar.exp()), dim=0)
_, sorted_ind = torch.topk(kld, num_kld)
sorted_ind = [42, 90, 12, 70, 108]
recon_motion = torch.zeros(num_kld, num_var_dim*num_datapoints, self.fea_dim, motion_z.shape[1])
for idx, val in enumerate(sorted_ind):
motion_z_one = motion_z[:,val].view(motion_z.shape[0])
# Get 90th percentile of motion_z_one
motion_cpu_copy = motion_z_one.cpu()
max_kld = np.percentile(motion_cpu_copy.numpy(), 90)
# Get 10th percentile of motion_z_one
min_kld = np.percentile(motion_cpu_copy.numpy(), 10)
motion_z_one.cuda()
recon_motion[idx,:,:,:] = self.model.decode_only(motion_z, towards, sorted_ind, min_kld, max_kld, num_var_dim, num_datapoints)
return recon_motion