smpl_torch.py

import numpy as np
import pickle
import torch
from torch.nn import Module
import os


class SMPLModel(Module):
  def __init__(self, device=None, model_path='./model.pkl'):
    
    super(SMPLModel, self).__init__()
    with open(model_path, 'rb') as f:
      params = pickle.load(f)
    self.J_regressor = torch.from_numpy(
      np.array(params['J_regressor'].todense())
    ).type(torch.float64)
    self.weights = torch.from_numpy(params['weights']).type(torch.float64)
    self.posedirs = torch.from_numpy(params['posedirs']).type(torch.float64)
    self.v_template = torch.from_numpy(params['v_template']).type(torch.float64)
    self.shapedirs = torch.from_numpy(params['shapedirs']).type(torch.float64)
    self.kintree_table = params['kintree_table']
    self.faces = params['f']
    self.device = device if device is not None else torch.device('cpu')
    for name in ['J_regressor', 'weights', 'posedirs', 'v_template', 'shapedirs']:
      _tensor = getattr(self, name)
      setattr(self, name, _tensor.to(device))

  @staticmethod
  def rodrigues(r):
    """
    Rodrigues' rotation formula that turns axis-angle tensor into rotation
    matrix in a batch-ed manner.

    Parameter:
    ----------
    r: Axis-angle rotation tensor of shape [batch_size, 1, 3].

    Return:
    -------
    Rotation matrix of shape [batch_size, 3, 3].

    """
    #r = r.to(self.device)
    eps = r.clone().normal_(std=1e-8)
    theta = torch.norm(r + eps, dim=(1, 2), keepdim=True)  # dim cannot be tuple
    theta_dim = theta.shape[0]
    r_hat = r / theta
    cos = torch.cos(theta)
    z_stick = torch.zeros(theta_dim, dtype=torch.float64).to(r.device)
    m = torch.stack(
      (z_stick, -r_hat[:, 0, 2], r_hat[:, 0, 1], r_hat[:, 0, 2], z_stick,
       -r_hat[:, 0, 0], -r_hat[:, 0, 1], r_hat[:, 0, 0], z_stick), dim=1)
    m = torch.reshape(m, (-1, 3, 3))
    i_cube = (torch.eye(3, dtype=torch.float64).unsqueeze(dim=0) \
             + torch.zeros((theta_dim, 3, 3), dtype=torch.float64)).to(r.device)
    A = r_hat.permute(0, 2, 1)
    dot = torch.matmul(A, r_hat)
    R = cos * i_cube + (1 - cos) * dot + torch.sin(theta) * m
    return R

  @staticmethod
  def with_zeros(x):
    """
    Append a [0, 0, 0, 1] tensor to a [3, 4] tensor.

    Parameter:
    ---------
    x: Tensor to be appended.

    Return:
    ------
    Tensor after appending of shape [4,4]

    """
    ones = torch.tensor([[0.0, 0.0, 0.0, 1.0]], dtype=torch.float64).to(x.device)
    ret = torch.cat((x, ones), dim=0)
    return ret

  @staticmethod
  def pack(x):
    """
    Append zero tensors of shape [4, 3] to a batch of [4, 1] shape tensor.

    Parameter:
    ----------
    x: A tensor of shape [batch_size, 4, 1]

    Return:
    ------
    A tensor of shape [batch_size, 4, 4] after appending.

    """
    zeros43 = torch.zeros((x.shape[0], 4, 3), dtype=torch.float64).to(x.device)
    ret = torch.cat((zeros43, x), dim=2)
    return ret

  def write_obj(self, verts, file_name):
    with open(file_name, 'w') as fp:
      for v in verts:
        fp.write('v %f %f %f\n' % (v[0], v[1], v[2]))

      for f in self.faces + 1:
        fp.write('f %d %d %d\n' % (f[0], f[1], f[2]))

  def forward(self, betas, pose, trans, simplify=False):
    """
          Construct a compute graph that takes in parameters and outputs a tensor as
          model vertices. Face indices are also returned as a numpy ndarray.

          Prameters:
          ---------
          pose: Also known as 'theta', a [24,3] tensor indicating child joint rotation
          relative to parent joint. For root joint it's global orientation.
          Represented in a axis-angle format.

          betas: Parameter for model shape. A tensor of shape [10] as coefficients of
          PCA components. Only 10 components were released by SMPL author.

          trans: Global translation tensor of shape [3].

          Return:
          ------
          A tensor for vertices, and a numpy ndarray as face indices.

    """
    id_to_col = {self.kintree_table[1, i]: i
                 for i in range(self.kintree_table.shape[1])}
    parent = {
      i: id_to_col[self.kintree_table[0, i]]
      for i in range(1, self.kintree_table.shape[1])
    }
    v_shaped = torch.tensordot(self.shapedirs, betas, dims=([2], [0])) + self.v_template
    J = torch.matmul(self.J_regressor, v_shaped)
    R_cube_big = self.rodrigues(pose.view(-1, 1, 3))

    if simplify:
      v_posed = v_shaped
    else:
      R_cube = R_cube_big[1:]
      I_cube = (torch.eye(3, dtype=torch.float64).unsqueeze(dim=0) + \
        torch.zeros((R_cube.shape[0], 3, 3), dtype=torch.float64)).to(self.device)
      lrotmin = torch.reshape(R_cube - I_cube, (-1, 1)).squeeze()
      v_posed = v_shaped + torch.tensordot(self.posedirs, lrotmin, dims=([2], [0]))

    results = []
    results.append(
      self.with_zeros(torch.cat((R_cube_big[0], torch.reshape(J[0, :], (3, 1))), dim=1))
    )
    for i in range(1, self.kintree_table.shape[1]):
      results.append(
        torch.matmul(
          results[parent[i]],
          self.with_zeros(
            torch.cat(
              (R_cube_big[i], torch.reshape(J[i, :] - J[parent[i], :], (3, 1))),
              dim=1
            )
          )
        )
      )

    stacked = torch.stack(results, dim=0)
    results = stacked - \
      self.pack(
        torch.matmul(
          stacked,
          torch.reshape(
            torch.cat((J, torch.zeros((24, 1), dtype=torch.float64).to(self.device)), dim=1),
            (24, 4, 1)
          )
        )
      )
    T = torch.tensordot(self.weights, results, dims=([1], [0]))
    rest_shape_h = torch.cat(
      (v_posed, torch.ones((v_posed.shape[0], 1), dtype=torch.float64).to(self.device)), dim=1
    )
    v = torch.matmul(T, torch.reshape(rest_shape_h, (-1, 4, 1)))
    v = torch.reshape(v, (-1, 4))[:, :3]
    result = v + torch.reshape(trans, (1, 3))
    return result


def test_gpu(gpu_id=[0]):
  if len(gpu_id) > 0 and torch.cuda.is_available():
    os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu_id[0])
    device = torch.device('cuda')
  else:
    device = torch.device('cpu')
  print(device)

  pose_size = 72
  beta_size = 10

  np.random.seed(9608)
  pose = torch.from_numpy((np.random.rand(pose_size) - 0.5) * 0.4)\
          .type(torch.float64).to(device)
  betas = torch.from_numpy((np.random.rand(beta_size) - 0.5) * 0.06) \
          .type(torch.float64).to(device)
  trans = torch.from_numpy(np.zeros(3)).type(torch.float64).to(device)
  outmesh_path = './smpl_torch.obj'

  model = SMPLModel(device=device)
  result = model(betas, pose, trans)
  model.write_obj(result, outmesh_path)


if __name__ == '__main__':
  test_gpu([1])