MultiGarment-Network under Conda Environment

How to configure Conda environments for MultiGarment-Network (both Python 2 & 3).

· 6 min read
MultiGarment-Network under Conda Environment
Multi-Garment Net: Learning to Dress 3D People from Images
Multi-Garment Net: Learning to Dress 3D People from Images
We present Multi-Garment Network (MGN), a method to predict body shape andclothing, layered on top of the SMPL model from a few frames (1-8) of a video.Several experiments demonstrate that this representation allows higher level ofcontrol when compared to single mesh or voxel representations of s…
bharat-b7/MultiGarmentNetwork
Repo for “Multi-Garment Net: Learning to Dress 3D People from Images, ICCV′19” - bharat-b7/MultiGarmentNetwork

Basic environment: CentOS 7, NVIDIA drivers (CUDA 10.1.243), Conda 4.10.

I hate Conda!

Python 2 Environment

First attempt is with Python 2.7 since the readme says

The code has been tested in python 2.7, Tensorflow 1.13

Create Conda environment:

$ conda create -n mgn-py27 python=2.7 cudatoolkit=10.1 cudnn=7.6.5
$ conda activate mgn-py27

MultiGarment-Network has 2 dependencies that we need to install manually.

Install dirt

pmh47/dirt
DIRT: a fast differentiable renderer for TensorFlow - pmh47/dirt

Install dependencies:

(mgn-py27)$ conda install cmake gcc_linux-64
(mgn-py27)$ conda install tensorflow-gpu=1.13

But pip list is not showing tensorflow-gpu package. So we need to install it again from pypi before building and installation:

(mgn-py27)$ pip install --ignore-installed tensorflow==1.13.1
(mgn-py27)$ pip install .

Install Mesh

MPI-IS/mesh
MPI-IS Mesh Processing Library. Contribute to MPI-IS/mesh development by creating an account on GitHub.
(mgn-py27)$ conda install gxx_linux-64 opencv
(mgn-py27)$ git checkout 1761d544686b3735991954947a8befa759891eb4
(mgn-py27)$ make
(mgn-py27)$ cd dist && pip install psbody_mesh-0.1-cp27-cp27mu-linux_x86_64.whl

Run MultiGarment-Network

(mgn-py27)$ conda install matlabplot
(mgn-py27)$ pip install "scikit-learn<0.18" chumpy
(mgn-py27)$ python test_network.py
Using dirt renderer.
....
Done
freeglut (mesh_viewer):  ERROR:  Internal error <FBConfig with necessary capabilities not found> in function fgOpenWindow

Python 3 Environment (WIP)

Create a Conda environment:

$ conda create -n mgn-py36 python=3.6 cudatoolkit=10.0 cudnn=7.6.5
$ conda activate mgn-py36

Install dirt

Note the troubleshooting section in the description:

If you are using TensorFlow 1.14, there are some binary compatibility issues when using older versions of python (e.g. 2.7 and 3.5), due to compiler version mismatches. These result in a segfault at tensorflow::shape_inference::InferenceContext::GetAttr or similar. To resolve, either upgrade python to 3.7, or downgrade TensorFlow to 1.13, or build DIRT with gcc 4.8

Thus we will use an older version 1.13 of Tensorflow to avoid this known issue. And since the package from conda-forge messes up the dependencies, we must specify channel to anaconda:

(mgn-py36)$ conda install -c anaconda tensorflow-gpu=1.13
(mgn-py36)$ pip install tensorflow-gpu==1.13.1

Then install rest dependencies for building:

(mgn-py36)$ conda install cmake gcc_linux-64

Finally build and install dirt, but with some environment variables:

(mgn-py36)$ export CUDA_HOME=/usr/local/cuda-10.1
(mgn-py36)$ export PATH=$CUDA_HOME/bin:$PATH
(mgn-py36)$ pip install .
(mgn-py36)$ python tests/square_test.py
....
successful: all pixels agree

Install Mesh

(mgn-py36)$ conda install boost
(mgn-py36)$ conda install pyopengl pillow pyzmq pyyaml
(mgn-py36)$ conda install gxx_linux-64
(mgn-py36)$ make all
(mgn-py36)$ make tests
....
Ran 28 tests in 5.882s

Convert MultiGarment-Network to Python 3

(mgn-py36)$ conda install matplotlib scikit-learn==0.17 chumpy

Since we are running Python 3, we will need to replace cPickle with _pickle as the former does not exist.

(mgn-py36)$ find ./ -type f -name \*.py -exec sed -i -e 's/cPickle/_pickle/g' {} \;
(mgn-py36)$ 2to3 -w -n MultiGarmentNetwork/

Due to some incompatibility, Pickle un-serialization will fail:

Traceback (most recent call last):
  File "test_network.py", line 171, in <module>
    _, faces = pkl.load(f)
UnicodeDecodeError: 'ascii' codec can't decode byte 0x8c in position 16: ordinal not in range(128)

You will need to change the Pickle loading to

_, faces = pkl.load(open(file_path, 'rb'), encoding='latin1')

A too-old scikit-learn is also a problem:

Traceback (most recent call last):
  File "test_network.py", line 176, in <module>
    pca_verts[garment] = pkl.load(f)
  File "$CONDA_HOME/lib/python3.6/site-packages/sklearn/decomposition/__init__.py", line 10, in <module>             
    from .kernel_pca import KernelPCA
  File "$CONDA_HOME/lib/python3.6/site-packages/sklearn/decomposition/kernel_pca.py", line 13, in <module>           
    from ..metrics.pairwise import pairwise_kernels
  File "$CONDA_HOME/lib/python3.6/site-packages/sklearn/metrics/__init__.py", line 33, in <module>                   
    from . import cluster
  File "$CONDA_HOME/lib/python3.6/site-packages/sklearn/metrics/cluster/__init__.py", line 8, in <module>
    from .supervised import adjusted_mutual_info_score
  File "$CONDA_HOME/lib/python3.6/site-packages/sklearn/metrics/cluster/supervised.py", line 14, in <module>         
    from scipy.misc import comb
ImportError: cannot import name 'comb'

Changing to from scipy.special import comb will solve the problem.

Until now the network should run without syntax error. Some report that the network could run after the above changes.

But here we will still get this error:

Traceback (most recent call last):
  File "test_network.py", line 185, in <module>
    pred = get_results(m, dat)
  File "test_network.py", line 55, in get_results
    out = m([images, vertex_label, J_2d])
  File "$CONDA_PREFIX/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py", line 592, in __call__
    outputs = self.call(inputs, *args, **kwargs)
  File "/public/wl4/clothing/MultiGarmentNetwork-py3/network/base_network.py", line 336, in call
    garm_model_outputs = [fe(latent_code_offset_ShapeMerged) for fe in self.garmentModels]
  File "/public/wl4/clothing/MultiGarmentNetwork-py3/network/base_network.py", line 336, in <listcomp>
    garm_model_outputs = [fe(latent_code_offset_ShapeMerged) for fe in self.garmentModels]
  File "$CONDA_PREFIX/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py", line 592, in __call__
    outputs = self.call(inputs, *args, **kwargs)
  File "/public/wl4/clothing/MultiGarmentNetwork-py3/network/base_network.py", line 65, in call
    x = self.PCA_(pca_comp)
  File "$CONDA_PREFIX/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py", line 592, in __call__
    outputs = self.call(inputs, *args, **kwargs)
  File "/public/wl4/clothing/MultiGarmentNetwork-py3/network/custom_layers.py", line 33, in call
    return tf.reshape(tf.matmul(x, self.components) + self.mean, (-1, K.int_shape(self.mean)[0] / 3, 3))
  File "$CONDA_PREFIX/lib/python3.6/site-packages/tensorflow/python/ops/gen_array_ops.py", line 7161, in reshape
    tensor, shape, name=name, ctx=_ctx)
  File "$CONDA_PREFIX/lib/python3.6/site-packages/tensorflow/python/ops/gen_array_ops.py", line 7206, in reshape_eager_fallback
    ctx=_ctx, name=name)
  File "$CONDA_PREFIX/lib/python3.6/site-packages/tensorflow/python/eager/execute.py", line 66, in quick_execute
    six.raise_from(core._status_to_exception(e.code, message), None)
  File "<string>", line 3, in raise_from
tensorflow.python.framework.errors_impl.InvalidArgumentError: Value for attr 'Tshape' of float is not in the list of allowed values: int32, int64
        ; NodeDef: {{node Reshape}}; Op<name=Reshape; signature=tensor:T, shape:Tshape -> output:T; attr=T:type; attr=Tshape:type,default=DT_INT32,allowed=[DT_INT32, DT_INT64]> [Op:Reshape]

There is also a repository containing Python 3 version of MultiGarment-Network:

minar09/MGN-Py3
Multi-Garment Network implementation for Python3. Contribute to minar09/MGN-Py3 development by creating an account on GitHub.

Related Articles

Rancher Server on RKE Deployment
· 5 min read
Using Linux Containers (LXC) on Fedora
· 3 min read

Notes on Redis Operation

Some common problems and solution to containerized Redis deployment.

· 1 min read