概要

Conda で仮想環境を作り、その仮想環境の Python で PyTorch を利用可能にし、その Python を PyCall に認識させ、Julia で呼び出す。

Conda https://conda.io/en/latest/
Python https://docs.python.org/3/
PyTorch https://pytorch.org/docs/stable/index.html
PyCall https://github.com/JuliaPy/PyCall.jl
Julia https://docs.julialang.org/en/v1/

手順は次の通りである。

Conda をシェルにインストールする。
Conda で仮想環境を作る。
仮想環境で PyTorch を利用できる状態にする。
Julia をインストールする。
PyCall を仮想環境の Python を参照するようにインストールする。

詳細

Conda のインストール方法は省略する。

conda create -n my_env python=3.8
conda activate my_env
conda install -c pytorch pytorch

実行するとこんなログになる。

paalon at paalon-mac in ~
↪ conda create -n my_env python=3.8                                      (base)
Collecting package metadata (current_repodata.json): done
Solving environment: done

## Package Plan ##

  environment location: /Users/paalon/conda/envs/my_env

  added / updated specs:
    - python=3.8


The following NEW packages will be INSTALLED:

  ca-certificates    pkgs/main/osx-64::ca-certificates-2020.1.1-0
  certifi            pkgs/main/osx-64::certifi-2019.11.28-py38_0
  libcxx             pkgs/main/osx-64::libcxx-4.0.1-hcfea43d_1
  libcxxabi          pkgs/main/osx-64::libcxxabi-4.0.1-hcfea43d_1
  libedit            pkgs/main/osx-64::libedit-3.1.20181209-hb402a30_0
  libffi             pkgs/main/osx-64::libffi-3.2.1-h475c297_4
  ncurses            pkgs/main/osx-64::ncurses-6.2-h0a44026_0
  openssl            pkgs/main/osx-64::openssl-1.1.1d-h1de35cc_4
  pip                pkgs/main/osx-64::pip-20.0.2-py38_1
  python             pkgs/main/osx-64::python-3.8.1-h359304d_1
  readline           pkgs/main/osx-64::readline-7.0-h1de35cc_5
  setuptools         pkgs/main/osx-64::setuptools-45.2.0-py38_0
  sqlite             pkgs/main/osx-64::sqlite-3.31.1-ha441bb4_0
  tk                 pkgs/main/osx-64::tk-8.6.8-ha441bb4_0
  wheel              pkgs/main/osx-64::wheel-0.34.2-py38_0
  xz                 pkgs/main/osx-64::xz-5.2.4-h1de35cc_4
  zlib               pkgs/main/osx-64::zlib-1.2.11-h1de35cc_3


Proceed ([y]/n)? y

Preparing transaction: done
Verifying transaction: done
Executing transaction: done
#
# To activate this environment, use
#
#     $ conda activate my_env
#
# To deactivate an active environment, use
#
#     $ conda deactivate

paalon at paalon-mac in ~
↪ conda activate my_env                                                  (base)
paalon at paalon-mac in ~
↪ conda install -c pytorch pytorch                                     (my_env)
Collecting package metadata (current_repodata.json): done
Solving environment: done

## Package Plan ##

  environment location: /Users/paalon/conda/envs/my_env

  added / updated specs:
    - pytorch


The following NEW packages will be INSTALLED:

  blas               pkgs/main/osx-64::blas-1.0-mkl
  intel-openmp       pkgs/main/osx-64::intel-openmp-2019.4-233
  libgfortran        pkgs/main/osx-64::libgfortran-3.0.1-h93005f0_2
  mkl                pkgs/main/osx-64::mkl-2019.4-233
  mkl-service        pkgs/main/osx-64::mkl-service-2.3.0-py38hfbe908c_0
  mkl_fft            pkgs/main/osx-64::mkl_fft-1.0.15-py38h5e564d8_0
  mkl_random         pkgs/main/osx-64::mkl_random-1.1.0-py38h6440ff4_0
  ninja              pkgs/main/osx-64::ninja-1.9.0-py38h04f5b5a_0
  numpy              pkgs/main/osx-64::numpy-1.18.1-py38h7241aed_0
  numpy-base         pkgs/main/osx-64::numpy-base-1.18.1-py38h6575580_1
  pytorch            pytorch/osx-64::pytorch-1.4.0-py3.8_0
  six                pkgs/main/osx-64::six-1.14.0-py38_0


Proceed ([y]/n)? y

Preparing transaction: done
Verifying transaction: done
Executing transaction: done

Julia を起動する。

paalon at paalon-mac in ~
↪ julia                                                                (my_env)
               _
   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.3.1 (2019-12-30)
 _/ |\__'_|_|_|\__'_|  |  Official https://julialang.org/ release
|__/                   |

PyCall を追加し、ENV["PYCALL_JL_RUNTIME_PYTHON"] と ENV["PYTHON"] を Sys.which("python") に設定し、ビルドすると利用可能になる。

(v1.3) pkg> add PyCall
  Updating registry at `~/.julia/registries/General`
  Updating git-repo `https://github.com/JuliaRegistries/General.git`
 Resolving package versions...
  Updating `~/.julia/environments/v1.3/Project.toml`
  [438e738f] + PyCall v1.91.4
  Updating `~/.julia/environments/v1.3/Manifest.toml`
  [8f4d0f93] + Conda v1.4.1
  [438e738f] + PyCall v1.91.4
  [81def892] + VersionParsing v1.2.0

julia> ENV["PYCALL_JL_RUNTIME_PYTHON"] = Sys.which("python")
"/Users/paalon/conda/envs/my_env/bin/python3.8"

julia> ENV["PYTHON"] = Sys.which("python")
"/Users/paalon/conda/envs/my_env/bin/python3.8"

(v1.3) pkg> build PyCall
  Building Conda ─→ `~/.julia/packages/Conda/3rPhK/deps/build.log`
  Building PyCall → `~/.julia/packages/PyCall/zqDXB/deps/build.log`

julia> using PyCall
[ Info: Precompiling PyCall [438e738f-606a-5dbb-bf0a-cddfbfd45ab0]

julia> torch = pyimport("torch")
PyObject <module 'torch' from '/Users/paalon/conda/envs/my_env/lib/python3.8/site-packages/torch/__init__.py'>

julia> x = torch.tensor([1, 2, 3])
PyObject tensor([1, 2, 3])

例

以下の公式チュートリアルの例に書かれているもの

import torch

dtype = torch.float
device = torch.device("cpu")
# device = torch.device("cuda:0") # Uncomment this to run on GPU

# N is batch size; D_in is input dimension;
# H is hidden dimension; D_out is output dimension.
N, D_in, H, D_out = 64, 1000, 100, 10

# Create random Tensors to hold input and outputs.
# Setting requires_grad=False indicates that we do not need to compute gradients
# with respect to these Tensors during the backward pass.
x = torch.randn(N, D_in, device=device, dtype=dtype)
y = torch.randn(N, D_out, device=device, dtype=dtype)

# Create random Tensors for weights.
# Setting requires_grad=True indicates that we want to compute gradients with
# respect to these Tensors during the backward pass.
w1 = torch.randn(D_in, H, device=device, dtype=dtype, requires_grad=True)
w2 = torch.randn(H, D_out, device=device, dtype=dtype, requires_grad=True)

learning_rate = 1e-6
for t in range(500):
    # Forward pass: compute predicted y using operations on Tensors; these
    # are exactly the same operations we used to compute the forward pass using
    # Tensors, but we do not need to keep references to intermediate values since
    # we are not implementing the backward pass by hand.
    y_pred = x.mm(w1).clamp(min=0).mm(w2)

    # Compute and print loss using operations on Tensors.
    # Now loss is a Tensor of shape (1,)
    # loss.item() gets the scalar value held in the loss.
    loss = (y_pred - y).pow(2).sum()
    if t % 100 == 99:
        print(t, loss.item())

    # Use autograd to compute the backward pass. This call will compute the
    # gradient of loss with respect to all Tensors with requires_grad=True.
    # After this call w1.grad and w2.grad will be Tensors holding the gradient
    # of the loss with respect to w1 and w2 respectively.
    loss.backward()

    # Manually update weights using gradient descent. Wrap in torch.no_grad()
    # because weights have requires_grad=True, but we don't need to track this
    # in autograd.
    # An alternative way is to operate on weight.data and weight.grad.data.
    # Recall that tensor.data gives a tensor that shares the storage with
    # tensor, but doesn't track history.
    # You can also use torch.optim.SGD to achieve this.
    with torch.no_grad():
        w1 -= learning_rate * w1.grad
        w2 -= learning_rate * w2.grad

        # Manually zero the gradients after updating weights
        w1.grad.zero_()
        w2.grad.zero_()

をほとんどそのまま Julia に移植すると

ENV["PYCALL_JL_RUNTIME_PYTHON"] = Sys.which("python")
ENV["PYTHON"] = Sys.which("python")
# python の構成を変えたときは次の行を実行してビルドする。
# using Pkg; Pkg.build("PyCall")
using PyCall

torch = pyimport("torch")

dtype = torch.float
device = torch.device("cpu")
# device = torch.device("cuda:0") # Uncomment this to run on GPU

# N is batch size; D_in is input dimension;
# H is hidden dimension; D_out is output dimension.
N, D_in, H, D_out = 64, 1000, 100, 10

# Create random Tensors to hold input and outputs.
# Setting requires_grad=False indicates that we do not need to compute gradients
# with respect to these Tensors during the backward pass.
x = torch.randn(N, D_in, device=device, dtype=dtype)
y = torch.randn(N, D_out, device=device, dtype=dtype)

# Create random Tensors for weights.
# Setting requires_grad=True indicates that we want to compute gradients with
# respect to these Tensors during the backward pass.
w1 = torch.randn(D_in, H, device=device, dtype=dtype, requires_grad=true)
w2 = torch.randn(H, D_out, device=device, dtype=dtype, requires_grad=true)

learning_rate = 1e-6
for t in 1:500
    # Forward pass: compute predicted y using operations on Tensors; these
    # are exactly the same operations we used to compute the forward pass using
    # Tensors, but we do not need to keep references to intermediate values since
    # we are not implementing the backward pass by hand.
    y_pred = x.mm(w1).clamp(min=0).mm(w2)

    # Compute and print loss using operations on Tensors.
    # Now loss is a Tensor of shape (1,)
    # loss.item() gets the scalar value held in the loss.
    loss = (y_pred - y).pow(2).sum()
    if t % 100 == 0
        println("$(t) $(loss.item())")
    end

    # Use autograd to compute the backward pass. This call will compute the
    # gradient of loss with respect to all Tensors with requires_grad=True.
    # After this call w1.grad and w2.grad will be Tensors holding the gradient
    # of the loss with respect to w1 and w2 respectively.
    loss.backward()

    # Manually update weights using gradient descent. Wrap in torch.no_grad()
    # because weights have requires_grad=True, but we don't need to track this
    # in autograd.
    # An alternative way is to operate on weight.data and weight.grad.data.
    # Recall that tensor.data gives a tensor that shares the storage with
    # tensor, but doesn't track history.
    # You can also use torch.optim.SGD to achieve this.
    @pywith torch.no_grad() begin
        # 代入してしまうと、置き換えてしまうので使わないこと。
        # w1 -= learning_rate * w1.grad
        # w2 -= learning_rate * w2.grad
        w1.sub_(learning_rate * w1.grad)
        w2.sub_(learning_rate * w2.grad)
        # Manually zero the gradients after updating weights
        w1.grad.zero_()
        w2.grad.zero_()
    end
end

となる。