MLflow 0.4.2 发布：使用Azure Blob存储，PyTorch和TensorBoard跟踪以及H20支持

问题导读

1.如何获取安装最新版本？
2.MLflow 0.4.2 新增了哪些功能？
3.MLflow 0.4.2修复了哪些功能？



关注最新经典文章，欢迎关注公众号



上一篇：
Spark新作MLflow 0.2发布：集成TensorFlow，Tracking Server更新和S3存储的新功能
http://www.aboutyun.com/forum.php?mod=viewthread&tid=24871


如果不知道MLflow 是什么，参考
Spark MLFlow 介绍
http://www.aboutyun.com/forum.php?mod=viewthread&tid=24862



MLflow 0.4.2已经在PyPI 有说明，相关文档也已经更新。如果按照MLflow快速入门指南中所述进行pip install mlflow，将获得最新版本。



Azure Blob存储组件（Artifact ）支持
作为MLflow 0.4.0的一部分，我们通过mlflow server命令中参数-default-artifact-root添加了对Azure Blob存储中存件的支持。 这样可以轻松地在多个Azure云虚拟机上运行MLflow训练作业，并跟踪它们的结果。 以下示例显示如何使用Azure Blob存储Artifact 库启动跟踪服务器。 需要设置AZURE_STORAGE_CONNECTION_STRING环境变量，如MLflow Tracking> Azure Blob Storage中所述。
[mw_shl_code=bash,true]mlflow server --default-artifact-root wasbs://$container@$account.blob.core.windows.net/
[/mw_shl_code]

使用MLflow与PyTorch和Tensorboard
新版添加了一些包含高级跟踪的示例，包括带有以下MLflow UI和TensorBoard输出的PyTorch TensorBoard示例。[mw_shl_code=python,true]#
# Trains an MNIST digit recognizer using PyTorch, and uses tensorboardX to log training metrics
# and weights in TensorBoard event format to the MLflow run's artifact directory. This stores the
# TensorBoard events in MLflow for later access using the TensorBoard command line tool.
#
# NOTE: This example requires you to first install PyTorch (using the instructions at pytorch.org)
#       and tensorboardX (using pip install tensorboardX).
#
# Code based on https://github.com/lanpa/tensorb ... ster/mnist/main.py.
#

from __future__ import print_function
import argparse
import os
import mlflow
import tempfile
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
from torch.autograd import Variable
from tensorboardX import SummaryWriter

# Command-line arguments
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size', type=int, default=64, metavar='N',
                    help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
                    help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs', type=int, default=10, metavar='N',
                    help='number of epochs to train (default: 10)')
parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
                    help='learning rate (default: 0.01)')
parser.add_argument('--momentum', type=float, default=0.5, metavar='M',
                    help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda', action='store_true', default=False,
                    help='disables CUDA training')
parser.add_argument('--seed', type=int, default=1, metavar='S',
                    help='random seed (default: 1)')
parser.add_argument('--log-interval', type=int, default=100, metavar='N',
                    help='how many batches to wait before logging training status')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()

torch.manual_seed(args.seed)
if args.cuda:
    torch.cuda.manual_seed(args.seed)

kwargs = {'num_workers': 1, 'pin_memory': True} if args.cuda else {}
train_loader = torch.utils.data.DataLoader(
    datasets.MNIST('../data', train=True, download=True,
                   transform=transforms.Compose([
                       transforms.ToTensor(),
                       transforms.Normalize((0.1307,), (0.3081,))
                   ])),
    batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
    datasets.MNIST('../data', train=False, transform=transforms.Compose([
                       transforms.ToTensor(),
                       transforms.Normalize((0.1307,), (0.3081,))
                   ])),
    batch_size=args.batch_size, shuffle=True, **kwargs)

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 10, kernel_size=5)
        self.conv2 = nn.Conv2d(10, 20, kernel_size=5)
        self.conv2_drop = nn.Dropout2d()
        self.fc1 = nn.Linear(320, 50)
        self.fc2 = nn.Linear(50, 10)

    def forward(self, x):
        x = F.relu(F.max_pool2d(self.conv1(x), 2))
        x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
        x = x.view(-1, 320)
        x = F.relu(self.fc1(x))
        x = F.dropout(x, training=self.training)
        x = self.fc2(x)
        return F.log_softmax(x, dim=0)

    def log_weights(self, step):
        writer.add_histogram('weights/conv1/weight', model.conv1.weight.data, step)
        writer.add_histogram('weights/conv1/bias', model.conv1.bias.data, step)
        writer.add_histogram('weights/conv2/weight', model.conv2.weight.data, step)
        writer.add_histogram('weights/conv2/bias', model.conv2.bias.data, step)
        writer.add_histogram('weights/fc1/weight', model.fc1.weight.data, step)
        writer.add_histogram('weights/fc1/bias', model.fc1.bias.data, step)
        writer.add_histogram('weights/fc2/weight', model.fc2.weight.data, step)
        writer.add_histogram('weights/fc2/bias', model.fc2.bias.data, step)

model = Net()
if args.cuda:
    model.cuda()

optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum)

writer = None # Will be used to write TensorBoard events

def train(epoch):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        if args.cuda:
            data, target = data.cuda(), target.cuda()
        data, target = Variable(data), Variable(target)
        optimizer.zero_grad()
        output = model(data)
        loss = F.nll_loss(output, target)
        loss.backward()
        optimizer.step()
        if batch_idx % args.log_interval == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                epoch, batch_idx * len(data), len(train_loader.dataset),
                100. * batch_idx / len(train_loader), loss.data.item()))
            step = epoch * len(train_loader) + batch_idx
            log_scalar('train_loss', loss.data.item(), step)
            model.log_weights(step)

def test(epoch):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for data, target in test_loader:
            if args.cuda:
                data, target = data.cuda(), target.cuda()
            data, target = Variable(data), Variable(target)
            output = model(data)
            test_loss += F.nll_loss(output, target, reduction='sum').data.item() # sum up batch loss
            pred = output.data.max(1)[1] # get the index of the max log-probability
            correct += pred.eq(target.data).cpu().sum().item()

    test_loss /= len(test_loader.dataset)
    test_accuracy = 100.0 * correct / len(test_loader.dataset)
    print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
        test_loss, correct, len(test_loader.dataset), test_accuracy))
    step = (epoch + 1) * len(train_loader)
    log_scalar('test_loss', test_loss, step)
    log_scalar('test_accuracy', test_accuracy, step)

def log_scalar(name, value, step):
    """Log a scalar value to both MLflow and TensorBoard"""
    writer.add_scalar(name, value, step)
    mlflow.log_metric(name, value)

with mlflow.start_run():
    # Log our parameters into mlflow
    for key, value in vars(args).items():
        mlflow.log_param(key, value)
   
    # Create a SummaryWriter to write TensorBoard events locally
    output_dir = dirpath = tempfile.mkdtemp()
    writer = SummaryWriter(output_dir)
    print("Writing TensorBoard events locally to %s\n" % output_dir)

    # Perform the training
    for epoch in range(1, args.epochs + 1):
        train(epoch)
        test(epoch)

    # Upload the TensorBoard event logs as a run artifact
    print("Uploading TensorBoard events as a run artifact...")
    mlflow.log_artifacts(output_dir, artifact_path="events")
    print("\nLaunch TensorBoard with:\n\ntensorboard --logdir=%s" %
        os.path.join(mlflow.get_artifact_uri(), "events"))[/mw_shl_code]




H2O整合

由于PR 170，MLflow现在包括对H2O模型export 和服务的支持; 看看h2o_example.ipynb Jupyter笔记本。


其他功能和错误修复

除了这些功能外，这些版本还包含其他项目，错误和文档修复程序。 值得注意的一些项目是：

• MLflow实验REST API和mlflow实验现在创建支持提供--artifact-location（问题＃232）
• [UI]在UI中显示从http（s）：// GitHub URL运行的项目的GitHub链接（问题＃235）
• 在向/从分布式文件系统保存/加载模型时修复Spark模型支持（问题＃180）
• [跟踪] GCS工件存储现在是可插拔的依赖项（默认情况下不再安装）。 要启用GCS支持，请通过pip在客户端和跟踪服务器上安装google-cloud-storage。 （问题＃202）
• [Projects]支持在Git repos的子目录中运行项目（问题＃153）
• [SageMaker]支持在部署到SageMaker时指定计算规范（问题＃185）