From 72e291f2567ddd63e7f320ed8010ae50e3085c0b Mon Sep 17 00:00:00 2001 From: Jeff Wang Date: Wed, 7 Mar 2018 14:15:16 -0800 Subject: [PATCH] Add English tutorials (#299) * Provide English tutorial on how to use VisualDL in Keras * Add the English Tutorial of PyTorch. * Add the English MXNet tutorial. * Remove links in the MXNet tutorial. --- demo/keras/TUTORIAL_EN.md | 90 ++++++++++++++++ demo/mxnet/TUTORIAL_CN.md | 10 +- demo/mxnet/TUTORIAL_EN.md | 84 +++++++++++++++ demo/pytorch/TUTORIAL_EN.md | 210 ++++++++++++++++++++++++++++++++++++ 4 files changed, 389 insertions(+), 5 deletions(-) create mode 100644 demo/keras/TUTORIAL_EN.md create mode 100644 demo/mxnet/TUTORIAL_EN.md create mode 100644 demo/pytorch/TUTORIAL_EN.md diff --git a/demo/keras/TUTORIAL_EN.md b/demo/keras/TUTORIAL_EN.md new file mode 100644 index 00000000..79c8bf60 --- /dev/null +++ b/demo/keras/TUTORIAL_EN.md @@ -0,0 +1,90 @@ +# How to use VisualDL in Keras + +Here we will show you how to use VisualDL with Keras so that you can visualize the training process of Keras. +We will use the Keras Convolution Neural Network to train the [MNIST](http://yann.lecun.com/exdb/mnist/) dataset as an example. + +The training program comes from the official GitHub [Example](https://github.com/keras-team/keras/blob/master/examples/mnist_cnn.py) of Keras. +We just need to create the VisualDL data collection loggers in the code + +```python +# create VisualDL logger +logdir = "/workspace" +logger = LogWriter(logdir, sync_cycle=100) + +# mark the components with 'train' label. +with logger.mode("train"): + # create a scalar component called 'scalars/' + scalar_keras_train_loss = logger.scalar( + "scalars/scalar_keras_train_loss") + image_input = logger.image("images/input", 1) + image0 = logger.image("images/image0", 1) + image1 = logger.image("images/image1", 1) + histogram0 = logger.histogram("histogram/histogram0", num_buckets=50) + histogram1 = logger.histogram("histogram/histogram1", num_buckets=50) + +``` + +Then we can insert our data loggers in the callback function provided by Keras. + +```python +train_step = 0 + +class LossHistory(keras.callbacks.Callback): + def on_train_begin(self, logs={}): + self.losses = [] + + def on_batch_end(self, batch, logs={}): + global train_step + + # Scalar + scalar_keras_train_loss.add_record(train_step, logs.get('loss')) + + # get weights for 2 layers + W0 = model.layers[0].get_weights()[0] # 3 x 3 x 1 x 32 + W1 = model.layers[1].get_weights()[0] # 3 x 3 x 32 x 64 + + weight_array0 = W0.flatten() + weight_array1 = W1.flatten() + + # histogram + histogram0.add_record(train_step, weight_array0) + histogram1.add_record(train_step, weight_array1) + + # image + image_input.start_sampling() + image_input.add_sample([28, 28], x_train[0].flatten()) + image_input.finish_sampling() + + image0.start_sampling() + image0.add_sample([9, 32], weight_array0) + image0.finish_sampling() + + image1.start_sampling() + image1.add_sample([288, 64], weight_array1) + image1.finish_sampling() + + train_step += 1 + self.losses.append(logs.get('loss')) +``` + +After the training, the visual results of each component are as follows: + +The scalar diagram of the error is as follows: + +

+ +

+ +The input picture and the first, second layer convolution weight after the training are as follows: + +

+ +

+ +The histograms of the training parameters is as follows: + +

+ +

+ +The full demonstration code can be downloaded in [here](./keras_mnist_demo.py). \ No newline at end of file diff --git a/demo/mxnet/TUTORIAL_CN.md b/demo/mxnet/TUTORIAL_CN.md index ae1e8bd8..b108b678 100644 --- a/demo/mxnet/TUTORIAL_CN.md +++ b/demo/mxnet/TUTORIAL_CN.md @@ -9,7 +9,7 @@ ## 安装MXNet -请按照MXNet的[官方网站](https://mxnet.incubator.apache.org/get_started/install.html)来安装MXNet，并验证安装成功。 +请按照MXNet的[官方网站](https://mxnet.incubator.apache.org/install/index.html)来安装MXNet，并验证安装成功。 >>> import mxnet as mx @@ -29,7 +29,7 @@ pip install --upgrade dist/visualdl-*.whl ## 开始编写训练MNIST的程序 -我们为您提供了一个演示程序 [mxnet_demo.py](./mxnet_demo.py)。里面展示了如何下载MNIST数据集以及编写MXNet程序来进行CNN的训练。MXNet的部分借鉴了MXNet[官方入门文件](https://mxnet.incubator.apache.org/tutorials/python/mnist.html) +我们为您提供了一个演示程序 [mxnet_demo.py](https://github.com/PaddlePaddle/VisualDL/blob/develop/demo/mxnet/mxnet_demo.py)。里面展示了如何下载MNIST数据集以及编写MXNet程序来进行CNN的训练。MXNet的部分借鉴了MXNet[官方入门文件](https://mxnet.incubator.apache.org/tutorials/python/mnist.html) 为了嵌入VisualDL程序，以便在MXNet训练时进行检测，我们需要声明一个LogWriter实例： ```python @@ -63,7 +63,7 @@ lenet_model.fit(train_iter, ## 用VisualDL展示模型图 -VisualDL的一个优点是能可视化深度学习模型，帮助用户更直观的了解模型的构成，都有哪些操作，哪些输入等等。VisualDL的模型图支持原生态的PaddlePaddle格式以及普遍适用的ONNX格式。在这里用户可以使用MXNet训练模型，然后用 [ONNX-MXNet] (https://github.com/onnx/onnx-mxnet) 工具将其转换成 ONNX 格式，然后进行可视化。 +VisualDL的一个优点是能可视化深度学习模型，帮助用户更直观的了解模型的构成，都有哪些操作，哪些输入等等。VisualDL的模型图支持原生态的PaddlePaddle格式以及普遍适用的ONNX格式。在这里用户可以使用MXNet训练模型，然后用 [ONNX-MXNet](https://github.com/onnx/onnx-mxnet) 工具将其转换成 ONNX 格式，然后进行可视化。 我们这里使用已经从MXNet转换到ONNX的现成模型 [Super_Resolution model](https://s3.amazonaws.com/onnx-mxnet/examples/super_resolution.onnx) VisualDL的使用很简单，在完成安装后只需要把模型文件（protobuf格式）用参数 -m 提供给VisualDL即可。 @@ -74,6 +74,6 @@ visualDL --logdir=/workspace -m /workspace/super_resolution_mnist.onnx --port=88 模型图的效果如下： -

+

-生成的完整效果图可以在[这里](./super_resolution_graph.png)下载。 +生成的完整效果图可以在[这里](https://github.com/PaddlePaddle/VisualDL/blob/develop/demo/mxnet/super_resolution_graph.png)下载。 diff --git a/demo/mxnet/TUTORIAL_EN.md b/demo/mxnet/TUTORIAL_EN.md new file mode 100644 index 00000000..64e38369 --- /dev/null +++ b/demo/mxnet/TUTORIAL_EN.md @@ -0,0 +1,84 @@ +# How to use VisualDL in MXNet + +Here we will show you how to use VisualDL in MXNet so that you can visualize the training process of MXNet. +We will use the MXNet Convolution Neural Network to train the [MNIST](http://yann.lecun.com/exdb/mnist/) dataset as an example. + +## Install MXNet +Please install MXNet according to MXNet's [official website](https://mxnet.incubator.apache.org/install/index.html) +and verify that the installation is successful. + + >>> import mxnet as mx + >>> a = mx.nd.ones((2, 3)) + >>> b = a * 2 + 1 + >>> b.asnumpy() + array([[ 3., 3., 3.], + [ 3., 3., 3.]], dtype=float32) + +## Install VisualDL +The installation of VisualDL is very simple. Please install it according to the [official website](https://github.com/PaddlePaddle/VisualDL) of VisualDL. +Only two steps are required. + +``` +python setup.py bdist_wheel +pip install --upgrade dist/visualdl-*.whl +``` + +## Start writing the program for training MNIST + +We have provided you with a demonstration program [mxnet_demo.py](https://github.com/PaddlePaddle/VisualDL/blob/develop/demo/mxnet/mxnet_demo.py). +It shows how to download MNIST data sets and write a MXNet programs for CNN training. +The training program is based on the [MXNet tutorial](https://mxnet.incubator.apache.org/tutorials/python/mnist.html) + +We need to create a VisualDL LogWriter instance to record MXNet training: + +```python +logger = LogWriter(logdir, sync_cycle=30) +``` + +The logger instance contains three modules, Scalar, Image, and Histogram. Here we use the Scalar module: + +```python +scalar0 = logger.scalar("scalars/scalar0") +``` + +The Tag can contain '/' in order to create a different namespace for complex model. + +MXNet provides a lot of [API](https://mxnet.incubator.apache.org/api/python/index.html) in the fit function. +We insert our callback function `add_scalar` into the corresponding API + +```python +lenet_model.fit(train_iter, + eval_data=val_iter, + optimizer='sgd', + optimizer_params={'learning_rate':0.1}, + eval_metric='acc', + # Here we embed our custom callback function + batch_end_callback=[add_scalar()], + num_epoch=2) +``` + +That's all. In the training process of MXNet, our callback function is called to record the accuracy at the end of each training batch. +The rate of accuracy will continue to rise until more than 95%. +The following is the accuracy of the two epochs: + +

+ +## Display the model graph with VisualDL + +VisualDL helps users understand the composition of the model more intuitively by visualizing deep learning models. +VisualDL can visualize ONNX format models, which is widely supported. +Users may use MXNet to train the model, then convert it into ONNX format with [ONNX-MXNet](https://github.com/onnx/onnx-mxnet) tool, and then visualize it. + +Here we use the existing model that has been transformed from MXNet to ONNX, [Super_Resolution model](https://s3.amazonaws.com/onnx-mxnet/examples/super_resolution.onnx). + +To display the model graph via VisualDL, pass the model file path with the parameter -m to the VisualDL + +```bash +visualDL --logdir=/workspace -m /workspace/super_resolution_mnist.onnx --port=8888 +``` + +The model graph is as follows: + +

+ +You can download the full size image [here](https://github.com/PaddlePaddle/VisualDL/blob/develop/demo/mxnet/super_resolution_graph.png). diff --git a/demo/pytorch/TUTORIAL_EN.md b/demo/pytorch/TUTORIAL_EN.md new file mode 100644 index 00000000..7801b604 --- /dev/null +++ b/demo/pytorch/TUTORIAL_EN.md @@ -0,0 +1,210 @@ +# How to use VisualDL in PyTorch + +Here we will show you how to use VisualDL in PyTorch so that you can visualize the training process of PyTorch. +We will use the PyTorch Convolution Neural Network to train the [Cifar10](https://www.cs.toronto.edu/~kriz/cifar.html) dataset as an example. + +The training program comes from the [PyTorch Tutorial](http://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html). + +```python +import torch +import torchvision +import torchvision.transforms as transforms +from torch.autograd import Variable +import torch.nn as nn +import torch.nn.functional as F +import torch.optim as optim + +import matplotlib +matplotlib.use('Agg') + +from visualdl import LogWriter + + +transform = transforms.Compose( + [transforms.ToTensor(), + transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]) + +trainset = torchvision.datasets.CIFAR10(root='./data', train=True, + download=True, transform=transform) +trainloader = torch.utils.data.DataLoader(trainset, batch_size=500, + shuffle=True, num_workers=2) + +testset = torchvision.datasets.CIFAR10(root='./data', train=False, + download=True, transform=transform) +testloader = torch.utils.data.DataLoader(testset, batch_size=500, + shuffle=False, num_workers=2) + +classes = ('plane', 'car', 'bird', 'cat', + 'deer', 'dog', 'frog', 'horse', 'ship', 'truck') + + +import matplotlib.pyplot as plt +import numpy as np + + +# functions to show an image +def imshow(img): + img = img / 2 + 0.5 # unnormalize + npimg = img.numpy() + fig, ax = plt.subplots() + plt.imshow(np.transpose(npimg, (1, 2, 0))) + # we can either show the image or save it locally + # plt.show() + fig.savefig('out' + str(np.random.randint(0, 10000)) + '.pdf') +``` + +We can preview the Cifar10 picture set to be analyzed: + +

+ +

+ +We just need to create the VisualDL data collection loggers in the code + +```python +logdir = "/workspace" +logger = LogWriter(logdir, sync_cycle=100) + +# mark the components with 'train' label. +with logger.mode("train"): + # create a scalar component called 'scalars/' + scalar_pytorch_train_loss = logger.scalar("scalars/scalar_pytorch_train_loss") + image1 = logger.image("images/image1", 1) + image2 = logger.image("images/image2", 1) + histogram0 = logger.histogram("histogram/histogram0", num_buckets=100) +``` + +There are 50000 training images and 10000 test images in Cifar10. We set the training set size to 500, +and picture sampling rate to 500. The training set (batch) dimension is: +500 x 3 x 32 x 32 + +Then we start to build the CNN model + +```python +# get some random training images +dataiter = iter(trainloader) +images, labels = dataiter.next() + +# show images +imshow(torchvision.utils.make_grid(images)) +# print labels +print(' '.join('%5s' % classes[labels[j]] for j in range(4))) + +# Define a Convolution Neural Network +class Net(nn.Module): + def __init__(self): + super(Net, self).__init__() + self.conv1 = nn.Conv2d(3, 6, 5) + self.pool = nn.MaxPool2d(2, 2) + self.conv2 = nn.Conv2d(6, 16, 5) + self.fc1 = nn.Linear(16 * 5 * 5, 120) + self.fc2 = nn.Linear(120, 84) + self.fc3 = nn.Linear(84, 10) + + def forward(self, x): + x = self.pool(F.relu(self.conv1(x))) + x = self.pool(F.relu(self.conv2(x))) + x = x.view(-1, 16 * 5 * 5) + x = F.relu(self.fc1(x)) + x = F.relu(self.fc2(x)) + x = self.fc3(x) + return x + + +net = Net() +``` + +Then we start to train and use VisualDL to collect data at the same time + +```python +# Train the network +for epoch in range(5): # loop over the dataset multiple times + running_loss = 0.0 + for i, data in enumerate(trainloader, 0): + # get the inputs + inputs, labels = data + + # wrap them in Variable + inputs, labels = Variable(inputs), Variable(labels) + + # zero the parameter gradients + optimizer.zero_grad() + + # forward + backward + optimize + outputs = net(inputs) + loss = criterion(outputs, labels) + + loss.backward() + optimizer.step() + + # use VisualDL to retrieve metrics + # scalar + scalar_pytorch_train_loss.add_record(train_step, float(loss)) + + # histogram + weight_list = net.conv1.weight.view(6*3*5*5, -1) + histogram0.add_record(train_step, weight_list) + + # image + image1.start_sampling() + image1.add_sample([96, 25], net.conv2.weight.view(16*6*5*5, -1)) + image1.finish_sampling() + + image2.start_sampling() + image2.add_sample([18, 25], net.conv1.weight.view(6*3*5*5, -1)) + image2.finish_sampling() + + + train_step += 1 + + # print statistics + running_loss += loss.data[0] + if i % 2000 == 1999: # print every 2000 mini-batches + print('[%d, %5d] loss: %.3f' % + (epoch + 1, i + 1, running_loss / 2000)) + running_loss = 0.0 + +print('Finished Training') +``` + +PyTorch support ONNX standard and it can export its model into ONNX. +PyTorch runs a single round of inference to trace the graph. We use a dummy input to run the model to produce the ONNX model + +```python +import torch.onnx +dummy_input = Variable(torch.randn(4, 3, 32, 32)) +torch.onnx.export(net, dummy_input, "pytorch_cifar10.onnx") + +print('Done') +``` + +After the training, the visual results of each component are as follows: + +The scalar diagram of the error is as follows: + +

+ +

+ +The picture of the first, second layer convolution weight after the training are as follows: + +

+ +

+ + +The histograms of the training parameters is as follows: + +

+ +

+ + +The model graph is as follows: + +

+ +

+ + +You can download the full size image [here](https://github.com/daming-lu/large_files/blob/master/pytorch_demo_figs/graph.png?Raw=true). -- GitLab