diff --git a/demo/mxnet/TUTORIAL_CN.md b/demo/mxnet/TUTORIAL_CN.md new file mode 100644 index 0000000000000000000000000000000000000000..9eb8f952a11acc11c9e9d6dede0057cb39feb773 --- /dev/null +++ b/demo/mxnet/TUTORIAL_CN.md @@ -0,0 +1,60 @@ +# 在MXNet中使用VisualDL + +下面我们演示一下如何在MXNet中使用VisualDL，从而可以把MXNet的训练过程以及最后的模型可视化出来。我们将以MXNet用卷积神经网络(CNN, Convolutional Neural Network)来训练[MNIST](http://yann.lecun.com/exdb/mnist/)数据集作为例子。 + +- [安装MXNet](#%E5%AE%89%E8%A3%85mxnet) +- [安装VisualDL](#%E5%AE%89%E8%A3%85visualdl) +- [开始编写训练MNIST的程序](#%E5%BC%80%E5%A7%8B%E7%BC%96%E5%86%99%E8%AE%AD%E7%BB%83mnist%E7%9A%84%E7%A8%8B%E5%BA%8F) + +## 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。 + + + >>> 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) + +## 安装VisualDL +VisualDL的安装很简单。请按照VisualDL的[官方网站](https://github.com/PaddlePaddle/VisualDL)进行安装。具体只需要两步 + +``` +python setup.py bdist_wheel +pip install --upgrade dist/visualdl-0.0.1-py2-none-any.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) +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： + +``` +logger = LogWriter(logdir, sync_cycle=10) +``` + +logger实例里面包含VisualDL的四个功能模块 Scalar， Image 以及 Histogram。这里我们使用 Scalar 模块： + +``` +scalar0 = logger.scalar("scalars/scalar0") +``` + +模块的命名可以有 '/', 以便对于复杂模型创建不同的命名空间。 + +MXNet在fit函数中提供了很多[API](https://mxnet.incubator.apache.org/api/python/index.html)。我们把自己编写好的回调函数 add_scalar 插入到相应的 API中 + +``` +lenet_model.fit(train_iter, + eval_data=val_iter, + optimizer='sgd', + optimizer_params={'learning_rate':0.1}, + eval_metric='acc', + # 在此嵌入我们自定义的回调函数 + batch_end_callback=[add_scalar()], + num_epoch=2) +``` + +这样就好了。在MXNet的训练过程中，每一个批次（batch）训练完后，都会调用我们的回调函数来对准确率进行记录。如您所料，随着训练的进行，准确率会不断上升直到95%以上。以下是两个epoch训练过后的准确率走向： + +

diff --git a/demo/mxnet/epoch2_small.png b/demo/mxnet/epoch2_small.png new file mode 100644 index 0000000000000000000000000000000000000000..0779166af4e0467e9131451f01e344129b9e629e Binary files /dev/null and b/demo/mxnet/epoch2_small.png differ diff --git a/demo/mxnet/mxnet_demo.py b/demo/mxnet/mxnet_demo.py new file mode 100644 index 0000000000000000000000000000000000000000..dc35bda8987c748d00ac4f58eed7dd9d0e7cdc77 --- /dev/null +++ b/demo/mxnet/mxnet_demo.py @@ -0,0 +1,93 @@ +import mxnet as mx + +# Here we import LogWriter so that we can write log data while MXNet is training +from visualdl import LogWriter + +# Download MNIST data +mnist = mx.test_utils.get_mnist() +batch_size = 100 + + +# Provide a folder to store data for log, model, image, etc. VisualDL's visualization will be +# based on this folder. +logdir = "./tmp" + +# Initialize a logger instance. Parameter 'sync_cycle' means write a log every 10 operations on +# memory. +logger = LogWriter(logdir, sync_cycle=10) + +# mark the components with 'train' label. +with logger.mode("train"): + # scalar0 is used to record scalar metrics while MXNet is training. We will record accuracy. + # In the visualization, we can see the accuracy is increasing as more training steps happen. + scalar0 = logger.scalar("scalars/scalar0") + +# Record training steps +cnt_step = 0 + + +# MXNet provides many callback interface. Here we define our own callback method and it is called +# after every batch. +# https://mxnet.incubator.apache.org/api/python/callback/callback.html +def add_scalar(): + def _callback(param): + with logger.mode("train"): + global cnt_step + # Here the value is the accuracy we want to record + # https://mxnet.incubator.apache.org/_modules/mxnet/callback.html + name_value = param.eval_metric.get_name_value() + for name, value in name_value: + scalar0.add_record(cnt_step, value) + cnt_step += 1 + return _callback + + +# Start to build CNN in MXNet, train MNIST dataset. For more info, check MXNet's official website: +# https://mxnet.incubator.apache.org/tutorials/python/mnist.html + +import logging +logging.getLogger().setLevel(logging.DEBUG) # logging to stdout + +train_iter = mx.io.NDArrayIter(mnist['train_data'], mnist['train_label'], batch_size, shuffle=True) +val_iter = mx.io.NDArrayIter(mnist['test_data'], mnist['test_label'], batch_size) + +data = mx.sym.var('data') +# first conv layer +conv1 = mx.sym.Convolution(data=data, kernel=(5, 5), num_filter=20) +tanh1 = mx.sym.Activation(data=conv1, act_type="tanh") +pool1 = mx.sym.Pooling(data=tanh1, pool_type="max", kernel=(2, 2), stride=(2, 2)) +# second conv layer +conv2 = mx.sym.Convolution(data=pool1, kernel=(5, 5), num_filter=50) +tanh2 = mx.sym.Activation(data=conv2, act_type="tanh") +pool2 = mx.sym.Pooling(data=tanh2, pool_type="max", kernel=(2, 2), stride=(2, 2)) +# first fullc layer +flatten = mx.sym.flatten(data=pool2) +fc1 = mx.symbol.FullyConnected(data=flatten, num_hidden=500) +tanh3 = mx.sym.Activation(data=fc1, act_type="tanh") +# second fullc +fc2 = mx.sym.FullyConnected(data=tanh3, num_hidden=10) +# softmax loss +lenet = mx.sym.SoftmaxOutput(data=fc2, name='softmax') + +# create a trainable module on CPU +lenet_model = mx.mod.Module(symbol=lenet, context=mx.cpu()) + + +# train with the same +lenet_model.fit(train_iter, + eval_data=val_iter, + optimizer='sgd', + optimizer_params={'learning_rate': 0.1}, + eval_metric='acc', + # integrate our customized callback method + batch_end_callback=[add_scalar()], + num_epoch=2) + +test_iter = mx.io.NDArrayIter(mnist['test_data'], None, batch_size) +prob = lenet_model.predict(test_iter) +test_iter = mx.io.NDArrayIter(mnist['test_data'], mnist['test_label'], batch_size) + +# predict accuracy for lenet +acc = mx.metric.Accuracy() +lenet_model.score(test_iter, acc) +print(acc)