diff --git a/.gitignore b/.gitignore index fee9a6ba..f1f2b400 100644 --- a/.gitignore +++ b/.gitignore @@ -53,6 +53,7 @@ examples/* # Generated documentation docs/_site +docs/gathered _site # Sublime Text settings diff --git a/docs/README.md b/docs/README.md index 81e1566b..8f1781e3 100644 --- a/docs/README.md +++ b/docs/README.md @@ -1,3 +1,5 @@ -To generate stuff you can paste in an .md page from an IPython notebook, run +# Caffe Documentation - ipython nbconvert --to markdown +To generate the documentation, run `$CAFFE_ROOT/scripts/build_docs.sh`. + +To push your changes to the documentation to the gh-pages branch of your or the BVLC repo, run `$CAFFE_ROOT/scripts/deploy_docs.sh `. diff --git a/docs/index.md b/docs/index.md index 64ffe8b2..6ffab7a0 100644 --- a/docs/index.md +++ b/docs/index.md @@ -9,14 +9,14 @@ Caffe is released under the [BSD 2-Clause license](https://github.com/BVLC/caffe Check out our web image classification [demo](http://demo.caffe.berkeleyvision.org)! -## Why +## Why use Caffe? **Clean architecture** enables rapid deployment. Networks are specified in simple config files, with no hard-coded parameters in the code. -Switching between CPU and GPU code is as simple as setting a flag -- so models can be trained on a GPU machine, and then used on commodity clusters. +Switching between CPU and GPU is as simple as setting a flag -- so models can be trained on a GPU machine, and then used on commodity clusters. **Readable & modifiable implementation** fosters active development. -In Caffe's first six months, it has been forked by over 300 developers on Github, and many have contributed significant changes. +In Caffe's first six months, it has been forked by over 300 developers on Github, and many have pushed significant changes. **Speed** makes Caffe perfect for industry use. Caffe can process over **40M images per day** with a single NVIDIA K40 or Titan GPU\*. @@ -31,27 +31,34 @@ There is an active discussion and support community on [Github](https://github.c Consult performance [details](/performance_hardware.html).

-## How +## Documentation -* [Introductory slides](http://dl.caffe.berkeleyvision.org/caffe-presentation.pdf): slides about the Caffe architecture, *updated 03/14*. -* [ACM MM paper](http://ucb-icsi-vision-group.github.io/caffe-paper/caffe.pdf): a 4-page report for the ACM Multimedia Open Source competition. -* [Installation instructions](/installation.html): tested on Ubuntu, Red Hat, OS X. -* [Pre-trained models](/getting_pretrained_models.html): BVLC provides ready-to-use models for non-commercial use. -* [Development](/development.html): Guidelines for development and contributing to Caffe. +- [Introductory slides](http://dl.caffe.berkeleyvision.org/caffe-presentation.pdf)
+Slides about the Caffe architecture, *updated 03/14*. +- [ACM MM paper](http://ucb-icsi-vision-group.github.io/caffe-paper/caffe.pdf)
+A 4-page report for the ACM Multimedia Open Source competition. +- [Installation instructions](/installation.html)
+Tested on Ubuntu, Red Hat, OS X. +* [Pre-trained models](/getting_pretrained_models.html)
+BVLC provides ready-to-use models for non-commercial use. +* [Development](/development.html)
+Guidelines for development and contributing to Caffe. -### Tutorials and Examples +### Examples -* [Image Classification \[notebook\]][imagenet_classification]: classify images with the pretrained ImageNet model by the Python interface. -* [Detection \[notebook\]][detection]: run a pretrained model as a detector in Python. -* [Visualizing Features and Filters \[notebook\]][visualizing_filters]: extracting features and visualizing trained filters with an example image, viewed layer-by-layer. -* [LeNet / MNIST Demo](/mnist.html): end-to-end training and testing of LeNet on MNIST. -* [CIFAR-10 Demo](/cifar10.html): training and testing on the CIFAR-10 data. -* [Training ImageNet](/imagenet_training.html): recipe for end-to-end training of an ImageNet classifier. -* [Feature extraction with C++](/feature_extraction.html): feature extraction using pre-trained model. +{% for page in site.pages %} +{% if page.category == 'example' %} +-
{{page.title}}
{{page.description}}
+{% endif %} +{% endfor %} -[imagenet_classification]: http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/examples/imagenet_classification.ipynb -[detection]: http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/examples/detection.ipynb -[visualizing_filters]: http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/examples/filter_visualization.ipynb +### Notebook examples + +{% for page in site.pages %} +{% if page.category == 'notebook' %} +-
{{page.title}}
{{page.description}}
+{% endif %} +{% endfor %} ## Citing Caffe diff --git a/docs/mnist.md b/docs/mnist.md deleted file mode 100644 index 9a9b46a4..00000000 --- a/docs/mnist.md +++ /dev/null @@ -1,91 +0,0 @@ ---- -layout: default -title: Caffe ---- - -Training MNIST with Caffe -================ - -We will assume that you have caffe successfully compiled. If not, please refer to the [Installation page](installation.html). In this tutorial, we will assume that your caffe installation is located at `CAFFE_ROOT`. - -Prepare Datasets ----------------- - -You will first need to download and convert the data format from the MNIST website. To do this, simply run the following commands: - - cd $CAFFE_ROOT/data/mnist - ./get_mnist.sh - cd $CAFFE_ROOT/examples/mnist - ./create_mnist.sh - -If it complains that `wget` or `gunzip` are not installed, you need to install them respectively. After running the script there should be two datasets, `mnist-train-leveldb`, and `mnist-test-leveldb`. - -LeNet: the MNIST Classification Model -------------------------------------- -Before we actually run the training program, let's explain what will happen. We will use the [LeNet](http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf) network, which is known to work well on digit classification tasks. We will use a slightly different version from the original LeNet implementation, replacing the sigmoid activations with Rectified Linear Unit (ReLU) activations for the neurons. - -The design of LeNet contains the essence of CNNs that are still used in larger models such as the ones in ImageNet. In general, it consists of a convolutional layer followed by a pooling layer, another convolution layer followed by a pooling layer, and then two fully connected layers similar to the conventional multilayer perceptrons. We have defined the layers in `CAFFE_ROOT/data/lenet.prototxt`. - -If you would like to read about step-by-step instruction on how the protobuf definitions are written, see [MNIST: Define the Network](mnist_prototxt.html) and [MNIST: Define the Solver](mnist_solver_prototxt.html)?. - -Training and Testing the Model ------------------------------- - -Training the model is simple after you have written the network definition protobuf and solver protobuf files. Simply run `train_mnist.sh`, or the following command directly: - - cd $CAFFE_ROOT/examples/mnist - ./train_lenet.sh - -`train_lenet.sh` is a simple script, but here are a few explanations: `GLOG_logtostderr=1` is the google logging flag that prints all the logging messages directly to stderr. The main tool for training is `train_net.bin`, with the solver protobuf text file as its argument. - -When you run the code, you will see a lot of messages flying by like this: - - I1203 net.cpp:66] Creating Layer conv1 - I1203 net.cpp:76] conv1 <- data - I1203 net.cpp:101] conv1 -> conv1 - I1203 net.cpp:116] Top shape: 20 24 24 - I1203 net.cpp:127] conv1 needs backward computation. - -These messages tell you the details about each layer, its connections and its output shape, which may be helpful in debugging. After the initialization, the training will start: - - I1203 net.cpp:142] Network initialization done. - I1203 solver.cpp:36] Solver scaffolding done. - I1203 solver.cpp:44] Solving LeNet - -Based on the solver setting, we will print the training loss function every 100 iterations, and test the network every 1000 iterations. You will see messages like this: - - I1203 solver.cpp:204] Iteration 100, lr = 0.00992565 - I1203 solver.cpp:66] Iteration 100, loss = 0.26044 - ... - I1203 solver.cpp:84] Testing net - I1203 solver.cpp:111] Test score #0: 0.9785 - I1203 solver.cpp:111] Test score #1: 0.0606671 - -For each training iteration, `lr` is the learning rate of that iteration, and `loss` is the training function. For the output of the testing phase, score 0 is the accuracy, and score 1 is the testing loss function. - -And after a few minutes, you are done! - - I1203 solver.cpp:84] Testing net - I1203 solver.cpp:111] Test score #0: 0.9897 - I1203 solver.cpp:111] Test score #1: 0.0324599 - I1203 solver.cpp:126] Snapshotting to lenet_iter_10000 - I1203 solver.cpp:133] Snapshotting solver state to lenet_iter_10000.solverstate - I1203 solver.cpp:78] Optimization Done. - -The final model, stored as a binary protobuf file, is stored at - - lenet_iter_10000 - -which you can deploy as a trained model in your application, if you are training on a real-world application dataset. - -Um... How about GPU training? ------------------------------ - -You just did! All the training was carried out on the GPU. In fact, if you would like to do training on CPU, you can simply change one line in `lenet_solver.prototxt`: - - # solver mode: CPU or GPU - solver_mode: CPU - -and you will be using CPU for training. Isn't that easy? - -MNIST is a small dataset, so training with GPU does not really introduce too much benefit due to communication overheads. On larger datasets with more complex models, such as ImageNet, the computation speed difference will be more significant. diff --git a/docs/mnist_prototxt.md b/docs/mnist_prototxt.md deleted file mode 100644 index aaff2b00..00000000 --- a/docs/mnist_prototxt.md +++ /dev/null @@ -1,153 +0,0 @@ ---- -layout: default -title: Caffe ---- - -Define the MNIST Network -========================= - -This page explains the prototxt file `lenet_train.prototxt` used in the MNIST demo. We assume that you are familiar with [Google Protobuf](https://developers.google.com/protocol-buffers/docs/overview), and assume that you have read the protobuf definitions used by Caffe, which can be found at [src/caffe/proto/caffe.proto](https://github.com/Yangqing/caffe/blob/master/src/caffe/proto/caffe.proto). - -Specifically, we will write a `caffe::NetParameter` (or in python, `caffe.proto.caffe_pb2.NetParameter`) protubuf. We will start by giving the network a name: - - name: "LeNet" - -Writing the Data Layer ----------------------- -Currently, we will read the MNIST data from the leveldb we created earlier in the demo. This is defined by a data layer: - - layers { - name: "mnist" - type: DATA - data_param { - source: "mnist-train-leveldb" - batch_size: 64 - scale: 0.00390625 - } - top: "data" - top: "label" - } - -Specifically, this layer has name `mnist`, type `data`, and it reads the data from the given leveldb source. We will use a batch size of 64, and scale the incoming pixels so that they are in the range \[0,1\). Why 0.00390625? It is 1 divided by 256. And finally, this layer produces two blobs, one is the `data` blob, and one is the `label` blob. - -Writing the Convolution Layer --------------------------------------------- -Let's define the first convolution layer: - - layers { - name: "conv1" - type: CONVOLUTION - blobs_lr: 1. - blobs_lr: 2. - convolution_param { - num_output: 20 - kernelsize: 5 - stride: 1 - weight_filler { - type: "xavier" - } - bias_filler { - type: "constant" - } - } - bottom: "data" - top: "conv1" - } - -This layer takes the `data` blob (it is provided by the data layer), and produces the `conv1` layer. It produces outputs of 20 channels, with the convolutional kernel size 5 and carried out with stride 1. - -The fillers allow us to randomly initialize the value of the weights and bias. For the weight filler, we will use the `xavier` algorithm that automatically determines the scale of initialization based on the number of input and output neurons. For the bias filler, we will simply initialize it as constant, with the default filling value 0. - -`blobs_lr` are the learning rate adjustments for the layer's learnable parameters. In this case, we will set the weight learning rate to be the same as the learning rate given by the solver during runtime, and the bias learning rate to be twice as large as that - this usually leads to better convergence rates. - -Writing the Pooling Layer -------------------------- -Phew. Pooling layers are actually much easier to define: - - layers { - name: "pool1" - type: POOLING - pooling_param { - kernel_size: 2 - stride: 2 - pool: MAX - } - bottom: "conv1" - top: "pool1" - } - -This says we will perform max pooling with a pool kernel size 2 and a stride of 2 (so no overlapping between neighboring pooling regions). - -Similarly, you can write up the second convolution and pooling layers. Check `data/lenet.prototxt` for details. - -Writing the Fully Connected Layer ----------------------------------- -Writing a fully connected layer is also simple: - - layers { - name: "ip1" - type: INNER_PRODUCT - blobs_lr: 1. - blobs_lr: 2. - inner_product_param { - num_output: 500 - weight_filler { - type: "xavier" - } - bias_filler { - type: "constant" - } - } - bottom: "pool2" - top: "ip1" - } - -This defines a fully connected layer (for some legacy reason, Caffe calls it an `innerproduct` layer) with 500 outputs. All other lines look familiar, right? - -Writing the ReLU Layer ----------------------- -A ReLU Layer is also simple: - - layers { - name: "relu1" - type: RELU - bottom: "ip1" - top: "ip1" - } - -Since ReLU is an element-wise operation, we can do *in-place* operations to save some memory. This is achieved by simply giving the same name to the bottom and top blobs. Of course, do NOT use duplicated blob names for other layer types! - -After the ReLU layer, we will write another innerproduct layer: - - layers { - name: "ip2" - type: INNER_PRODUCT - blobs_lr: 1. - blobs_lr: 2. - inner_product_param { - num_output: 10 - weight_filler { - type: "xavier" - } - bias_filler { - type: "constant" - } - } - bottom: "ip1" - top: "ip2" - } - -Writing the Loss Layer -------------------------- -Finally, we will write the loss! - - layers { - name: "loss" - type: SOFTMAX_LOSS - bottom: "ip2" - bottom: "label" - } - -The `softmax_loss` layer implements both the softmax and the multinomial logistic loss (that saves time and improves numerical stability). It takes two blobs, the first one being the prediction and the second one being the `label` provided by the data layer (remember it?). It does not produce any outputs - all it does is to compute the loss function value, report it when backpropagation starts, and initiates the gradient with respect to `ip2`. This is where all magic starts. - -Now that we have demonstrated how to write the MNIST layer definition prototxt, maybe check out [how we write a solver prototxt](mnist_solver_prototxt.html)? diff --git a/docs/mnist_solver_prototxt.md b/docs/mnist_solver_prototxt.md deleted file mode 100644 index aa3578f1..00000000 --- a/docs/mnist_solver_prototxt.md +++ /dev/null @@ -1,37 +0,0 @@ ---- -layout: default -title: Caffe ---- - -Define the MNIST Solver -======================= - -The page is under construction. For now, check out the comments in the solver prototxt file, which explains each line in the prototxt: - - # The training protocol buffer definition - train_net: "lenet_train.prototxt" - # The testing protocol buffer definition - test_net: "lenet_test.prototxt" - # test_iter specifies how many forward passes the test should carry out. - # In the case of MNIST, we have test batch size 100 and 100 test iterations, - # covering the full 10,000 testing images. - test_iter: 100 - # Carry out testing every 500 training iterations. - test_interval: 500 - # The base learning rate, momentum and the weight decay of the network. - base_lr: 0.01 - momentum: 0.9 - weight_decay: 0.0005 - # The learning rate policy - lr_policy: "inv" - gamma: 0.0001 - power: 0.75 - # Display every 100 iterations - display: 100 - # The maximum number of iterations - max_iter: 10000 - # snapshot intermediate results - snapshot: 5000 - snapshot_prefix: "lenet" - # solver mode: 0 for CPU and 1 for GPU - solver_mode: 1 diff --git a/docs/cifar10.md b/examples/cifar10/readme.md similarity index 97% rename from docs/cifar10.md rename to examples/cifar10/readme.md index dd85667d..9d5bd7b2 100644 --- a/docs/cifar10.md +++ b/examples/cifar10/readme.md @@ -1,6 +1,9 @@ --- +title: CIFAR-10 tutorial +category: example +description: Train and test Caffe on CIFAR-10 data. +include_in_docs: true layout: default -title: Caffe --- Alex's CIFAR-10 tutorial, Caffe style diff --git a/examples/detection.ipynb b/examples/detection.ipynb index feb3e36f..5ec986f6 100644 --- a/examples/detection.ipynb +++ b/examples/detection.ipynb @@ -1,6 +1,8 @@ { "metadata": { - "name": "" + "name": "ImageNet detection", + "description": "Run a pretrained model as a detector in Python.", + "include_in_docs": true }, "nbformat": 3, "nbformat_minor": 0, @@ -652,4 +654,4 @@ "metadata": {} } ] -} \ No newline at end of file +} diff --git a/docs/feature_extraction.md b/examples/feature_extraction/readme.md similarity index 93% rename from docs/feature_extraction.md rename to examples/feature_extraction/readme.md index fa23e9c8..c336e718 100644 --- a/docs/feature_extraction.md +++ b/examples/feature_extraction/readme.md @@ -1,6 +1,9 @@ --- +title: Feature extraction with Caffe C++ code. +description: Extract AlexNet features using the Caffe binary. +category: example +include_in_docs: true layout: default -title: Caffe --- Extracting Features @@ -57,7 +60,7 @@ The last parameter above is the number of data mini-batches. The features are stored to LevelDB `examples/_temp/features`, ready for access by some other code. -If you meet with the error "Check failed: status.ok() Failed to open leveldb examples/_temp/features", it is because the directory examples/_temp/features has been created the last time you run the command. Remove it and run again. +If you meet with the error "Check failed: status.ok() Failed to open leveldb examples/_temp/features", it is because the directory examples/_temp/features has been created the last time you run the command. Remove it and run again. rm -rf examples/_temp/features/ diff --git a/examples/filter_visualization.ipynb b/examples/filter_visualization.ipynb index 6f494821..22d7fc21 100644 --- a/examples/filter_visualization.ipynb +++ b/examples/filter_visualization.ipynb @@ -1,6 +1,8 @@ { "metadata": { - "name": "" + "name": "Filter visualization", + "description": "Extracting features and visualizing trained filters with an example image, viewed layer-by-layer.", + "include_in_docs": true }, "nbformat": 3, "nbformat_minor": 0, @@ -611,4 +613,4 @@ "metadata": {} } ] -} \ No newline at end of file +} diff --git a/docs/imagenet_training.md b/examples/imagenet/readme.md similarity index 98% rename from docs/imagenet_training.md rename to examples/imagenet/readme.md index 9e0076cf..55348727 100644 --- a/docs/imagenet_training.md +++ b/examples/imagenet/readme.md @@ -1,6 +1,9 @@ --- +title: ImageNet tutorial +description: Train and test "AlexNet" on ImageNet challenge data. +category: example +include_in_docs: true layout: default -title: Caffe --- Yangqing's Recipe on Brewing ImageNet diff --git a/examples/imagenet_classification.ipynb b/examples/imagenet_classification.ipynb index 0e0e06bb..7ac140d9 100644 --- a/examples/imagenet_classification.ipynb +++ b/examples/imagenet_classification.ipynb @@ -1,6 +1,8 @@ { "metadata": { - "name": "" + "description": "Use the pre-trained ImageNet model to classify images with the Python interface.", + "name": "ImageNet Classification", + "include_in_docs": true }, "nbformat": 3, "nbformat_minor": 0, @@ -407,4 +409,4 @@ "metadata": {} } ] -} \ No newline at end of file +} diff --git a/examples/mnist/readme.md b/examples/mnist/readme.md new file mode 100644 index 00000000..d609cfff --- /dev/null +++ b/examples/mnist/readme.md @@ -0,0 +1,266 @@ +--- +title: MNIST Tutorial +description: Train and test "LeNet" on MNIST data. +category: example +include_in_docs: true +layout: default +--- + +# Training MNIST with Caffe + +We will assume that you have caffe successfully compiled. If not, please refer to the [Installation page](installation.html). In this tutorial, we will assume that your caffe installation is located at `CAFFE_ROOT`. + +## Prepare Datasets + +You will first need to download and convert the data format from the MNIST website. To do this, simply run the following commands: + + cd $CAFFE_ROOT/data/mnist + ./get_mnist.sh + cd $CAFFE_ROOT/examples/mnist + ./create_mnist.sh + +If it complains that `wget` or `gunzip` are not installed, you need to install them respectively. After running the script there should be two datasets, `mnist-train-leveldb`, and `mnist-test-leveldb`. + +## LeNet: the MNIST Classification Model + +Before we actually run the training program, let's explain what will happen. We will use the [LeNet](http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf) network, which is known to work well on digit classification tasks. We will use a slightly different version from the original LeNet implementation, replacing the sigmoid activations with Rectified Linear Unit (ReLU) activations for the neurons. + +The design of LeNet contains the essence of CNNs that are still used in larger models such as the ones in ImageNet. In general, it consists of a convolutional layer followed by a pooling layer, another convolution layer followed by a pooling layer, and then two fully connected layers similar to the conventional multilayer perceptrons. We have defined the layers in `CAFFE_ROOT/data/lenet.prototxt`. + +## Define the MNIST Network + +This section explains the prototxt file `lenet_train.prototxt` used in the MNIST demo. We assume that you are familiar with [Google Protobuf](https://developers.google.com/protocol-buffers/docs/overview), and assume that you have read the protobuf definitions used by Caffe, which can be found at [src/caffe/proto/caffe.proto](https://github.com/Yangqing/caffe/blob/master/src/caffe/proto/caffe.proto). + +Specifically, we will write a `caffe::NetParameter` (or in python, `caffe.proto.caffe_pb2.NetParameter`) protubuf. We will start by giving the network a name: + + name: "LeNet" + +### Writing the Data Layer + +Currently, we will read the MNIST data from the leveldb we created earlier in the demo. This is defined by a data layer: + + layers { + name: "mnist" + type: DATA + data_param { + source: "mnist-train-leveldb" + batch_size: 64 + scale: 0.00390625 + } + top: "data" + top: "label" + } + +Specifically, this layer has name `mnist`, type `data`, and it reads the data from the given leveldb source. We will use a batch size of 64, and scale the incoming pixels so that they are in the range \[0,1\). Why 0.00390625? It is 1 divided by 256. And finally, this layer produces two blobs, one is the `data` blob, and one is the `label` blob. + +### Writing the Convolution Layer + +Let's define the first convolution layer: + + layers { + name: "conv1" + type: CONVOLUTION + blobs_lr: 1. + blobs_lr: 2. + convolution_param { + num_output: 20 + kernelsize: 5 + stride: 1 + weight_filler { + type: "xavier" + } + bias_filler { + type: "constant" + } + } + bottom: "data" + top: "conv1" + } + +This layer takes the `data` blob (it is provided by the data layer), and produces the `conv1` layer. It produces outputs of 20 channels, with the convolutional kernel size 5 and carried out with stride 1. + +The fillers allow us to randomly initialize the value of the weights and bias. For the weight filler, we will use the `xavier` algorithm that automatically determines the scale of initialization based on the number of input and output neurons. For the bias filler, we will simply initialize it as constant, with the default filling value 0. + +`blobs_lr` are the learning rate adjustments for the layer's learnable parameters. In this case, we will set the weight learning rate to be the same as the learning rate given by the solver during runtime, and the bias learning rate to be twice as large as that - this usually leads to better convergence rates. + +### Writing the Pooling Layer + +Phew. Pooling layers are actually much easier to define: + + layers { + name: "pool1" + type: POOLING + pooling_param { + kernel_size: 2 + stride: 2 + pool: MAX + } + bottom: "conv1" + top: "pool1" + } + +This says we will perform max pooling with a pool kernel size 2 and a stride of 2 (so no overlapping between neighboring pooling regions). + +Similarly, you can write up the second convolution and pooling layers. Check `data/lenet.prototxt` for details. + +### Writing the Fully Connected Layer + +Writing a fully connected layer is also simple: + + layers { + name: "ip1" + type: INNER_PRODUCT + blobs_lr: 1. + blobs_lr: 2. + inner_product_param { + num_output: 500 + weight_filler { + type: "xavier" + } + bias_filler { + type: "constant" + } + } + bottom: "pool2" + top: "ip1" + } + +This defines a fully connected layer (for some legacy reason, Caffe calls it an `innerproduct` layer) with 500 outputs. All other lines look familiar, right? + +### Writing the ReLU Layer + +A ReLU Layer is also simple: + + layers { + name: "relu1" + type: RELU + bottom: "ip1" + top: "ip1" + } + +Since ReLU is an element-wise operation, we can do *in-place* operations to save some memory. This is achieved by simply giving the same name to the bottom and top blobs. Of course, do NOT use duplicated blob names for other layer types! + +After the ReLU layer, we will write another innerproduct layer: + + layers { + name: "ip2" + type: INNER_PRODUCT + blobs_lr: 1. + blobs_lr: 2. + inner_product_param { + num_output: 10 + weight_filler { + type: "xavier" + } + bias_filler { + type: "constant" + } + } + bottom: "ip1" + top: "ip2" + } + +### Writing the Loss Layer + +Finally, we will write the loss! + + layers { + name: "loss" + type: SOFTMAX_LOSS + bottom: "ip2" + bottom: "label" + } + +The `softmax_loss` layer implements both the softmax and the multinomial logistic loss (that saves time and improves numerical stability). It takes two blobs, the first one being the prediction and the second one being the `label` provided by the data layer (remember it?). It does not produce any outputs - all it does is to compute the loss function value, report it when backpropagation starts, and initiates the gradient with respect to `ip2`. This is where all magic starts. + +## Define the MNIST Solver + +Check out the comments explaining each line in the prototxt: + + # The training protocol buffer definition + train_net: "lenet_train.prototxt" + # The testing protocol buffer definition + test_net: "lenet_test.prototxt" + # test_iter specifies how many forward passes the test should carry out. + # In the case of MNIST, we have test batch size 100 and 100 test iterations, + # covering the full 10,000 testing images. + test_iter: 100 + # Carry out testing every 500 training iterations. + test_interval: 500 + # The base learning rate, momentum and the weight decay of the network. + base_lr: 0.01 + momentum: 0.9 + weight_decay: 0.0005 + # The learning rate policy + lr_policy: "inv" + gamma: 0.0001 + power: 0.75 + # Display every 100 iterations + display: 100 + # The maximum number of iterations + max_iter: 10000 + # snapshot intermediate results + snapshot: 5000 + snapshot_prefix: "lenet" + # solver mode: 0 for CPU and 1 for GPU + solver_mode: 1 + +## Training and Testing the Model + +Training the model is simple after you have written the network definition protobuf and solver protobuf files. Simply run `train_mnist.sh`, or the following command directly: + + cd $CAFFE_ROOT/examples/mnist + ./train_lenet.sh + +`train_lenet.sh` is a simple script, but here are a few explanations: `GLOG_logtostderr=1` is the google logging flag that prints all the logging messages directly to stderr. The main tool for training is `train_net.bin`, with the solver protobuf text file as its argument. + +When you run the code, you will see a lot of messages flying by like this: + + I1203 net.cpp:66] Creating Layer conv1 + I1203 net.cpp:76] conv1 <- data + I1203 net.cpp:101] conv1 -> conv1 + I1203 net.cpp:116] Top shape: 20 24 24 + I1203 net.cpp:127] conv1 needs backward computation. + +These messages tell you the details about each layer, its connections and its output shape, which may be helpful in debugging. After the initialization, the training will start: + + I1203 net.cpp:142] Network initialization done. + I1203 solver.cpp:36] Solver scaffolding done. + I1203 solver.cpp:44] Solving LeNet + +Based on the solver setting, we will print the training loss function every 100 iterations, and test the network every 1000 iterations. You will see messages like this: + + I1203 solver.cpp:204] Iteration 100, lr = 0.00992565 + I1203 solver.cpp:66] Iteration 100, loss = 0.26044 + ... + I1203 solver.cpp:84] Testing net + I1203 solver.cpp:111] Test score #0: 0.9785 + I1203 solver.cpp:111] Test score #1: 0.0606671 + +For each training iteration, `lr` is the learning rate of that iteration, and `loss` is the training function. For the output of the testing phase, score 0 is the accuracy, and score 1 is the testing loss function. + +And after a few minutes, you are done! + + I1203 solver.cpp:84] Testing net + I1203 solver.cpp:111] Test score #0: 0.9897 + I1203 solver.cpp:111] Test score #1: 0.0324599 + I1203 solver.cpp:126] Snapshotting to lenet_iter_10000 + I1203 solver.cpp:133] Snapshotting solver state to lenet_iter_10000.solverstate + I1203 solver.cpp:78] Optimization Done. + +The final model, stored as a binary protobuf file, is stored at + + lenet_iter_10000 + +which you can deploy as a trained model in your application, if you are training on a real-world application dataset. + +### Um... How about GPU training? + +You just did! All the training was carried out on the GPU. In fact, if you would like to do training on CPU, you can simply change one line in `lenet_solver.prototxt`: + + # solver mode: CPU or GPU + solver_mode: CPU + +and you will be using CPU for training. Isn't that easy? + +MNIST is a small dataset, so training with GPU does not really introduce too much benefit due to communication overheads. On larger datasets with more complex models, such as ImageNet, the computation speed difference will be more significant. diff --git a/scripts/build_docs.sh b/scripts/build_docs.sh index 1faf8bd5..c5680151 100755 --- a/scripts/build_docs.sh +++ b/scripts/build_docs.sh @@ -1,11 +1,17 @@ #!/bin/bash +# Build documentation for display in web browser. PORT=${1:-4000} -echo "usage: build_docs.sh [port]" +echo "usage: build.sh [port]" # Find the docs dir, no matter where the script is called -DIR="$( cd "$(dirname "$0")" ; pwd -P )" -cd $DIR/../docs +ROOT_DIR="$( cd "$(dirname "$0")"/.. ; pwd -P )" +cd $ROOT_DIR +# Gather docs. +scripts/gather_examples.sh + +# Display docs using web server. +cd docs jekyll serve -w -s . -d _site --port=$PORT diff --git a/scripts/copy_notebook.py b/scripts/copy_notebook.py new file mode 100755 index 00000000..4f60cfd0 --- /dev/null +++ b/scripts/copy_notebook.py @@ -0,0 +1,32 @@ +#!/usr/bin/env python +""" +Takes as arguments: +1. the path to a JSON file (such as an IPython notebook). +2. the path to output file + +If 'metadata' dict in the JSON file contains 'include_in_docs': true, +then copies the file to output file, appending the 'metadata' property +as YAML front-matter, adding the field 'category' with value 'notebook'. +""" +import os +import sys +import json + +filename = sys.argv[1] +output_filename = sys.argv[2] +content = json.load(open(filename)) + +if 'include_in_docs' in content['metadata'] and content['metadata']['include_in_docs']: + yaml_frontmatter = ['---'] + for key, val in content['metadata'].iteritems(): + if key == 'name': + key = 'title' + if val == '': + val = os.path.basename(filename) + yaml_frontmatter.append('{}: {}'.format(key, val)) + yaml_frontmatter += ['category: notebook'] + yaml_frontmatter += ['original_path: ' + filename] + + with open(output_filename, 'w') as fo: + fo.write('\n'.join(yaml_frontmatter + ['---']) + '\n') + fo.write(open(filename).read()) diff --git a/scripts/deploy_docs.sh b/scripts/deploy_docs.sh index b6029613..fdf97f71 100755 --- a/scripts/deploy_docs.sh +++ b/scripts/deploy_docs.sh @@ -1,5 +1,5 @@ -#!/usr/bin/env sh -# Publish/ Pull-request documentation to the gh-pages site. +#!/bin/bash +# Publish documentation to the gh-pages site. # The remote for pushing the docs (defaults to origin). # This is where you will submit the PR to BVLC:gh-pages from. diff --git a/scripts/gather_examples.sh b/scripts/gather_examples.sh new file mode 100755 index 00000000..f5192375 --- /dev/null +++ b/scripts/gather_examples.sh @@ -0,0 +1,28 @@ +#!/bin/bash +# Assemble documentation for the project into one directory via symbolic links. + +# Find the docs dir, no matter where the script is called +ROOT_DIR="$( cd "$(dirname "$0")"/.. ; pwd -P )" +cd $ROOT_DIR + +# Gather docs from examples/**/readme.md +rm -r docs/gathered +mkdir docs/gathered +for README_FILENAME in $(find examples -iname "readme.md"); do + # Only use file if it is to be included in docs. + if grep -Fxq "include_in_docs: true" $README_FILENAME; then + # Make link to readme.md in docs/gathered/. + # Since everything is called readme.md, rename it by its dirname. + README_DIRNAME=`dirname $README_FILENAME` + DOCS_FILENAME=docs/gathered/$README_DIRNAME.md + mkdir -p `dirname $DOCS_FILENAME` + ln -s $ROOT_DIR/$README_FILENAME $DOCS_FILENAME + fi +done + +# Gather docs from examples/*.ipynb and add YAML front-matter. +for NOTEBOOK_FILENAME in $(find examples -d 1 -iname "*.ipynb"); do + DOCS_FILENAME=docs/gathered/$NOTEBOOK_FILENAME + mkdir -p `dirname $DOCS_FILENAME` + python scripts/copy_notebook.py $NOTEBOOK_FILENAME $DOCS_FILENAME +done