Blob是一個四維連續數組(4-D contiguous array, type = float32), 使用(n, k, h, w)表示的話，那么每一維的意思分別是：

• n: number. 輸入數據量，比如進行sgd時候的mini-batch大小。
• c: channel. 如果是圖像數據的話可以認為是通道數量。
• h,w: height, width. 如果是圖像數據的話可以認為是圖片的高度和寬度。

當然Blob不一定就是用來表示圖像輸入數據。理解這些維度最重要的一點是，下標w是變化最快的。主頁里面舉了幾個例子：

• the shape of blob holding 1000 vectors of 16 feature dimensions is 1000 x 16 x 1 x 1.
• For a convolution layer with 96 filters of 11 x 11 spatial dimension and 3 inputs the blob is 96 x 3 x 11 x 11.
• For an inner product / fully-connected layer with 1000 output channels and 1024 input channels the parameter blob is 1 x 1 x 1000 x 1024.

Blob內部其實有兩個字段data, diff. data表示流動數據(輸出數據)，而diff則存儲BP的梯度。data/diff可以存儲于cpu, 也可以存儲于gpu. 如果某個layer不支持gpu的話，那么就需要將gpu數據copy到cpu上，造成性能開銷。對于python/numpy用戶來說，可以用reshape函數來轉換為blob: data = data.reshape((-1, c, h, w))

caffe提供了許多內置layer，比如convolution layer, pool layer, dropout layer, nonlinearity layer等。這些層說明以及具體參數都可以在?這里?查到（文檔比代碼有一些滯后，文檔里面沒有說支持了dropout但是實際已經提供）。每個layer有輸入一些'bottom' blobs, 輸出一些'top' blobs. 輸入層是"data"和"label" blobs.

Each layer type defines three critical computations: setup, forward, and backward.

• Setup: initialize the layer and its connections once at model initialization. # 初始化工作
• Forward: given input from bottom compute the output and send to the top. # 前向轉播
• Backward: given the gradient w.r.t. the top output compute the gradient w.r.t. to the input and send to the bottom. A layer with parameters computes the gradient w.r.t. to its parameters and stores it internally. # 反向轉播/計算梯度

caffe支持的layer完整在?http://caffe.berkeleyvision.org/tutorial/layers.html, 部分data layer還支持?預處理?操作

#note: 有可能文檔上名字和實際代碼對不上，如果是這樣的話可以閱讀src/caffe/layers/*_layer.cpp找到REGISTER_LAYER_CLASS(name). 其中name就是注冊的字符串

net是layers組成的DAG, 并且可以使用文本格式來描述(protocol-buffers TextFormat). 比如下面文本生成的是logistic regression.

name: "LogReg"
layers {name: "mnist"type: DATAtop: "data"top: "label"data_param {source: "input_leveldb"batch_size: 64}
}
layers {name: "ip"type: INNER_PRODUCTbottom: "data"top: "ip"inner_product_param {num_output: 2}
}
layers {name: "loss"type: SOFTMAX_LOSSbottom: "ip"bottom: "label"top: "loss"
}

Net有個初始化函數Init(). 它的作用有兩個：1. 創建blosb和layers; 2. 調用layers的SetUp函數來初始化layers. 在這個過程中會打印日志來說明。注意在這個階段并沒有指明說是用GPU還是CPU來訓練，指定使用什么訓練是在solver層面的事情，這樣可以將模型和實現分離。Net還有Forward和Backward兩個函數，分別調用各個Layers的forward/backward. 最周如果我們進行預測的話，我們先填充好input blobs, 然后調用forward函數，最后獲取output blobs作為預測結果。

I0902 22:52:17.931977 2079114000 net.cpp:39] Initializing net from parameters:
name: "LogReg"
[...model prototxt printout...]
# construct the network layer-by-layer
I0902 22:52:17.932152 2079114000 net.cpp:67] Creating Layer mnist
I0902 22:52:17.932165 2079114000 net.cpp:356] mnist -> data
I0902 22:52:17.932188 2079114000 net.cpp:356] mnist -> label
I0902 22:52:17.932200 2079114000 net.cpp:96] Setting up mnist
I0902 22:52:17.935807 2079114000 data_layer.cpp:135] Opening leveldb input_leveldb
I0902 22:52:17.937155 2079114000 data_layer.cpp:195] output data size: 64,1,28,28
I0902 22:52:17.938570 2079114000 net.cpp:103] Top shape: 64 1 28 28 (50176)
I0902 22:52:17.938593 2079114000 net.cpp:103] Top shape: 64 1 1 1 (64)
I0902 22:52:17.938611 2079114000 net.cpp:67] Creating Layer ip
I0902 22:52:17.938617 2079114000 net.cpp:394] ip <- data
I0902 22:52:17.939177 2079114000 net.cpp:356] ip -> ip
I0902 22:52:17.939196 2079114000 net.cpp:96] Setting up ip
I0902 22:52:17.940289 2079114000 net.cpp:103] Top shape: 64 2 1 1 (128)
I0902 22:52:17.941270 2079114000 net.cpp:67] Creating Layer loss
I0902 22:52:17.941305 2079114000 net.cpp:394] loss <- ip
I0902 22:52:17.941314 2079114000 net.cpp:394] loss <- label
I0902 22:52:17.941323 2079114000 net.cpp:356] loss -> loss
# set up the loss and configure the backward pass
I0902 22:52:17.941328 2079114000 net.cpp:96] Setting up loss
I0902 22:52:17.941328 2079114000 net.cpp:103] Top shape: 1 1 1 1 (1)
I0902 22:52:17.941329 2079114000 net.cpp:109]     with loss weight 1
I0902 22:52:17.941779 2079114000 net.cpp:170] loss needs backward computation.
I0902 22:52:17.941787 2079114000 net.cpp:170] ip needs backward computation.
I0902 22:52:17.941794 2079114000 net.cpp:172] mnist does not need backward computation.
# determine outputs
I0902 22:52:17.941800 2079114000 net.cpp:208] This network produces output loss
# finish initialization and report memory usage
I0902 22:52:17.941810 2079114000 net.cpp:467] Collecting Learning Rate and Weight Decay.
I0902 22:52:17.941818 2079114000 net.cpp:219] Network initialization done.
I0902 22:52:17.941824 2079114000 net.cpp:220] Memory required for data: 201476

如果閱讀caffe/models會發現，這些例子下面有train.prototxt，還有一個deploy.prototxt. 差別僅僅在于deploy.txt沒有data-layer，而是在指定輸入的shape.

input: "data"
input_dim: 10
input_dim: 1
input_dim: 28
input_dim: 28

從字面上來看train.prototxt是用來訓練出model的，而deploy.prototxt則是用來進行預測的。下面是使用python進行預測的代碼:

#note: 我沒有使用caffe自身提供的classifier.py, 因為我發現Classifier會對input做一些處理。在進行實驗的時候我發現使用Classifier得到的結果比直接使用Net::forward_all接口要差很多。

caffe.set_mode_cpu()
net = caffe.Net('caffe-conf/test.prototxt','uv_iter_10000.caffemodel',caffe.TEST)
data = data.reshape((-1, 1, 28, 28))
out = net.forward_all(**{'data': data})
rs = out['prob'] # 得到的是softmax.
print_timer(<span class="org-string">"predict"</span>)

solver做了下面這些事情：

• scaffolds the optimization bookkeeping and creates the training network for learning and test network(s) for evaluation.
• iteratively optimizes by calling forward / backward and updating parameters # Solver::ComputeUpdateValue()
• (periodically) evaluates the test networks
• snapshots the model and solver state throughout the optimization
  • Solver::Snapshot() / Solver::Restore() # 保存和恢復網絡參數, 后綴.caffemodel
  • Solver::SnapshotSolverState() / Solver::RestoreSolverState() # 保存和恢復運行狀態，后綴.solverstate
  • 文件名稱是<prefix>_iter_<N>，其中prefix是指定前綴，N表示迭代輪數。

solver每輪迭代做了下面這些事情：

• calls network forward to compute the output and loss
• calls network backward to compute the gradients
  • Stochastic Gradient Descent (SGD),
  • Adaptive Gradient (ADAGRAD),
  • and Nesterov’s Accelerated Gradient (NESTEROV).
  • 如何選擇和設置參數可以看?這里
• incorporates the gradients into parameter updates according to the solver method
• updates the solver state according to learning rate, history, and method

下面是solver.prototxt的一個示例（從examples/mnist/修改過來的）

# The train/test net protocol buffer definition
net: "caffe-conf/train.prototxt"# 如果test數據量是10000，而bacth_size = 100的話，那么test_iter就應該設置100
# 這樣每次進行test就可以把所有的cases都使用上了
test_iter: 90
# Carry out testing every 500 training iterations.
# 每進行500輪迭代進行一次測試
test_interval: 500# 下面這些是訓練使用參數
# 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: 500
# The maximum number of iterations
max_iter: 10000
# snapshot intermediate results
# 每進行500輪做一次snapshot.
# 每一輪使用的數據量大小為batch_size.
snapshot: 500
snapshot_prefix: "uv"
snapshot_after_train: true
# solver mode: CPU or GPU
# 使用CPU訓練
solver_mode: CPU

"net"表示train和test使用同一個net. 在net.prototxt中可以使用include語法來聲明說，某個layer是否需要包含在train/test階段.

如果你在訓練時候不想進行test的話，那么可以指定上面的"net"為"train_net". 當然你也可以使用"test_nets"來指定多個test_net.