caffe.proto檔案位於 ..\caffe-fast-rcnn\src\caffe\proto檔案目錄下 ，根目錄為 py_faster_rcnn資料夾 syntax = "proto2"; package caffe; //  repeated  required optional  //    類似陣列     必要的      可選的 // Specifies the shape (dimensions) of a Blob. message BlobShape {   repeated int64 dim = 1 [packed = true]; } message BlobProto {   optional BlobShape shape = 7; //下文中替代4D描述符的結構   repeated float data = 5 [packed = true];

  repeated float diff = 6 [packed = true];   repeated double double_data = 8 [packed = true];   repeated double double_diff = 9 [packed = true];   // 4D dimensions -- deprecated.  Use "shape" instead.   // 4維的描述方式捨棄掉，改用"BlobShape"結構替代   optional int32 num = 1 [default = 0];   optional int32 channels = 2 [default = 0];

  optional int32 height = 3 [default = 0];   optional int32 width = 4 [default = 0]; } // The BlobProtoVector is simply a way to pass multiple blobproto instances // around. message BlobProtoVector {   repeated BlobProto blobs = 1; } //影象資料結構 message Datum {   optional int32 channels = 1;   optional int32 height = 2;

  optional int32 width = 3;   // the actual image data, in bytes   //實際上以位元組儲存影象內容   optional bytes data = 4;   optional int32 label = 5;   // Optionally, the datum could also hold float data.   repeated float float_data = 6;   // If true data contains an encoded image that need to be decoded   optional bool encoded = 7 [default = false]; } message FillerParameter {   // The filler type.   optional string type = 1 [default = 'constant'];   optional float value = 2 [default = 0]; // the value in constant filler   optional float min = 3 [default = 0]; // the min value in uniform filler   optional float max = 4 [default = 1]; // the max value in uniform filler   optional float mean = 5 [default = 0]; // the mean value in Gaussian filler   optional float std = 6 [default = 1]; // the std value in Gaussian filler   // The expected number of non-zero output weights for a given input in   // Gaussian filler -- the default -1 means don't perform sparsification.   //非零的輸出值以高斯濾波係數值的方式填充   //預設值為-1，不進行稀疏化   optional int32 sparse = 7 [default = -1];   // Normalize the filler variance by fan_in, fan_out, or their average.   // Applies to 'xavier' and 'msra' fillers.   //Protocol Buffer中的列舉和C++中類似   enum VarianceNorm {     FAN_IN = 0;     FAN_OUT = 1;     AVERAGE = 2;   }   optional VarianceNorm variance_norm = 8 [default = FAN_IN]; } //網路引數 message NetParameter {   optional string name = 1; // consider giving the network a name   // The input blobs to the network.   repeated string input = 3;   // The shape of the input blobs.   repeated BlobShape input_shape = 8;   // 4D input dimensions -- deprecated.  Use "shape" instead.   // If specified, for each input blob there should be four   // values specifying thenum, channels, height and width of the input blob.   // Thus, there should be a total of (4 * #input) numbers.   repeated int32 input_dim = 4;   // Whether the network will force every layer to carry out backward operation.   // If set False, then whether to carry out backward is determined   // automatically according to the net structure and learning rates.   //網路是否會迫使每一層都進行反向操作。   //如果設定為False，則根據網路結構和學習速率自動確定是否執行向後。   optional bool force_backward = 5 [default = false];   // The current "state" of the network, including thephase, level, and stage.   //當前的網路狀態有phase，level，stage三種狀態。   // Some layers may be included/excluded depending on this state and the states   // specified in the layers' include and exclude fields.   //根據此狀態和圖層的包含和非包含欄位中指定的狀態，可以包括/排除某些網路層。   optional NetState state = 6;   // Print debugging information about results while running Net::Forward,   // Net::Backward, and Net::Update.   //當執行前向網路，後向網路，更新網路的時候列印除錯資訊，預設不列印   optional bool debug_info = 7 [default = false];   // The layers that make up the net.  Each of their configurations, including   // connectivity and behavior, is specified as a LayerParameter.   //很多層就構成了網.連線性和行為等配置引數構成了層引數.最後打印出來   repeated LayerParameter layer = 100;  // ID 100 so layers are printed last.   // DEPRECATED: use 'layer' instead.   //此後改用'layer'結構   repeated V1LayerParameter layers = 2; } // NOTE // Update the next available ID when you add a new SolverParameter field. //注意 //當你新增新的求解器引數物件時,更新了新的可用ID ，為 ID 41  type // // SolverParameter next available ID: 41 (last added: type) //求解器引數 message SolverParameter {   //////////////////////////////////////////////////////////////////////////////   // Specifying the train and test networks   //   // Exactly onetrain net must be specified using one of the following fields:   //     train_net_param, train_net, net_param, net   // One or moretest nets may be specified using any of the following fields:   //    test_net_param, test_net, net_param, net   // If more than one test net field is specified (e.g., both net and   // test_net are specified), they will be evaluated in the field order given   // above: (1) test_net_param, (2) test_net, (3) net_param/net.   //如果指定了多個測試網路欄位（例如，指定了net和test_net），則將以上面給出的欄位順序對它們求值：   //（1）test_net_param，（2）test_net，（3）net_param / net。   // A test_iter must be specified for each test_net.   // 必須為每個test_net 指定 test_iter   // A test_level and/or a test_stage may also be specified for each test_net.   // 還可以為每個test_net指定test_level和/或test_stage   //////////////////////////////////////////////////////////////////////////////   // Proto filename for the train net, possibly combined with one or more   // test nets.   //對於訓練網路的原型檔名可能由一個或者多個訓練網路組成。   optional string net = 24;   // Inline train net param, possibly combined with one or more test nets.   // 內聯訓練網路引數可能含有一個或者多個測試網路   optional NetParameter net_param = 25;   optional string train_net = 1; // Proto filename for the train net.   repeated string test_net = 2; // Proto filenames for the test nets.   optional NetParameter train_net_param = 21; // Inline train net params.   repeated NetParameter test_net_param = 22; // Inline test net params.   // The states for the train/test nets. Must be unspecified or   // specified once per net.   //要麼確定，要麼不確定，一旦確定，要麼全是測試網路要麼全是訓練網路   //   // By default, all states will have solver = true;   // train_state will have phase = TRAIN,   // and all test_state's will have phase = TEST.   // Other defaults are set according to the NetState defaults.   //預設的，所有求解器的狀態為真.訓練網路 phase = TRAIN，測試網路phase = TEST，   //其他情況有網路狀態的預設值決定   optional NetState train_state = 26;   repeated NetState test_state = 27;   // The number of iterations for each test net.   repeated int32 test_iter = 3;   // The number of iterations between two testing phases.   // 兩個測試階段之間的迭代次數。   optional int32 test_interval = 4 [default = 0];   optional bool test_compute_loss = 19 [default = false];   // If true, run an initial test pass before the first iteration,   // ensuring memory availability and printing the starting value of the loss.   // 若為真，在執行第一次迭代之前，先得執行初始化測試通過來確保有足夠儲存資源和列印初始值的loss資訊   optional bool test_initialization = 32 [default = true];   optional float base_lr = 5; // The base learning rate //基準學習率   // the number of iterations between displaying info. If display = 0, no info   // will be displayed.   // 顯示迭代之間展示資訊，如果display = 0，則沒有資訊顯示   optional int32 display = 6;   // Display the loss averaged over the last average_loss iterations   // 顯示上次average_loss迭代的平均損失   optional int32 average_loss = 33 [default = 1];   optional int32 max_iter = 7; // the maximum number of iterations   // accumulate gradients over `iter_size` x `batch_size` instances   optional int32 iter_size = 36 [default = 1];   // The learning rate decay policy. The currently implemented learning rate   // policies are as follows:   //    - fixed: always returnbase_lr.   //    - step: returnbase_lr *gamma ^ (floor(iter / step))   //    - exp: returnbase_lr *gamma ^ iter   //    - inv: returnbase_lr * (1 +gamma * iter) ^ (- power)   //    - multistep: similar to step but it allows non uniform steps defined by   //      stepvalue   //    - poly: the effective learning rate follows a polynomial decay, to be   //      zero by the max_iter. returnbase_lr (1 -iter/max_iter) ^ (power)   //    - sigmoid: the effective learning rate follows a sigmod decay   //      returnbase_lr ( 1/(1 + exp(-gamma * (iter - stepsize))))   //   // where base_lr, max_iter, gamma, step, stepvalue and power are defined   // in the solver parameter protocol buffer, and iter is the current iteration.   optional string lr_policy = 8;   optional float gamma = 9; // The parameter to compute the learning rate.   optional float power = 10; // The parameter to compute the learning rate.   optional float momentum = 11; // The momentum value. //動量值？   optional float weight_decay = 12; // The weight decay. //權重衰減   // regularization types supported: L1 and L2   // controlled by weight_decay   //正則化方式支援：L1 和 L2   //由權值衰減變數控制   optional string regularization_type = 29 [default = "L2"]; //預設正則化方式為L2   // the stepsize for learning rate policy "step"   optional int32 stepsize = 13;   // the stepsize for learning rate policy "multistep"   repeated int32 stepvalue = 34;   // Set clip_gradients to >= 0 to clip parameter gradients to that L2 norm,   // whenever their actual L2 norm is larger.   // 設定clip_gradients大於零，只要它比實際的L2範數大，那麼它就等於L2範數    optional float clip_gradients = 35 [default = -1];   optional int32 snapshot = 14 [default = 0]; // The snapshot interval  //snapshot：快照   optional string snapshot_prefix = 15; // The prefix for the snapshot. //prefix：字首   // whether to snapshot diff in the results or not. Snapshotting diff will help   // debugging but the final protocol buffer size will be much larger.   // 無論快照在結果中有無差值，快照的差值將會有助於除錯，但是最終的protocol buffer的尺寸會大很多   //   optional bool snapshot_diff = 16 [default = false];   enum SnapshotFormat {     HDF5 = 0;     BINARYPROTO = 1;   }   optional SnapshotFormat snapshot_format = 37 [default = BINARYPROTO];   // the mode solver will use: 0 for CPU and 1 for GPU. Use GPU in default.   enum SolverMode {     CPU = 0;     GPU = 1;   }   optional SolverMode solver_mode = 17 [default = GPU];   // the device_id will that be used in GPU mode. Use device_id = 0 in default.   optional int32 device_id = 18 [default = 0];   // If non-negative, the seed with which the Solver will initialize the Caffe   // random number generator -- useful for reproducible results. Otherwise,   // (and by default) initialize using a seed derived from the system clock.   optional int64 random_seed = 20 [default = -1];   // type of the solver   //求解器的型別 預設型別為SGD   optional string type = 40 [default = "SGD"]; //string 型別   // numerical stability for RMSProp, AdaGrad and AdaDelta and Adam   // 對於RMSProp, AdaGrad and AdaDelta and Adam的數值穩定性預設閾值為 1e-8   optional float delta = 31 [default = 1e-8];   // parameters for the Adam solver   // 自適應動量求解器的衰減的預設取值為0.999   optional float momentum2 = 39 [default = 0.999];   // RMSProp decay value   // RMSProp的衰減值   // MeanSquare(t) = rms_decay*MeanSquare(t-1) + (1-rms_decay)*SquareGradient(t)   // 均方差的迭代求解關係   // MeanSquare(t) = rms_decay*MeanSquare(t-1) + (1-rms_decay)*SquareGradient(t)   optional float rms_decay = 38;   // If true, print information about the state of the net that may help with   // debugging learning problems.   // 是否列印除錯資訊，預設設定為否   optional bool debug_info = 23 [default = false];   // If false, don't save a snapshot after training finishes.   //如何設定為否，則不儲存每次訓練結束後的快照   optional bool snapshot_after_train = 28 [default = true];   // DEPRECATED: old solver enum types, use string instead   //捨棄舊的求解器列舉型別,使用string代替   enum SolverType {     SGD = 0;     NESTEROV = 1;     ADAGRAD = 2;     RMSPROP = 3;     ADADELTA = 4;     ADAM = 5;   }   // DEPRECATED: use type instead of solver_type   // 捨棄solver_type, 改用 type   optional SolverType solver_type = 30 [default = SGD]; } // A message that stores the solver snapshots message SolverState {   optional int32 iter = 1; // The current iteration //當前迭代   optional string learned_net = 2; // The file that stores the learned net. //儲存學習網路的檔案   repeated BlobProto history = 3; // The history for sgd solvers //sgd求解器的歷史記錄   optional int32 current_step = 4 [default = 0]; // The current step for learning rate //當前學習率的步進 } //狀態列舉：訓練或者測試 enum Phase {     TRAIN = 0;    TEST = 1; } //網路狀態 message NetState {   optional Phase phase = 1 [default = TEST];   optional int32 level = 2 [default = 0];   repeated string stage = 3; } //Rule網路狀態 message NetStateRule {   // Set phase to require the NetState have a particular phase (TRAIN or TEST)   // to meet this rule.   optional Phase phase = 1;   // Set the minimum and/or maximum levels in which the layer should be used.   // Leave undefined to meet the rule regardless of level.   //設定Rule層需使用的最大與/或最小層，其他未定義的層需滿足rule規則。   optional int32 min_level = 2;   optional int32 max_level = 3;   // Customizable sets of stages to include or exclude.   //包含或排除使用者自定義集的狀態   // The net must have ALL of the specified stages and NONE of the specified   // "not_stage"s to meet the rule.   //網路必須含有所有具體的狀態，使用多層網路Rlue用於連線特定狀態   // (Use multiple NetStateRules to specify conjunctions of stages.)   repeated string stage = 4;   repeated string not_stage = 5; } // Specifies training parameters (multipliers on global learning constants, // and the name and other settings used for weight sharing). // 指定訓練引數（多層網路的全域性學習常數，以及用於權重分配的名稱和其他設定）。 message ParamSpec {   // The names of the parameter blobs -- useful for sharing parameters among   // layers, but never required otherwise.  To share a parameter between two   // layers, give it a (non-empty) name.   // blobs引數的名稱-用於在圖層之間共享引數，但從不需要。為了共享一個引數給兩層網路，給它一個名字   optional string name = 1;   // Whether to require shared weights to have the same shape, or just the same   // count -- defaults to STRICT if unspecified.   // 無論是為了相同的shape而共享權值，或者僅僅只是計數。預設情況下如果未指定則為STRICT（限制）   //    optional DimCheckMode share_mode = 2;   enum DimCheckMode {     // STRICT (default) requires thatnum, channels, height, width each match.     //STRICT 限制 (預設)num, channels, height, width為shape的四個引數，對應一一匹配     STRICT = 0;     // PERMISSIVE requires only the count (num*channels*height*width) to match.     // PERMISSIVE 允許 僅需要num*channels*height*width的數相同即可     PERMISSIVE = 1;   }   // The multiplier on the global learning rate for this parameter.   //該引數在全域性上的學習率的乘數   optional float lr_mult = 3 [default = 1.0];   // The multiplier on the global weight decay for this parameter.   // 該引數在全域性上的權值衰減的乘數