/*- * * * Copyright 2016 Skymind,Inc. * * * * Licensed under the Apache License, Version 2.0 (the "License"); * * you may not use this file except in compliance with the License. * * You may obtain a copy of the License at * * * * http://www.apache.org/licenses/LICENSE-2.0 * * * * Unless required by applicable law or agreed to in writing, software * * distributed under the License is distributed on an "AS IS" BASIS, * * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * * See the License for the specific language governing permissions and * * limitations under the License. * */ package org.deeplearning4j.nn.graph.vertex; import org.deeplearning4j.berkeley.Pair; import org.deeplearning4j.nn.api.Layer; import org.deeplearning4j.nn.api.MaskState; import org.deeplearning4j.nn.gradient.Gradient; import org.nd4j.linalg.api.ndarray.INDArray; import java.io.Serializable; /** A GraphVertex is a vertex in the computation graph. It may contain Layer, or define some arbitrary forward/backward pass * behaviour based on the inputs.<br> * The purposes of GraphVertex instances are as follows: * 1. To track the (local) network connection structure: i.e., a GraphVertex knows about the vertices on the input and output sides * 2. To store intermediate results (activations and epsilons) * 3. To allow forward pass and backward pass to be conducted, once the intermediate results are set * @author Alex Black */ public interface GraphVertex extends Serializable { /** Get the name/label of the GraphVertex */ String getVertexName(); /** Get the index of the GraphVertex */ int getVertexIndex(); /** Get the number of input arrays. For example, a Layer may have only one input array, but in general a GraphVertex * may have an arbtrary (>=1) number of input arrays (for example, from multiple other layers) */ int getNumInputArrays(); /** Get the number of outgoing connections from this GraphVertex. A GraphVertex may only have a single output (for * example, the activations out of a layer), but this output may be used as the input to an arbitrary number of other * GraphVertex instances. This method returns the number of GraphVertex instances the output of this GraphVertex is input for. */ int getNumOutputConnections(); /** A representation of the vertices that are inputs to this vertex (inputs duing forward pass)<br> * Specifically, if inputVertices[X].getVertexIndex() = Y, and inputVertices[X].getVertexEdgeNumber() = Z * then the Zth output connection (see {@link #getNumOutputConnections()} of vertex Y is the Xth input to this vertex */ VertexIndices[] getInputVertices(); /** Sets the input vertices. * @see #getInputVertices() */ void setInputVertices(VertexIndices[] inputVertices); /** A representation of the vertices that this vertex is connected to (outputs duing forward pass) * Specifically, if outputVertices[X].getVertexIndex() = Y, and outputVertices[X].getVertexEdgeNumber() = Z * then the Xth output of this vertex is connected to the Zth input of vertex Y */ VertexIndices[] getOutputVertices(); /** set the output vertices. * @see #getOutputVertices() */ void setOutputVertices(VertexIndices[] outputVertices); /** Whether the GraphVertex contains a {@link Layer} object or not */ boolean hasLayer(); /** Whether the GraphVertex is an input vertex */ boolean isInputVertex(); /** Whether the GraphVertex is an output vertex */ boolean isOutputVertex(); /** Get the Layer (if any). Returns null if {@link #hasLayer()} == false */ Layer getLayer(); /** Set the input activations. * * @param inputNumber Must be in range 0 to {@link #getNumInputArrays()}-1 * @param input The input array */ void setInput(int inputNumber, INDArray input); /** Set the errors (epsilon - aka dL/dActivation) for this GraphVertex */ void setEpsilon(INDArray epsilon); /** Clear the internal state (if any) of the GraphVertex. For example, any stored inputs/errors */ void clear(); /** Whether the GraphVertex can do forward pass. Typically, this is just whether all inputs are set. */ boolean canDoForward(); /** Whether the GraphVertex can do backward pass. Typically, this is just whether all errors/epsilons are set */ boolean canDoBackward(); /** Do forward pass using the stored inputs * @param training if true: forward pass at training time. If false: forward pass at test time * @return The output (for example, activations) of the GraphVertex */ INDArray doForward(boolean training); /** Do backward pass * @param tbptt If true: do backprop using truncated BPTT * @return The gradients (may be null), and the errors/epsilons for all inputs to this GraphVertex */ Pair<Gradient, INDArray[]> doBackward(boolean tbptt); /** Get the array of inputs previously set for this GraphVertex */ INDArray[] getInputs(); /** Get the epsilon/error (i.e., dL/dOutput) array previously set for this GraphVertex */ INDArray getEpsilon(); /** Set all inputs for this GraphVertex * @see #setInput(int, INDArray) */ void setInputs(INDArray... inputs); /** * See {@link Layer#setBackpropGradientsViewArray(INDArray)} * @param backpropGradientsViewArray */ void setBackpropGradientsViewArray(INDArray backpropGradientsViewArray); Pair<INDArray, MaskState> feedForwardMaskArrays(INDArray[] maskArrays, MaskState currentMaskState, int minibatchSize); /** * Only applies to layer vertices. Will throw exceptions on others. * If applied to a layer vertex it will treat the parameters of the layer within it as constant. * Activations through these will be calculated as they would as test time regardless of training mode */ void setLayerAsFrozen(); /** * This method clears inpjut for this vertex */ void clearVertex(); }