Exemplo n.º 1
0
func main() {
	flag.Parse()
	if flag.NArg() != 3 {
		flag.Usage()
		os.Exit(1)
	}

	vecs := common.ReadEmbeddingsOrExit(common.Embedding{flag.Arg(0), false, false})
	network := common.ReadModelOrExit(flag.Arg(1), cblas.Implementation{})

	out, err := os.Create(flag.Arg(2))
	common.ExitIfError("Cannot open output vectors for writing: ", err)
	defer out.Close()

	if network.Layers() != 1 {
		fmt.Fprintf(os.Stderr, "Weight file contains %d layers, expected 1", network.Layers())
		os.Exit(1)
	}

	layer := network.Layer(0)
	weights := layer.W()

	if layer.Inputs() != uint(vecs.Size()) {
		fmt.Fprintf(os.Stderr, "Embedding layer and one-hot size mismatch: %d - %d", layer.Inputs(), vecs.Size())
		os.Exit(1)
	}

	mergedVecs := go2vec.NewEmbeddings(int(layer.Outputs()))

	wordIdx := 0
	vec := make([]float32, layer.Outputs())
	vecs.Iterate(func(word string, vector []float32) bool {
		for idx := range vec {
			vec[idx] = weights[uint(idx)*layer.Inputs()+uint(wordIdx)]
		}

		mergedVecs.Put(word, vec)
		wordIdx++
		return true
	})

	writer := bufio.NewWriter(out)
	mergedVecs.Write(writer)
	writer.Flush()
}
Exemplo n.º 2
0
func main() {
	flag.Parse()
	if flag.NArg() != 1 {
		flag.Usage()
		os.Exit(1)
	}

	config := common.ReadConfigOrExit(flag.Arg(0))

	transitionSystem, ok := common.TransitionSystems[config.Parser.System]
	if !ok {
		log.Fatalf("Unknown transition system: %s", config.Parser.System)
	}

	labelNumberer := common.ReadTransitionsOrExit(config.Parser.Transitions, transitionSystem)

	var layers []*caffe.LayerParameter
	layers = []*caffe.LayerParameter{
		createDataLayer("test/train-lmdb", caffe.Phase_TRAIN, caffe.DataParameter_LMDB),
		createDataLayer("test/validation-lmdb", caffe.Phase_TEST, caffe.DataParameter_LMDB),
		createDropoutLayer("data_dropout", "data", "data", 0.1),
	}

	ilas := common.ReadIlasOrExit(config.Parser.Inputs)
	vecs := common.ReadEmbeddingsOrExit(config.Embeddings.Word)
	tvecs := common.ReadEmbeddingsOrExit(config.Embeddings.Tag)
	rvecs := common.ReadEmbeddingsOrExit(config.Embeddings.DepRel)
	fvecs := common.ReadEmbeddingsOrExit(config.Embeddings.Feature)
	cvecs := common.ReadEmbeddingsOrExit(config.Embeddings.Char)

	layerVecs := map[addr.Layer]*go2vec.Embeddings{
		addr.TOKEN:   vecs,
		addr.TAG:     tvecs,
		addr.DEPREL:  rvecs,
		addr.FEATURE: fvecs,
		addr.CHAR:    cvecs,
	}

	wordSplits, wordNames := splitLayers(layerVecs, ilas, addr.TOKEN, "words", *wordEmbeddings)
	tagSplits, tagNames := splitLayers(layerVecs, ilas, addr.TAG, "tags", *tagEmbeddings)
	relSplits, relNames := splitLayers(layerVecs, ilas, addr.DEPREL, "deprels", *deprelEmbeddings)
	featureSplits, featureNames := splitLayers(layerVecs, ilas, addr.FEATURE, "features", *featureEmbeddings)
	charSplits, charNames := splitCharLayers(layerVecs, ilas, "chars")

	splits := make([]uint, 0, len(wordSplits)+len(tagSplits)+len(relSplits))
	splits = append(splits, wordSplits...)
	splits = append(splits, tagSplits...)
	splits = append(splits, relSplits...)
	splits = append(splits, featureSplits...)
	splits = append(splits, charSplits...)

	splitNames := make([]string, 0, len(wordNames)+len(tagNames)+len(relNames)+len(featureNames))
	splitNames = append(splitNames, wordNames...)
	splitNames = append(splitNames, tagNames...)
	splitNames = append(splitNames, relNames...)
	splitNames = append(splitNames, featureNames...)
	splitNames = append(splitNames, charNames...)

	layers = append(layers, createSliceLayer("slice", "data", splits, splitNames))

	concatLayers := make([]string, 0)

	if *wordEmbeddings {
		layers = append(layers, createEmbeddingLayers("words", "wordembed", wordSplits, 50, false)...)
		concatLayers = append(concatLayers, names("wordembed", wordSplits)...)
	} else {
		layers = append(layers, createActivationLayer("flat_words", "words", "flat_words", "Flatten"))
		concatLayers = append(concatLayers, "flat_words")
	}

	if *tagEmbeddings {
		layers = append(layers, createEmbeddingLayers("tags", "tagembed", tagSplits, 50, false)...)
		concatLayers = append(concatLayers, names("tagembed", tagSplits)...)
	} else {
		layers = append(layers, createActivationLayer("flat_tags", "tags", "flat_tags", "Flatten"))
		concatLayers = append(concatLayers, "flat_tags")
	}

	if *deprelEmbeddings {
		layers = append(layers, createEmbeddingLayers("deprels", "relembed", relSplits, 50, false)...)
		concatLayers = append(concatLayers, names("relembed", relSplits)...)
	} else {
		layers = append(layers, createActivationLayer("flat_deprels", "deprels", "flat_deprels", "Flatten"))
		concatLayers = append(concatLayers, "flat_deprels")
	}

	if *featureEmbeddings {
		layers = append(layers, createEmbeddingLayers("features", "featureembed", featureSplits, 50, false)...)
		concatLayers = append(concatLayers, names("featureembed", relSplits)...)
	} else if layerVecs[addr.FEATURE] != nil {
		layers = append(layers, createActivationLayer("flat_features", "features", "flat_features", "Flatten"))
		concatLayers = append(concatLayers, "flat_features")
	}

	layers = append(layers, createEmbeddingLayers("chars", "morph", charSplits, 210, true)...)
	for _, name := range names("morph", charSplits) {
		layers = append(layers, createActivationLayer(fmt.Sprintf("%s_activation", name), name, name, "Sigmoid"))
	}
	concatLayers = append(concatLayers, names("morph", charSplits)...)

	layers = append(layers,
		createConcatLayer("concat", "concat", concatLayers),
		createInnerProductLayer("ip", "concat", "ip", 200),
		createActivationLayer("tanh", "ip", "ip", "TanH"),
		createDropoutLayer("dropout", "ip", "ip", 0.05),
		createInnerProductLayer("ip2", "ip", "ip2", uint32(labelNumberer.Size())),
		createLossLayer("softmax", "ip2", "label", "softmax", "SoftmaxWithLoss"),
		createAccuracyLayer("accuracy", "ip2", "label", "accuracy", "Accuracy"),
	)

	net := &caffe.NetParameter{
		Name:  proto.String("dparnn-net"),
		Layer: layers,
	}

	proto.MarshalText(os.Stdout, net)
}