Embedding doc

frontend/src/common/component/AppMenu.vue

@@ -63,13 +63,11 @@ export default {
           title: 'TEXTS',
           name: 'texts',
         },
-        /* // Hide the top menu
-            {
-                url: '/HighDimensional',
-                title: 'HighDimensional',
-                name: 'HighDimensional'
-            }
-            */
+        {
+          url: '/HighDimensional',
+          title: 'HighDimensional',
+          name: 'HighDimensional',
+        },
       ],
     };
   },

frontend/src/high-dimensional/HighDimensional.vue

@@ -7,6 +7,7 @@
         :search-text="config.searchText"
         :dimension="config.dimension"
         :embedding-data="embeddingData"
+        :show-loading="showLoading"
       />
     </div>
     <div class="visual-dl-page-right">
@@ -39,11 +40,12 @@ export default {
         searchText: '',
         displayWordLabel: true,
         dimension: '2',
-        reduction: 'tsne',
+        reduction: 'pca',
         selectedRun: '',
         running: true,
       },
       embeddingData: [],
+      showLoading: false,
     };
   },
   created() {
@@ -83,13 +85,17 @@ export default {
   },
   methods: {
     fetchDatasets() {
+      this.showLoading = true;
+
       // Fetch the data from the server. Passing dimension and reduction method
       let params = {
         dimension: this.config.dimension,
         reduction: this.config.reduction,
         run: this.config.selectedRun,
       };
       getHighDimensionalDatasets(params).then(({errno, data}) => {
+        this.showLoading = false;
+
         let vectorData = data.embedding;
         let labels = data.labels;
 

frontend/src/high-dimensional/ui/Chart.vue

@@ -35,6 +35,10 @@ export default {
       type: String,
       required: true,
     },
+    showLoading: {
+      type: Boolean,
+      required: true,
+    },
   },
   data() {
     return {
@@ -53,15 +57,11 @@ export default {
   created() {},
   mounted() {
     this.createChart();
-    this.myChart.showLoading();
-
     this.set2DChartOptions();
     this.setDisplayWordLabel();
   },
   watch: {
     embeddingData: function(val) {
-      this.myChart.hideLoading();
-
       // Got new data, pass to the filter function to render the 'matched' set and 'not matched' set
       this.filterSeriesDataAndSetOption(this.searchText);
     },
@@ -70,7 +70,6 @@ export default {
     },
     dimension: function(val) {
       this.myChart.clear();
-      this.myChart.showLoading();
       if (val === '2') {
         this.set2DChartOptions();
         this.setDisplayWordLabel();
@@ -82,6 +81,13 @@ export default {
     searchText: function(val) {
       this.filterSeriesDataAndSetOption(val);
     },
+    showLoading: function(val) {
+      if (val) {
+        this.myChart.showLoading();
+      } else {
+        this.myChart.hideLoading();
+      }
+    },
   },
   methods: {
     createChart() {

frontend/src/high-dimensional/ui/Config.vue

@@ -31,12 +31,13 @@
       label="Reduction Method"
       v-model="config.reduction"
       dark>
-      <v-radio
-        label="T-SNE"
-        value="tsne"/>
       <v-radio
         label="PCA"
         value="pca"/>
+      <v-radio
+        label="T-SNE"
+        value="tsne"/>
+
     </v-radio-group>
 
     <v-radio-group

visualdl/logic/pybind.cc

@@ -253,10 +253,19 @@ PYBIND11_MODULE(core, m) {
       .def("total_records", &cp::TextReader::total_records)
       .def("size", &cp::TextReader::size);
 
-  py::class_<cp::Embedding>(m, "EmbeddingWriter")
+  py::class_<cp::Embedding>(m, "EmbeddingWriter", R"pbdoc(
+        PyBind class. Must instantiate through the LogWriter.
+      )pbdoc")
       .def("set_caption", &cp::Embedding::SetCaption)
-      .def("add_embeddings_with_word_list",
-           &cp::Embedding::AddEmbeddingsWithWordList);
+      .def(
+          "add_embeddings_with_word_list"
+          R"pbdoc(
+            Add embedding record. Each run can only store one embedding data.
+
+            :param embedding: hot vector of embedding words
+            :type embedding: list
+          )pbdoc",
+          &cp::Embedding::AddEmbeddingsWithWordList);
 
   py::class_<cp::EmbeddingReader>(m, "EmbeddingReader")
       .def("get_all_labels", &cp::EmbeddingReader::get_all_labels)

visualdl/python/storage.py

@@ -143,6 +143,9 @@ def text(self, tag):
         return self.reader.get_text(tag)
 
     def embedding(self):
+        """
+        Get the embedding reader.
+        """
         return self.reader.get_embedding(EMBEDDING_TAG)
 
     def audio(self, tag):
@@ -292,9 +295,19 @@ def text(self, tag):
         return self.writer.new_text(tag)
 
     def embedding(self):
+        """
+        Create an embedding writer that used to write
+        embedding data.
+
+        :return: A embedding writer to record embedding data
+        :rtype: embeddingWriter
+        """
         return self.writer.new_embedding(EMBEDDING_TAG)
 
     def save(self):
+        """
+        Force the VisualDL to sync with the file system.
+        """
         self.writer.save()
 
     def __enter__(self):

visualdl/server/lib.py

@@ -307,19 +307,18 @@ def get_embeddings(storage, mode, reduction, dimension=2, num_records=5000):
     with storage.mode(mode) as reader:
         embedding = reader.embedding()
         labels = embedding.get_all_labels()
-        high_dimensional_vectors = embedding.get_all_embeddings()
+        high_dimensional_vectors = np.array(embedding.get_all_embeddings())
 
-        # TODO: Move away from sklearn
         if reduction == 'tsne':
-            from sklearn.manifold import TSNE
-            tsne = TSNE(
-                perplexity=30, n_components=dimension, init='pca', n_iter=5000)
-            low_dim_embs = tsne.fit_transform(high_dimensional_vectors)
+            import tsne
+            low_dim_embs = tsne.tsne(
+                high_dimensional_vectors,
+                dimension,
+                initial_dims=50,
+                perplexity=30.0)
 
         elif reduction == 'pca':
-            from sklearn.decomposition import PCA
-            pca = PCA(n_components=3)
-            low_dim_embs = pca.fit_transform(high_dimensional_vectors)
+            low_dim_embs = simple_pca(high_dimensional_vectors, dimension)
 
         return {"embedding": low_dim_embs.tolist(), "labels": labels}
 
@@ -393,3 +392,23 @@ def _handler(key, func, *args, **kwargs):
         return data
 
     return _handler
+
+
+# A simple PCA implementaiton to do the dimension reduction.
+def simple_pca(x, dimension):
+    # Center the data.
+    x -= np.mean(x, axis=0)
+
+    # Computing the Covariance Matrix
+    cov = np.cov(x, rowvar=False)
+
+    # Get eigenvectors and eigenvalues from the covariance matrix
+    eigvals, eigvecs = np.linalg.eig(cov)
+
+    # Sort the eigvals from high to low
+    order = np.argsort(eigvals)[::-1]
+
+    # Drop the eigenvectors with low eigenvalues
+    eigvecs = eigvecs[:, order[:dimension]]
+
+    return np.dot(x, eigvecs)