Serve tensors from server as bytes to avoid the ~300MB browser string limit.

Also store data points as Float32Array[] instead of number[][], i.e. each data point is Float32Array. This reduces memory usage by 2x.
Change: 138550170

10 files changed, 91 insertions, 57 deletions

diff --git a/tensorflow/tensorboard/backend/server_test.py b/tensorflow/tensorboard/backend/server_test.py
index 714796644e..c3088491a9 100644
--- a/tensorflow/tensorboard/backend/server_test.py
+++ b/tensorflow/tensorboard/backend/server_test.py
@@ -29,8 +29,8 @@ import os
 import shutil
 import tempfile
 import threading
-import zlib
 
+import numpy as np
 from six import BytesIO
 from six.moves import http_client
 from six.moves import xrange  # pylint: disable=redefined-builtin
@@ -264,9 +264,11 @@ class TensorboardServerTest(tf.test.TestCase):
     if 'projector' not in REGISTERED_PLUGINS:
       return
 
-    tensor_tsv = (self._get('/data/plugin/projector/tensor?run=run1&name=var1')
-                  .read())
-    self.assertEqual(tensor_tsv, b'6.0\t6.0')
+    url = '/data/plugin/projector/tensor?run=run1&name=var1'
+    tensor_bytes = self._get(url).read()
+    tensor = np.reshape(np.fromstring(tensor_bytes, dtype='float32'), [1, 2])
+    expected_tensor = np.array([[6, 6]], dtype='float32')
+    self.assertTrue(np.array_equal(tensor, expected_tensor))
 
   def testAcceptGzip_compressesResponse(self):
     response = self._get('/data/graph?run=run1&limit_attr_size=1024'
diff --git a/tensorflow/tensorboard/components/vz_projector/data-provider-proto.ts b/tensorflow/tensorboard/components/vz_projector/data-provider-proto.ts
index 656173f265..cbb8b412ea 100644
--- a/tensorflow/tensorboard/components/vz_projector/data-provider-proto.ts
+++ b/tensorflow/tensorboard/components/vz_projector/data-provider-proto.ts
@@ -79,13 +79,9 @@ export class ProtoDataProvider implements DataProvider {
       throw 'The shape doesn\'t match the length of the flattened array';
     }
     for (let i = 0; i < n; i++) {
-      let vector: number[] = [];
       let offset = i * d;
-      for (let j = 0; j < d; j++) {
-        vector.push(tensor[offset++]);
-      }
       points.push({
-        vector: vector,
+        vector: new Float32Array(tensor.slice(offset, offset + d)),
         metadata: {},
         projections: null,
         index: i
diff --git a/tensorflow/tensorboard/components/vz_projector/data-provider-server.ts b/tensorflow/tensorboard/components/vz_projector/data-provider-server.ts
index f663388eb7..a617a981d9 100644
--- a/tensorflow/tensorboard/components/vz_projector/data-provider-server.ts
+++ b/tensorflow/tensorboard/components/vz_projector/data-provider-server.ts
@@ -82,17 +82,30 @@ export class ServerDataProvider implements DataProvider {
       callback: (ds: DataSet) => void) {
     // Get the tensor.
     logging.setModalMessage('Fetching tensor values...', TENSORS_MSG_ID);
-    d3.text(
-        `${this.routePrefix}/tensor?run=${run}&name=${tensorName}`,
-        (err: any, tsv: string) => {
-          if (err) {
-            logging.setModalMessage('Error: ' + err.responseText);
-            return;
-          }
-          dataProvider.parseTensors(tsv).then(dataPoints => {
-            callback(new DataSet(dataPoints));
-          });
+    let xhr = new XMLHttpRequest();
+    xhr.open('GET', `${this.routePrefix}/tensor?run=${run}&name=${tensorName}`);
+    xhr.responseType = 'arraybuffer';
+    xhr.onprogress = (ev) => {
+      if (ev.lengthComputable) {
+        let percent = (ev.loaded * 100 / ev.total).toFixed(1);
+        logging.setModalMessage('Fetching tensor values: ' + percent + '%',
+                                TENSORS_MSG_ID);
+      }
+    };
+    xhr.onload = () => {
+      let data = new Float32Array(xhr.response);
+      this.getEmbeddingInfo(run, tensorName, embedding => {
+        let dim = embedding.tensorShape[1];
+        dataProvider.parseTensorsFromFloat32Array(data, dim).then(
+            dataPoints => {
+          callback(new DataSet(dataPoints));
         });
+      });
+    };
+    xhr.onerror = () => {
+      logging.setModalMessage('Error: ' + xhr.responseText);
+    };
+    xhr.send(null);
   }
 
   retrieveSpriteAndMetadata(run: string, tensorName: string,
diff --git a/tensorflow/tensorboard/components/vz_projector/data-provider.ts b/tensorflow/tensorboard/components/vz_projector/data-provider.ts
index 4f74a44e8b..c774c50aa4 100644
--- a/tensorflow/tensorboard/components/vz_projector/data-provider.ts
+++ b/tensorflow/tensorboard/components/vz_projector/data-provider.ts
@@ -122,9 +122,9 @@ export function parseTensors(
       // If the first label is not a number, take it as the label.
       if (isNaN(row[0] as any) || numDim === row.length - 1) {
         dataPoint.metadata['label'] = row[0];
-        dataPoint.vector = row.slice(1).map(Number);
+        dataPoint.vector = new Float32Array(row.slice(1).map(Number));
       } else {
-        dataPoint.vector = row.map(Number);
+        dataPoint.vector = new Float32Array(row.map(Number));
       }
       data.push(dataPoint);
       if (numDim == null) {
@@ -148,6 +148,29 @@ export function parseTensors(
   });
 }
 
+/** Parses a tsv text file. */
+export function parseTensorsFromFloat32Array(data: Float32Array,
+    dim: number): Promise<DataPoint[]> {
+  return runAsyncTask('Parsing tensors...', () => {
+    let N = data.length / dim;
+    let dataPoints: DataPoint[] = [];
+    let offset = 0;
+    for (let i = 0; i < N; ++i) {
+      dataPoints.push({
+        metadata: {},
+        vector: data.subarray(offset, offset + dim),
+        index: i,
+        projections: null,
+      });
+      offset += dim;
+    }
+    return dataPoints;
+  }, TENSORS_MSG_ID).then(dataPoints => {
+    logging.setModalMessage(null, TENSORS_MSG_ID);
+    return dataPoints;
+  });
+}
+
 export function analyzeMetadata(
     columnNames, pointsMetadata: PointMetadata[]): ColumnStats[] {
   let columnStats: ColumnStats[] = columnNames.map(name => {
diff --git a/tensorflow/tensorboard/components/vz_projector/data.ts b/tensorflow/tensorboard/components/vz_projector/data.ts
index 0f5cf31dca..f7766ac623 100644
--- a/tensorflow/tensorboard/components/vz_projector/data.ts
+++ b/tensorflow/tensorboard/components/vz_projector/data.ts
@@ -60,7 +60,7 @@ export interface SpriteAndMetadataInfo {
 
 export interface DataPoint extends scatterPlot.DataPoint {
   /** The point in the original space. */
-  vector: number[];
+  vector: Float32Array;
 
   /*
    * Metadata for each point. Each metadata is a set of key/value pairs
diff --git a/tensorflow/tensorboard/components/vz_projector/data_test.ts b/tensorflow/tensorboard/components/vz_projector/data_test.ts
index 34795b33e2..6d92287637 100644
--- a/tensorflow/tensorboard/components/vz_projector/data_test.ts
+++ b/tensorflow/tensorboard/components/vz_projector/data_test.ts
@@ -28,7 +28,7 @@ function makePointsWithTraces(traces: number[]) {
     let metadata: {[key: string]: any} = {};
     metadata[nextAttr] = t >= 0 ? t : null;
     points.push({
-      vector: [],
+      vector: new Float32Array(0),
       metadata: metadata,
       projections: {},
       index: i
diff --git a/tensorflow/tensorboard/components/vz_projector/knn.ts b/tensorflow/tensorboard/components/vz_projector/knn.ts
index 4db45d207b..906e077b5d 100644
--- a/tensorflow/tensorboard/components/vz_projector/knn.ts
+++ b/tensorflow/tensorboard/components/vz_projector/knn.ts
@@ -16,6 +16,7 @@ limitations under the License.
 import {runAsyncTask} from './util';
 import * as logging from './logging';
 import {KMin} from './heap';
+import {Vector} from './vector';
 import * as vector from './vector';
 
 export type NearestEntry = {
@@ -46,7 +47,7 @@ const KNN_GPU_MSG_ID = 'knn-gpu';
  */
 export function findKNNGPUCosine<T>(
     dataPoints: T[], k: number,
-    accessor: (dataPoint: T) => number[]): Promise<NearestEntry[][]> {
+    accessor: (dataPoint: T) => Float32Array): Promise<NearestEntry[][]> {
   let N = dataPoints.length;
   let dim = accessor(dataPoints[0]).length;
 
@@ -139,8 +140,8 @@ export function findKNNGPUCosine<T>(
  *   distance is above the limit.
  */
 export function findKNN<T>(
-    dataPoints: T[], k: number, accessor: (dataPoint: T) => number[],
-    dist: (a: number[], b: number[], limit: number) =>
+    dataPoints: T[], k: number, accessor: (dataPoint: T) => Float32Array,
+    dist: (a: Vector, b: Vector, limit: number) =>
         number): Promise<NearestEntry[][]> {
   return runAsyncTask<NearestEntry[][]>('Finding nearest neighbors...', () => {
     let N = dataPoints.length;
@@ -218,8 +219,8 @@ function minDist(
  */
 export function findKNNofPoint<T>(
     dataPoints: T[], pointIndex: number, k: number,
-    accessor: (dataPoint: T) => number[],
-    distance: (a: number[], b: number[]) => number) {
+    accessor: (dataPoint: T) => Float32Array,
+    distance: (a: Vector, b: Vector) => number) {
   let kMin = new KMin<NearestEntry>(k);
   let a = accessor(dataPoints[pointIndex]);
   for (let i = 0; i < dataPoints.length; ++i) {
diff --git a/tensorflow/tensorboard/components/vz_projector/vector.ts b/tensorflow/tensorboard/components/vz_projector/vector.ts
index 7f7d9f30f7..fd092bc4d1 100644
--- a/tensorflow/tensorboard/components/vz_projector/vector.ts
+++ b/tensorflow/tensorboard/components/vz_projector/vector.ts
@@ -19,7 +19,7 @@ import {assert} from './util';
  * @fileoverview Useful vector utilities.
  */
 
-export type Vector = number[];
+export type Vector = Float32Array | number[];
 export type Point2D = [number, number];
 export type Point3D = [number, number, number];
 
@@ -43,9 +43,9 @@ export function sum(a: Vector): number {
 }
 
 /** Returns the sum of two vectors, i.e. a + b */
-export function add(a: Vector, b: Vector): Vector {
+export function add(a: Vector, b: Vector): Float32Array {
   assert(a.length === b.length, 'Vectors a and b must be of same length');
-  let result = new Array(a.length);
+  let result = new Float32Array(a.length);
   for (let i = 0; i < a.length; ++i) {
     result[i] = a[i] + b[i];
   }
@@ -53,9 +53,9 @@ export function add(a: Vector, b: Vector): Vector {
 }
 
 /** Subtracts vector b from vector a, i.e. returns a - b */
-export function sub(a: Vector, b: Vector): Vector {
+export function sub(a: Vector, b: Vector): Float32Array {
   assert(a.length === b.length, 'Vectors a and b must be of same length');
-  let result = new Array(a.length);
+  let result = new Float32Array(a.length);
   for (let i = 0; i < a.length; ++i) {
     result[i] = a[i] - b[i];
   }
@@ -151,12 +151,13 @@ export function unit(a: Vector): void {
  * @param vectors Array of vectors to be projected.
  * @param newDim The resulting dimension of the vectors.
  */
-export function projectRandom(vectors: number[][], newDim: number): number[][] {
+export function projectRandom(vectors: Float32Array[], newDim: number):
+    Float32Array[] {
   let dim = vectors[0].length;
   let N = vectors.length;
-  let newVectors: number[][] = new Array(N);
+  let newVectors: Float32Array[] = new Array(N);
   for (let i = 0; i < N; ++i) {
-    newVectors[i] = new Array(newDim);
+    newVectors[i] = new Float32Array(newDim);
   }
   // Make nDim projections.
   for (let k = 0; k < newDim; ++k) {
@@ -175,15 +176,6 @@ export function project2d(a: Vector, dir1: Vector, dir2: Vector): Point2D {
   return [dot(a, dir1), dot(a, dir2)];
 }
 
-/** Returns a vector filled with zeros */
-export function zeros(length: number): Vector {
-  let result = new Array(length);
-  for (let i = 0; i < length; ++i) {
-    result[i] = 0;
-  }
-  return result;
-}
-
 /**
  * Computes the centroid of the data points. If the provided data points are not
  * vectors, an accessor function needs to be provided.
@@ -197,7 +189,7 @@ export function centroid<T>(dataPoints: T[], accessor?: (a: T) => Vector):
     accessor = (a: T) => <any>a;
   }
   assert(dataPoints.length >= 0, '`vectors` must be of length >= 1');
-  let centroid = zeros(accessor(dataPoints[0]).length);
+  let centroid = new Float32Array(accessor(dataPoints[0]).length);
   for (let i = 0; i < dataPoints.length; ++i) {
     let dataPoint = dataPoints[i];
     let vector = accessor(dataPoint);
@@ -215,9 +207,9 @@ export function centroid<T>(dataPoints: T[], accessor?: (a: T) => Vector):
  * Generates a vector of the specified size where each component is drawn from
  * a random (0, 1) gaussian distribution.
  */
-export function rn(size: number): Vector {
+export function rn(size: number): Float32Array {
   let normal = d3.random.normal();
-  let result = new Array(size);
+  let result = new Float32Array(size);
   for (let i = 0; i < size; ++i) {
     result[i] = normal();
   }
@@ -249,7 +241,7 @@ export function cosSim(a: Vector, b: Vector): number {
  * typed array with row-first order.
  */
 export function toTypedArray<T>(
-    dataPoints: T[], accessor: (dataPoint: T) => number[]): Float32Array {
+    dataPoints: T[], accessor: (dataPoint: T) => Float32Array): Float32Array {
   let N = dataPoints.length;
   let dim = accessor(dataPoints[0]).length;
   let result = new Float32Array(N * dim);
diff --git a/tensorflow/tensorboard/components/vz_projector/vz-projector-projections-panel.ts b/tensorflow/tensorboard/components/vz_projector/vz-projector-projections-panel.ts
index 7b28b94060..6a26c985e8 100644
--- a/tensorflow/tensorboard/components/vz_projector/vz-projector-projections-panel.ts
+++ b/tensorflow/tensorboard/components/vz_projector/vz-projector-projections-panel.ts
@@ -14,6 +14,7 @@ limitations under the License.
 ==============================================================================*/
 
 import {DataSet, SpriteAndMetadataInfo, PCA_SAMPLE_DIM, Projection, SAMPLE_SIZE, State} from './data';
+import {Vector} from './vector';
 import * as vector from './vector';
 import {Projector} from './vz-projector';
 import {ProjectorInput} from './vz-projector-input';
@@ -44,12 +45,16 @@ export let ProjectionsPanelPolymer = PolymerElement({
 type InputControlName = 'xLeft' | 'xRight' | 'yUp' | 'yDown';
 
 type CentroidResult = {
-  centroid?: number[]; numMatches?: number;
+  centroid?: Vector;
+  numMatches?: number;
 };
 
 type Centroids = {
-  [key: string]: number[]; xLeft: number[]; xRight: number[]; yUp: number[];
-  yDown: number[];
+  [key: string]: Vector;
+  xLeft: Vector;
+  xRight: Vector;
+  yUp: Vector;
+  yDown: Vector;
 };
 
 /**
diff --git a/tensorflow/tensorboard/plugins/projector/plugin.py b/tensorflow/tensorboard/plugins/projector/plugin.py
index f6dd2bcad7..12619637ce 100644
--- a/tensorflow/tensorboard/plugins/projector/plugin.py
+++ b/tensorflow/tensorboard/plugins/projector/plugin.py
@@ -20,6 +20,7 @@ from __future__ import print_function
 
 import imghdr
 import os
+import numpy as np
 
 from google.protobuf import json_format
 from google.protobuf import text_format
@@ -57,8 +58,8 @@ def _read_tensor_file(fpath):
     tensor = []
     for line in f:
       if line:
-        tensor.append(line.rstrip('\n').split('\t'))
-  return tensor
+        tensor.append(map(float, line.rstrip('\n').split('\t')))
+  return np.array(tensor, dtype='float32')
 
 
 def _latest_checkpoints_changed(configs, run_path_pairs):
@@ -341,9 +342,10 @@ class ProjectorPlugin(TBPlugin):
 
     # Sample the tensor
     tensor = tensor[:LIMIT_NUM_POINTS]
-    # Stream it as TSV.
-    tsv = '\n'.join(['\t'.join([str(val) for val in row]) for row in tensor])
-    request.respond(tsv, 'text/tab-separated-values')
+    if tensor.dtype != 'float32':
+      tensor = tensor.astype(dtype='float32', copy=False)
+    data_bytes = tensor.tobytes()
+    request.respond(data_bytes, 'application/octet-stream')
 
   def _serve_bookmarks(self, request, query_params):
     run = query_params.get('run')