idmission-web-sdk 1.0.194 → 1.0.196

package/dist/sdk2.cjs.development.js

@@ -48,7 +48,7 @@ var LanguageDetector__default = /*#__PURE__*/_interopDefaultLegacy(LanguageDetec
 var i18n__default = /*#__PURE__*/_interopDefaultLegacy(i18n);
 var SignatureCanvas__default = /*#__PURE__*/_interopDefaultLegacy(SignatureCanvas);
 
-var webSdkVersion = '1.0.194';
+var webSdkVersion = '1.0.196';
 
 function getPlatform() {
   if (typeof window === 'undefined') return;
@@ -1918,20 +1918,18 @@ function loadDocumentDetectionModel(url) {
           return [4 /*yield*/, tf__namespace.loadGraphModel(url)];
         case 1:
           documentDetectionModel = _c.sent();
-          return [4 /*yield*/, tf__namespace.ready()];
+          return [4 /*yield*/, tf__namespace.ready()
+          // warm the model by predicting on an empty image
+          ];
         case 2:
           _c.sent();
-          _c.label = 3;
-        case 3:
-          _c.trys.push([3,, 5, 6]);
-          tf__namespace.engine().startScope();
           zeroes = tf__namespace.zeros([448, 448, 3]);
           tfInt32 = tf__namespace.cast(zeroes, 'int32');
           tf4d_ = tf__namespace.tensor4d(Array.from(tfInt32.dataSync()), [1, 448, 448, 3]);
           tf4d = tf__namespace.cast(tf4d_, 'int32');
           _b = (_a = tf__namespace).dispose;
           return [4 /*yield*/, documentDetectionModel.executeAsync(tf4d)];
-        case 4:
+        case 3:
           _b.apply(_a, [_c.sent()]);
           tfInt32.dispose();
           tf4d.dispose();
@@ -1945,11 +1943,6 @@ function loadDocumentDetectionModel(url) {
             scores: outputKeys.indexOf('Identity_4'),
             labels: outputKeys.indexOf('Identity_2')
           };
-          return [3 /*break*/, 6];
-        case 5:
-          tf__namespace.engine().endScope();
-          return [7 /*endfinally*/];
-        case 6:
           return [2 /*return*/, documentDetectionModel];
       }
     });
@@ -1992,22 +1985,22 @@ function loadFocusModel(url) {
           return [4 /*yield*/, tf__namespace.loadLayersModel(url)];
         case 1:
           focusModel = _e.sent();
-          return [4 /*yield*/, tf__namespace.ready()];
+          return [4 /*yield*/, tf__namespace.ready()
+          // warm the model by predicting on an empty image...
+          ];
         case 2:
           _e.sent();
-          _e.label = 3;
-        case 3:
-          _e.trys.push([3,, 6, 7]);
-          tf__namespace.engine().startScope();
           zeroes = tf__namespace.zeros([448, 448, 3]);
           tfInt32 = tf__namespace.cast(zeroes, 'float32');
           _b = (_a = tf__namespace).tensor4d;
           _d = (_c = Array).from;
           return [4 /*yield*/, tfInt32.data()];
-        case 4:
+        case 3:
           tf4d_ = _b.apply(_a, [_d.apply(_c, [_e.sent()]), [1, 448, 448, 3]]);
           tf4d = tf__namespace.cast(tf4d_, 'float32');
-          warmupResult = focusModel.predict(tf4d);
+          return [4 /*yield*/, focusModel.predict(tf4d)];
+        case 4:
+          warmupResult = _e.sent();
           return [4 /*yield*/, warmupResult.data()];
         case 5:
           _e.sent();
@@ -2016,11 +2009,6 @@ function loadFocusModel(url) {
           tf4d.dispose();
           tf4d_.dispose();
           tf__namespace.dispose(zeroes);
-          return [3 /*break*/, 7];
-        case 6:
-          tf__namespace.engine().endScope();
-          return [7 /*endfinally*/];
-        case 7:
           return [2 /*return*/, focusModel];
       }
     });
@@ -9044,30 +9032,26 @@ function makeDocumentDetectionPredictions(image, logPredictions) {
           _e.label = 1;
         case 1:
           _e.trys.push([1, 7, 8, 9]);
+          console.log('  Tensors before:', tf__namespace.memory().numTensors);
           tf__namespace.engine().startScope();
           tfImg = tf__namespace.browser.fromPixels(image);
           if (logPredictions) console.time('resize');
           tfResize = tf__namespace.image.resizeBilinear(tfImg, [448, 448]);
           if (logPredictions) console.timeEnd('resize');
-          tfImg.dispose();
           if (logPredictions) console.time('convert to int32');
           tfInt32 = tf__namespace.cast(tfResize, 'int32');
-          tfResize.dispose();
           _b = (_a = tf__namespace).tensor4d;
           _d = (_c = Array).from;
           return [4 /*yield*/, tfInt32.data()];
         case 2:
           tf4d_ = _b.apply(_a, [_d.apply(_c, [_e.sent()]), [1, 448, 448, 3]]);
-          tfInt32.dispose();
           tf4d = tf__namespace.cast(tf4d_, 'int32');
           if (logPredictions) console.timeEnd('convert to int32');
-          tf4d_.dispose();
           if (logPredictions) console.time('predictions');
           return [4 /*yield*/, documentDetectionModel.executeAsync(tf4d)];
         case 3:
           result = _e.sent();
           if (logPredictions) console.timeEnd('predictions');
-          tf4d.dispose();
           predictions = result;
           if (logPredictions) {
             console.log('predictions made!', predictions.map(function (r) {
@@ -9086,7 +9070,6 @@ function makeDocumentDetectionPredictions(image, logPredictions) {
           labels = labels.map(function (l) {
             return Math.round(l);
           });
-          tf__namespace.dispose(result);
           time = new Date().getTime() - startedAt.getTime();
           return [2 /*return*/, {
             boxes: boxes,
@@ -9105,6 +9088,8 @@ function makeDocumentDetectionPredictions(image, logPredictions) {
           }];
         case 8:
           tf__namespace.engine().endScope();
+          // tf.engine().disposeVariables()
+          console.log('Tensors afters:', tf__namespace.memory().numTensors);
           return [7 /*endfinally*/];
         case 9:
           return [2 /*return*/];
@@ -9152,27 +9137,21 @@ function makeFocusPrediction(imageData, canvases, bbox, logPredictions) {
             tfImg = tf__namespace.browser.fromPixels(imageData);
           }
           tfResize = tf__namespace.image.resizeBilinear(tfImg, [448, 448]).div(tf__namespace.scalar(127.5)).sub(tf__namespace.scalar(1));
-          tfImg.dispose();
           tfInt32 = tf__namespace.cast(tfResize, 'float32');
-          tfResize.dispose();
           _d = (_c = tf__namespace).tensor4d;
           _f = (_e = Array).from;
           return [4 /*yield*/, tfInt32.data()];
         case 2:
           tf4d_ = _d.apply(_c, [_f.apply(_e, [_g.sent()]), [1, 448, 448, 3]]);
-          tfInt32.dispose();
           tf4d = tf__namespace.cast(tf4d_, 'float32');
-          tf4d_.dispose();
           if (logPredictions) console.time('focus prediction');
           return [4 /*yield*/, focusModel.predict(tf4d)];
         case 3:
           predictions = _g.sent();
-          tf4d.dispose();
           return [4 /*yield*/, predictions.array()];
         case 4:
           focusPrediction = _g.sent()[0];
           if (logPredictions) console.timeEnd('focus prediction');
-          tf__namespace.dispose(predictions);
           predictionTime = new Date().getTime() - startedAt.getTime();
           return [2 /*return*/, [focusPrediction, predictionTime]];
         case 5: