Unable to load model using asyncStorageIO API in react-native

To get help from the community, we encourage using Stack Overflow and the tensorflow.js tag.

TensorFlow.js version

0.3.0

Browser version

React Native: v0.61.5

Describe the problem or feature request

I am getting the following error while loading a model using AsyncStorageIO API.

Row too big to fit into CursorWindow requiredPos=0, totalRows=1
* http://192.168.1.7:19001/index.bundle?platform=android&dev=true&minify=false&hot=false:203062:24 in convertError
* [native code]:null in map
- node_modules/@react-native-community/async-storage/src/AsyncStorage.native.js:97:32 in RCTAsyncStorage.multiGet$argument_1
- node_modules/react-native/Libraries/BatchedBridge/MessageQueue.js:483:4 in __invokeCallback
- node_modules/react-native/Libraries/BatchedBridge/MessageQueue.js:135:28 in __guard$argument_0
- node_modules/react-native/Libraries/BatchedBridge/MessageQueue.js:384:10 in __guard
- node_modules/react-native/Libraries/BatchedBridge/MessageQueue.js:134:17 in __guard$argument_0
* [native code]:null in invokeCallbackAndReturnFlushedQueue

In my project, I have a model with json and weights which I have loaded using bundleResourceIO API. Then, I am updating the model using the given code and saving it with asyncStorageIO API. This flow goes well and good. But, I try to load the model using asyncStorageIO API, I get the previously described error.

 const optimizer = tf.train.adam();
      model.compile({
        optimizer: optimizer,
        loss: "categoricalCrossentropy",
        metrics: ["accuracy"],
      });
model
        .trainOnBatch(
          tf
            .tensor2d(imageTensors, [numOfElems, 784])
            .reshape([numOfElems, 28, 28, 1]),
          tf.tensor2d(labelTensors, [numOfElems, 10]).reshape([numOfElems, 10])
        )
        .then((d) => {
          console.log("Model updated");
          model
            .save(asyncStorageIO("mnist-model"))
            .then((res) => console.log("Model Saved", res))
            .catch((e) => console.log(e));
        });

[geetesh-gupta]

Hi, thanks for your responses 😊. Currently due to time constraint, I created a different model with less layers that can work with the project, so reducing the size to 172 KB. But, when I will get time, I will try to incorporate the above suggested solutions and will respond back.

[nixoschu]

Hey, sorry for bringing this up again, but I have the same problem as @geetesh-gupta on Android only. iOS works just fine.

For the first run, I download my model directly from the cloud using const model = tf.loadLayersModel(modelUri). Then, I save it on asyncStorage doing model.save(asyncStorageIO('custom-model')). For the second run and on, I have a logic to check if I have any model saved on the asyncStorage and load it if there is. This is working very nicely on iOS, though I’m facing the same problems that @geetesh-gupta reported on Android.

I have a 20MB sized TF custom model. I have updated my gradle.properties file, as @tafsiri suggested from the AsyncStorage docs, to have a max of 50 MB in the SQLite DB: AsyncStorage_db_size_in_MB=50, but even though, it looks like this is an inner limitation of SQLite that can only work with a maximum 2MB in the Cursor Window buffer.

Even though it feels like I can save the model in the first trial, when I try to load it I get this: [Error: Row too big to fit into CursorWindow requiredPos=0, totalRows=1]

And if I try to save again the Model with the same path Key (to overwrite it), then I get this: Out of memory stringify@[native code]

@tafsiri - right now, I’m trying to save/retrieve model to the FileSystem, using react-native-fs. Do you think this approach could work? Or do you have any other workaround suggestion?

Best regards.

Exact same for me, my model is 12MB