diff --git a/benchmark/configs/android/mnn.yml b/benchmark/configs/android/mnn.yml
new file mode 100644
index 00000000..363cedcc
--- /dev/null
+++ b/benchmark/configs/android/mnn.yml
@@ -0,0 +1,32 @@
+# Configuration file of running tensorflow backend
+
+# ========== Cluster configuration ==========
+# ip address of the parameter server (need 1 GPU process)
+ps_ip: localhost
+
+exp_path: $FEDSCALE_HOME/fedscale/cloud
+
+aggregator_entry: aggregation/aggregator_mnn.py
+
+auth:
+    ssh_user: ""
+    ssh_private_key: ~/.ssh/id_rsa
+
+# cmd to run before we can indeed run FAR (in order)
+setup_commands:
+    - source $HOME/anaconda3/bin/activate fedscale
+
+# ========== Additional job configuration ==========
+# Default parameters are specified in config_parser.py, wherein more description of the parameter can be found
+
+job_conf:
+    - job_name: android-mnn              # Generate logs under this folder: log_path/job_name/time_stamp
+    - log_path: $FEDSCALE_HOME/benchmark # Path of log files
+    - experiment_mode: mobile
+    - num_participants: 1                # Number of participants per round, we use K=100 in our paper, large K will be much slower
+    - model: linear                      # Need to define the model in aggregator_mnn.py
+    - learning_rate: 0.01
+    - batch_size: 32
+    - input_shape: 32 32 3
+    - num_classes: 10
+    - test_bsz: 16
diff --git a/benchmark/configs/android/tflite.yml b/benchmark/configs/android/tflite.yml
new file mode 100644
index 00000000..a853c2c3
--- /dev/null
+++ b/benchmark/configs/android/tflite.yml
@@ -0,0 +1,33 @@
+# Configuration file of running tensorflow backend
+
+# ========== Cluster configuration ==========
+# ip address of the parameter server (need 1 GPU process)
+ps_ip: localhost
+
+exp_path: $FEDSCALE_HOME/fedscale/cloud
+
+aggregator_entry: aggregation/aggregator_tflite.py
+
+auth:
+    ssh_user: ""
+    ssh_private_key: ~/.ssh/id_rsa
+
+# cmd to run before we can indeed run FAR (in order)
+setup_commands:
+    - source $HOME/anaconda3/bin/activate fedscale
+
+# ========== Additional job configuration ==========
+# Default parameters are specified in config_parser.py, wherein more description of the parameter can be found
+
+job_conf:
+    - job_name: android-tflite           # Generate logs under this folder: log_path/job_name/time_stamp
+    - log_path: $FEDSCALE_HOME/benchmark # Path of log files
+    - experiment_mode: mobile
+    - num_participants: 1                # Number of participants per round, we use K=100 in our paper, large K will be much slower
+    - model: linear                      # Need to define the model in tf_aggregator.py
+    - learning_rate: 0.01
+    - batch_size: 32
+    - input_shape: 32 32 3
+    - num_classes: 10
+    - test_bsz: 16
+    - engine: 'tensorflow'
diff --git a/docs/fedscale-deploy.png b/docs/fedscale-deploy.png
new file mode 100644
index 00000000..736e9037
Binary files /dev/null and b/docs/fedscale-deploy.png differ
diff --git a/fedscale/cloud/aggregation/README.md b/fedscale/cloud/aggregation/README.md
deleted file mode 100644
index f86e89f3..00000000
--- a/fedscale/cloud/aggregation/README.md
+++ /dev/null
@@ -1,49 +0,0 @@
-# Aggregation for Mobiles
-
-This document contains explanation and instruction of aggregation for mobiles.
-
-An example android aggregator accompanied by 
-- [MNN](https://github.com/SymbioticLab/FedScale/fedscale/edge/mnn/). The android app has [MNN](https://github.com/alibaba/MNN) backend support for training and testing.
-- [TFLite](https://github.com/SymbioticLab/FedScale/fedscale/edge/tflite/). The android app has [TFLite](https://www.tensorflow.org/lite) backend support for training and testing.
-
-## MNN
-
-`fedscale/cloud/aggregation/aggregator_mnn.py` contains an inherited version of aggregator. While keeping all functionalities of the original [aggregator](https://github.com/SymbioticLab/FedScale/blob/master/fedscale/cloud/aggregation/aggregator.py), it adds support to do bijective conversion between PyTorch model and MNN model. It uses JSON to communicate with android client.
-
-**Note**:
-MNN does not support direct conversion from MNN to PyTorch model, so we did a manual conversion from MNN to JSON, then from JSON to PyTorch model. We currently only support Convolution (including Linear) and BatchNorm conversion. We welcome contribution to support more conversion for operators with trainable parameters.
-
-`fedscale/utils/models/mnn_model_provider.py` contains all the code necessary for MNN<->PyTorch model conversion and currently supported PyTorch models that can be converted to MNN without bugs.
-
-In order to run this aggregator with default setting in order to test sample app, please run
-```
-git clone https://github.com/SymbioticLab/FedScale.git
-cd FedScale
-source install.sh
-pip install -e .
-cd fedscale/cloud/aggregation
-python3 aggregator_mnn.py --experiment_mode mobile --num_participants 1 --num_classes 10 --input_shape 32 32 3 --model linear
-```
-and configure your android app according to the [tutorial](https://github.com/SymbioticLab/FedScale/fedscale/edge/mnn/README.md).
-
-## TFLite
-
-`fedscale/cloud/aggregation/aggregator_tflite.py` contains an inherited version of aggregator. While keeping all functionalities of the original [aggregator](https://github.com/SymbioticLab/FedScale/blob/master/fedscale/cloud/aggregation/aggregator.py), it adds support to do bijective conversion between tensorflow model and TFLite model.
-
-`fedscale/utils/models/tflite_model_provider.py` contains a simple linear model with Flatten->Dense->Dense, used for simple test of our sample android app. Please feel free to contribute to it and add more models.
-
-`fedscale/cloud/internal/tflite_model_adapter.py` defer from `fedscale/cloud/internal/tensorflow_model_adapter.py` in the way that TFLite adapter skip layers without weights, such as Flatten.
-
-In order to run this aggregator with default setting in order to test sample app, please run
-```
-git clone https://github.com/SymbioticLab/FedScale.git
-cd FedScale
-source install.sh
-pip install -e .
-cd fedscale/cloud/aggregation
-python3 aggregator_tflite.py --experiment_mode mobile --num_participants 1 --num_classes 10 --input_shape 32 32 3 --engine tensorflow --model [linear|mobilenetv3|resnet50|mobilenetv3_finetune|resnet50_finetune] --learning_rate 1e-2
-```
-and configure your android app according to the [tutorial](https://github.com/SymbioticLab/FedScale/fedscale/edge/tflite/README.md).
-
----
-If you need any other help, feel free to contact FedScale team or the developer [website](https://continue-revolution.github.io) [email](mailto:continuerevolution@gmail.com) of this android aggregator.
diff --git a/fedscale/edge/android/README.md b/fedscale/edge/android/README.md
index 753e76bb..b1918a00 100644
--- a/fedscale/edge/android/README.md
+++ b/fedscale/edge/android/README.md
@@ -1,12 +1,53 @@
-# Android TFLite Sample App
+# FedScale Deployment
 
-This directory contains minimum files modified from [MNN Android Demo](https://github.com/alibaba/MNN/tree/master/project/android/demo) and [TFLite Android Demo](https://www.tensorflow.org/lite/examples/on_device_training/overview). The training and testing will be conducted by TFLite backend, while the task execution and communication with server will be managed by Java. The sample has been tested upon image classification with a simple linear model and a small subset of [ImageNet-MINI](https://www.kaggle.com/datasets/ifigotin/imagenetmini-1000). This documentation contains a step-by-step tutorial on how to download, build and config this app on your own device, and modify this app for your own implementation and deployment.
+FedScale provides a cloud-based [aggregation service](https://github.com/SymbioticLab/FedScale/blob/master/fedscale/cloud/aggregation/README.md) and an SDK for smartphones on the edge that currently supports TensorflowLite and Alibaba MNN on Android (iOS support coming soon!). In this tutorial, we introduce how to:
 
-## Download and build sample android app
+- Initiate FedScale Cloud Service
+- Import FedScale SDK to locally fine tune models
+- Connect to FedScale cloud for federated training
+
+![fedscale deployment](../../../docs/fedscale-deploy.png)
+
+
+
+## FedScale Cloud Aggregation
+
+You may follow these steps to deploy and run the cloud server specific to [Alibaba MNN](https://github.com/SymbioticLab/FedScale/fedscale/edge/mnn/) or [TFLite](https://github.com/SymbioticLab/FedScale/fedscale/edge/tflite/).
+
+- Specify number of executors `num_participants: 1`. You may add more mobile participants to any number you want. We currently only support all participant in one single training backend.
+
+- Specify `model`. Currently we have only tested `linear` models for Alibaba MNN backend because Alibaba MNN does not support Dropout. However, you may choose one of `linear`|`mobilenetv3`|`resnet50`|`mobilenetv3_finetune`|`resnet50_finetune` models. `finetune` means that only the last 2 linear layers will be trained, but the backbone layers will be frozen.
+
+- Set `use_cuda` flag to `True` if you want to use GPU for aggregation. However, as aggregation process is sequential addition of several small tensors, GPU acceleration is very little.
+
+- Submit job
+
+	```
+	cd $FEDSCALE_HOME/docker
+	python3 driver.py submit $FEDSCALE_HOME/benchmark/configs/android/mnn.yml # If you want to run MNN backend on mobile.
+	python3 driver.py submit $FEDSCALE_HOME/benchmark/configs/android/tflite.yml # If you want to run TFLite backend on mobile.
+	```
+
+- Check logs: FedScale will generate logs under `data_path` you provided by default. Keep in mind that k8s may load balancing your job to any node on the cluster, so make sure you are checking the `data_path` on the correct node.
+
+- Stop job
+
+	```
+	cd $FEDSCALE_HOME/docker
+	python3 driver.py stop $YOUR_JOB
+	```
+
+## FedScale Mobile Runtime
+
+We provide a sample app which you can choose to 
+- Train/test models with TFLite or Alibaba MNN.
+- Fine-tune models locally **after** receiving model from the cloud.
+
+Please follow these steps to download and build the sample android app.
 
 1. Download and unzip [sample dataset (TrainTest.zip)](https://drive.google.com/file/d/1nfi3SVzjaE0LPxwj_5DNdqi6rK7BU8kb/view?usp=sharing) to `assets/` directory. Remove `TrainTest.zip` after unzip to save space on your mobile device. After unzip, you should see 3 files and 2 directories under `assets/`:
-   1. `TrainSet`: Training set directory, contains 316 images.
-   2. `TestSet`: Testing set directory, contains 34 images.
+   1. `TrainSet`: Training set directory, contains 320 images.
+   2. `TestSet`: Testing set directory, contains 32 images.
    3. `conf.json`: Configuration file for mobile app.
    4. `train_labels.txt`: Training label file with format `<filename> <label>`, where `<filename>` is the path after `TrainSet/`.
    5. `test_labels.txt`: Testing label file with the same format as `train_labels.txt`.
@@ -16,23 +57,13 @@ This directory contains minimum files modified from [MNN Android Demo](https://g
     - Gradle Version: 5.4.1
     - Source Compatibility: Java 8
     - Target Compatibility: Java 8
-3. Make project. Android Studio will compile and build the app for you.
-
-## Test this app with default setting
-
-1. ssh to your own server and run
-```
-cd fedscale/cloud/aggregation/android
-python3 aggregator_tflite.py --experiment_mode=mobile --num_participants=1 --engine=tensorflow
-```
-2. Change aggregator IP address inside `assets/conf.json` and click `Run` inside Android Studio.
-
-## Customize your own app
-
-1. If you want to use your own dataset, please put your data under `assets/TrainSet` and `assets/TestSet`, make sure that your label has the same format as my label file.
-   1. If you want to change the file/dir name under `assets`, please make sure to change the corresponding config in `assets` attribute inside `assets/conf.json`. 
-2. If you want to use your own model for **image classification**, please either change `channel`, `width` and `height` inside `assets/conf.json` to your own input and change `num_classes` to your own classes, or override these attributes when sending `CLIENT_TRAIN` request.
-3. If you want to use your own model for tasks other than image classification, you may need to write your own TFLite trainer and tester. Please refer to [TFLite](https://www.tensorflow.org/lite/api_docs) for further development guide. You may also need to change `channel`, `width` and `height` inside `assets/conf.json` to your own input and change or remove `num_classes`.
+3. Modify `conf.json` and dataset.
+   - You must modify `aggregator.ip` & `aggregator.port` to your own server.
+   - You can choose your training framework by modifying `model_conf.backend` to `tflite` or `mnn`.
+   - You may config your dataset information at `training_conf` and `testing_conf`.
+   - You may put your own **image classification** dataset under `/TrainSet` and `/TestSet` directories and modify `train_labels.txt` and `test_labels.txt` accordingly. The format of labels must be \<filename\> \<label\>.
+   - If you want to perform tasks other than image classification, you should modify framework-specific code [MNN](https://github.com/SymbioticLab/FedScale/fedscale/edge/android/app/src/main/java/com/fedscale/android/mnn) [TFLite](https://github.com/SymbioticLab/FedScale/fedscale/edge/android/app/src/main/java/com/fedscale/android/tflite). If you are using TFLite, you should also write your own signatures similar to [our TFLite model provider](https://github/com/SymbioticLab/FedScale/fedscale/cloud/internal/tflite_model_adapter.py)
+4. Make Project. Android Studio will compile and build the app for you. Click Run if you want to run the app on a real android device.
 
 ----
-If you need any other help, feel free to contact FedScale team or the developer [website](https://continue-revolution.github.io) [email](mailto:continuerevolution@gmail.com) of this app.
+If you need any further help, feel free to contact FedScale team or the developer [website](https://continue-revolution.github.io) [email](mailto:continuerevolution@gmail.com) of this app.
diff --git a/fedscale/edge/android/app/CMakeLists.txt b/fedscale/edge/android/app/CMakeLists.txt
index 389d8e05..9610dade 100644
--- a/fedscale/edge/android/app/CMakeLists.txt
+++ b/fedscale/edge/android/app/CMakeLists.txt
@@ -4,12 +4,10 @@ include_directories(src/main/jni/include)
 					    							  
 add_library(MNN SHARED IMPORTED)
 add_library(MNNTrain SHARED IMPORTED)
-add_library(MNNConvertDeps SHARED IMPORTED)
 add_library(MNNExpress SHARED IMPORTED)
 
 set_target_properties(MNN PROPERTIES IMPORTED_LOCATION ${CMAKE_SOURCE_DIR}/src/main/jniLibs/${ANDROID_ABI}/libMNN.so)
 set_target_properties(MNNTrain PROPERTIES IMPORTED_LOCATION ${CMAKE_SOURCE_DIR}/src/main/jniLibs/${ANDROID_ABI}/libMNNTrain.so)
-set_target_properties(MNNConvertDeps PROPERTIES IMPORTED_LOCATION ${CMAKE_SOURCE_DIR}/src/main/jniLibs/${ANDROID_ABI}/libMNNConvertDeps.so)
 set_target_properties(MNNExpress PROPERTIES IMPORTED_LOCATION ${CMAKE_SOURCE_DIR}/src/main/jniLibs/${ANDROID_ABI}/libMNN_Express.so)
 
 add_library(mnncore SHARED src/main/jni/mnntrainnative.cpp)
@@ -18,4 +16,4 @@ find_library(log-lib log)
 find_library(jnigraphics-lib jnigraphics)
 
 add_definitions(-DMNN_USE_LOGCAT)
-target_link_libraries(mnncore MNN MNNTrain MNNConvertDeps MNNExpress ${log-lib} ${jnigraphics-lib})
+target_link_libraries(mnncore MNN MNNTrain MNNExpress ${log-lib} ${jnigraphics-lib})
diff --git a/fedscale/edge/android/app/src/main/java/com/fedscale/android/app/FLApp.java b/fedscale/edge/android/app/src/main/java/com/fedscale/android/app/FLApp.java
new file mode 100644
index 00000000..406dfc50
--- /dev/null
+++ b/fedscale/edge/android/app/src/main/java/com/fedscale/android/app/FLApp.java
@@ -0,0 +1,213 @@
+package com.fedscale.android.app;
+
+import android.content.Context;
+import android.content.Intent;
+import android.content.res.Resources;
+import android.os.Bundle;
+import android.os.Handler;
+import android.os.HandlerThread;
+import android.view.WindowManager;
+import android.widget.Button;
+import android.widget.TextView;
+
+import androidx.appcompat.app.AppCompatActivity;
+
+import com.fedscale.android.R;
+import com.fedscale.android.executor.FLExecutor;
+import com.fedscale.android.utils.Common;
+
+import java.io.File;
+
+public class FLApp extends AppCompatActivity {
+
+    private TextView mUserId;
+    private TextView mRound;
+    private TextView mExecuteStatus;
+    private TextView mExecuteResult;
+    private TextView mExecuteMsg;
+
+    private Button flButton;
+    private Button finetuneButton;
+
+    private Resources res;
+
+    HandlerThread mThread;
+    Handler mHandle;
+
+    private File cacheDir;
+    private Context baseContext;
+
+    private FLExecutor executor;
+
+    private Boolean hasModel;
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        getWindow().setFlags(WindowManager.LayoutParams.FLAG_KEEP_SCREEN_ON,
+                WindowManager.LayoutParams.FLAG_KEEP_SCREEN_ON);
+        getWindow().addFlags(WindowManager.LayoutParams.FLAG_FULLSCREEN);
+        setContentView(R.layout.activity_main);
+
+        this.mUserId = findViewById(R.id.userId);
+        this.mRound = findViewById(R.id.executeRound);
+        this.mExecuteStatus = findViewById(R.id.executeStatus);
+        this.mExecuteResult = findViewById(R.id.executeResult);
+        this.mExecuteMsg = findViewById(R.id.executeMsg);
+
+        this.flButton = findViewById(R.id.flButton);
+        this.finetuneButton = findViewById(R.id.finetuneButton);
+
+        this.res = getResources();
+
+        this.cacheDir = getCacheDir();
+        this.baseContext = getBaseContext();
+
+        this.executor = new FLExecutor(this);
+        try {
+            this.executor.initExecutor();
+        } catch (Exception e) {
+            e.printStackTrace();
+        }
+
+        this.hasModel = false;
+
+        this.flButton.setOnClickListener(v -> {
+            if (this.flButton.getText().equals(this.res.getString(R.string.start_fl))) {
+                if (this.finetuneButton.isEnabled()) {
+                    this.finetuneButton.setEnabled(false);
+                }
+                this.mThread = new HandlerThread("FL");
+                this.mThread.start();
+                this.mHandle = new Handler(this.mThread.getLooper());
+                this.mHandle.post(() -> {
+                    try {
+                        this.executor.FLStart();
+                    } catch (Exception e) {
+                        e.printStackTrace();
+                    }
+                });
+                this.flButton.setText(this.res.getString(R.string.stop_fl));
+            } else if (this.flButton.getText().equals(this.res.getString(R.string.stop_fl))) {
+                try {
+                    this.executor.FLStop();
+                } catch (InterruptedException e) {
+                    e.printStackTrace();
+                }
+//                this.mThread.quit();
+                this.flButton.setText(this.res.getString(R.string.start_fl));
+                if (this.hasModel) {
+                    this.finetuneButton.setEnabled(true);
+                }
+            }
+        });
+
+        this.finetuneButton.setOnClickListener(v -> {
+            this.flButton.setEnabled(false);
+            this.mThread = new HandlerThread("Fine-tune");
+            this.mThread.start();
+            this.mHandle = new Handler(this.mThread.getLooper());
+            this.mHandle.post(() -> {
+                try {
+                    this.executor.LocalTrain();
+                } catch (Exception e) {
+                    e.printStackTrace();
+                }
+            });
+            this.finetuneButton.setEnabled(false);
+        });
+
+    }
+
+    @Override
+    protected void onPause() {
+        super.onPause();
+    }
+
+    @Override
+    protected void onResume() {
+        super.onResume();
+    }
+
+    @Override
+    public void onBackPressed() {
+        Intent home = new Intent(Intent.ACTION_MAIN);
+        home.addCategory(Intent.CATEGORY_HOME);
+        startActivity(home);
+    }
+
+    @Override
+    protected void onDestroy() {
+        try {
+            this.executor.FLStop();
+        } catch (InterruptedException e) {
+            e.printStackTrace();
+        }
+//        mThread.quit();
+        super.onDestroy();
+    }
+
+    /**
+     * Change Status on User Interface.
+     *
+     * @param newStatus The value of the updated status.
+     */
+    public void onChangeStatus(final String newStatus){
+        runOnUiThread(() -> this.mExecuteStatus.setText(newStatus));
+        if (newStatus.equals(Common.CLIENT_TRAIN_LOCALLY_FIN)) {
+            this.finetuneButton.setEnabled(true);
+            this.flButton.setEnabled(true);
+        } else if (newStatus.equals(Common.SHUT_DOWN)) {
+            this.flButton.setText(this.res.getString(R.string.start_fl));
+            if (this.hasModel) {
+                this.finetuneButton.setEnabled(true);
+            }
+        }
+    }
+
+    /**
+     * Change Round on User Interface.
+     *
+     * @param newRound The value of the updated round.
+     */
+    public void onChangeRound(final int newRound){
+        runOnUiThread(() -> this.mRound.setText(newRound));
+    }
+
+    /**
+     * When Executor get model from cloud, the local finetune button will be enabled when FL stop.
+     */
+    public void onGetModel(){
+        this.mExecuteMsg.setText(this.res.getString(R.string.has_model));
+        this.hasModel = true;
+    }
+
+    /**
+     * When Executor write model to the disk, the local finetune button will be disabled when FL stop.
+     */
+    public void onWriteModel() {
+        this.mExecuteMsg.setText(this.res.getString(R.string.no_model));
+        this.finetuneButton.setEnabled(false);
+        this.hasModel = false;
+    }
+
+    /**
+     * Initialization of the User Interface.
+     * Three lines of UI: executor id, status, result (not changed)
+     */
+    public void initUI(String userIdMsg) {
+        runOnUiThread(()->{
+            this.mUserId.setText(userIdMsg);
+            this.mExecuteStatus.setText(Common.CLIENT_CONNECT);
+            this.mExecuteResult.setText("");
+        });
+    }
+
+    public File getCacheDir() {
+        return this.cacheDir;
+    }
+
+    public Context getBaseContext() {
+        return this.baseContext;
+    }
+}
diff --git a/fedscale/edge/android/app/src/main/java/com/fedscale/android/executor/FLExecutor.java b/fedscale/edge/android/app/src/main/java/com/fedscale/android/executor/FLExecutor.java
index 393b01ec..d4c9de2b 100644
--- a/fedscale/edge/android/app/src/main/java/com/fedscale/android/executor/FLExecutor.java
+++ b/fedscale/edge/android/app/src/main/java/com/fedscale/android/executor/FLExecutor.java
@@ -1,15 +1,8 @@
 package com.fedscale.android.executor;
 
-import android.content.Intent;
-import android.os.Handler;
-import android.os.HandlerThread;
-import androidx.appcompat.app.AppCompatActivity;
-import android.os.Bundle;
 import android.util.Log;
-import android.view.WindowManager;
-import android.widget.TextView;
 
-import com.fedscale.android.R;
+import com.fedscale.android.app.FLApp;
 import com.fedscale.android.mnn.MNNBackend;
 import com.fedscale.android.utils.Backend;
 import com.fedscale.android.utils.MessageProcessor;
@@ -46,7 +39,7 @@
  * Training and executing will be handled inside MNN C++.
  * Server-client communication will be handled in JAVA.
  */
-public class FLExecutor extends AppCompatActivity {
+public class FLExecutor {
     private JSONObject config;
 
     private String mExecutorID;
@@ -57,24 +50,10 @@ public class FLExecutor extends AppCompatActivity {
     private Queue<ServerResponse> eventQueue = new LinkedList<>();
     private Backend backend;
 
-    private TextView mUserId;
-    private TextView mExecuteStatus;
-    private TextView mExecuteResult;
+    private FLApp app;
 
-    HandlerThread mThread;
-    Handler mHandle;
-
-    /**
-     * Initialization of the User Interface.
-     * Three lines of UI: executor id, status, result (not changed)
-     */
-    private void initUI() {
-        runOnUiThread(()->{
-            String userIdMsg = this.mExecutorID + ": Round " + this.round;
-            this.mUserId.setText(userIdMsg);
-            this.mExecuteStatus.setText(Common.CLIENT_CONNECT);
-            this.mExecuteResult.setText("");
-        });
+    public FLExecutor(FLApp app) {
+        this.app = app;
     }
 
     /**
@@ -121,7 +100,7 @@ private void initControlCommunication() {
      */
     private void initData() throws IOException {
         Log.i(Common.TAG, "Data movement starts ...");
-        Common.copyDir(getBaseContext(), "dataset", getCacheDir());
+        Common.copyDir(app.getBaseContext(), "dataset", app.getCacheDir());
         Log.i(Common.TAG, "Data movement completes ...");
     }
 
@@ -130,7 +109,7 @@ private void initData() throws IOException {
      */
     private void initAsset() throws Exception {
         this.config = new JSONObject(Common.readFile(
-                getBaseContext().getAssets().open("conf.json")).toString());
+                app.getBaseContext().getAssets().open("conf.json")).toString());
         this.mExecutorID = this.initExecutorId(this.config.getString("username"));
         String aggregatorIP = this.config.getJSONObject("aggregator").getString("ip");
         int aggregatorPort = this.config.getJSONObject("aggregator").getInt("port");
@@ -147,12 +126,10 @@ private void initAsset() throws Exception {
      * Start running the executor by setting up execution and communication environment,
      * and monitoring the grpc message.
      */
-    private void runExecutor() throws Exception {
+    public void initExecutor() throws Exception {
         this.initData();
         this.initAsset();
-        this.initUI();
-        this.setupCommunication();
-        this.eventMonitor();
+        this.app.initUI(this.config.getString("username"));
     }
 
     /**
@@ -193,13 +170,15 @@ private ByteString serializeResponse(Map<String, Object> responses) throws IOExc
      * @param model The broadcast global model config.
      */
     public void FLUpdateModel(byte[] model) throws JSONException, IOException {
+        this.app.onWriteModel();
         this.round++;
-        this.setText(this.mExecuteStatus, Common.UPDATE_MODEL);
-        this.setText(this.mUserId, this.mExecutorID + ": Round " + this.round);
+        this.app.onChangeStatus(Common.UPDATE_MODEL);
+        this.app.onChangeRound(this.round);
         final String fileName = this.config.getJSONObject("model_conf").getString("path");
-        final String modelPath = getCacheDir() + "/" + fileName;
+        final String modelPath = app.getCacheDir() + "/" + fileName;
         InputStream is = new ByteArrayInputStream(model);
         Common.inputStream2File(is, modelPath);
+        this.app.onGetModel();
     }
 
     /**
@@ -208,12 +187,12 @@ public void FLUpdateModel(byte[] model) throws JSONException, IOException {
      * @param config The client training config.
      */
     public void FLTrain(Map<String, Object> config) throws Exception {
-        this.setText(this.mExecuteStatus, Common.CLIENT_TRAIN);
+        this.app.onChangeStatus(Common.CLIENT_TRAIN);
         JSONObject newTrainingConf = this.overrideConf(
                 this.config.getJSONObject("training_conf"),
                 config);
         Map<String, Object> trainResult = this.backend.MLTrain(
-                getCacheDir().toString(),
+                this.app.getCacheDir().toString(),
                 this.config.getJSONObject("model_conf").getString("path"),
                 this.config.getJSONObject("training_data"),
                 newTrainingConf);
@@ -228,18 +207,31 @@ public void FLTrain(Map<String, Object> config) throws Exception {
         this.sendRequest(() -> this.communicator.stub.cLIENTEXECUTECOMPLETION(request));
     }
 
+    /**
+     * Load train config and data to start training on that client without connecting to the cloud.
+     */
+    public void LocalTrain() throws Exception {
+        this.app.onChangeStatus(Common.CLIENT_TRAIN_LOCALLY);
+        Map<String, Object> trainResult = this.backend.MLTrain(
+                this.app.getCacheDir().toString(),
+                this.config.getJSONObject("model_conf").getString("path"),
+                this.config.getJSONObject("training_data"),
+                this.config.getJSONObject("training_conf"));
+        this.app.onChangeStatus(Common.CLIENT_TRAIN_LOCALLY_FIN);
+    }
+
     /**
      * Model Testing. By default, we test the accuracy on all data of clients in the test group
      *
      * @param config The client testing config.
      */
     public void FLTest(Map<String, Object> config) throws Exception {
-        this.setText(this.mExecuteStatus, Common.MODEL_TEST);
+        this.app.onChangeStatus(Common.MODEL_TEST);
         JSONObject newTestingConf = this.overrideConf(
                 this.config.getJSONObject("testing_conf"),
                 config);
         Map<String, Object> testResult = this.backend.MLTest(
-                getCacheDir().toString(),
+                app.getCacheDir().toString(),
                 this.config.getJSONObject("model_conf").getString("path"),
                 this.config.getJSONObject("testing_data"),
                 newTestingConf);
@@ -256,11 +248,20 @@ public void FLTest(Map<String, Object> config) throws Exception {
         this.sendRequest(() -> this.communicator.stub.cLIENTEXECUTECOMPLETION(request));
     }
 
+    /**
+     * Start the current executor
+     */
+    public void FLStart() throws Exception {
+        this.app.onChangeStatus(Common.CLIENT_CONNECT);
+        this.setupCommunication();
+        this.eventMonitor();
+    }
+
     /**
      * Stop the current executor
      */
     public void FLStop() throws InterruptedException {
-        this.setText(this.mExecuteStatus, Common.SHUT_DOWN);
+        this.app.onChangeStatus(Common.SHUT_DOWN);
         this.communicator.CloseServerConnection();
         this.receivedStopRequest = true;
     }
@@ -311,7 +312,7 @@ private JSONObject overrideConf(JSONObject oldConfJSON, Map<String, Object> newC
      * Register the executor information to the aggregator.
      */
     private void clientRegister() throws Exception {
-        this.setText(this.mExecuteStatus, "Registering");
+        this.app.onChangeStatus("Registering");
         RegisterRequest request = RegisterRequest.newBuilder()
                 .setExecutorId(this.mExecutorID)
                 .setClientId(this.mExecutorID)
@@ -325,7 +326,7 @@ private void clientRegister() throws Exception {
      * Ping the aggregator for new task.
      */
     private void clientPing() throws InterruptedException {
-        this.setText(this.mExecuteStatus, "Pinging");
+        this.app.onChangeStatus("Pinging");
         PingRequest request = PingRequest.newBuilder()
                 .setClientId(this.mExecutorID)
                 .setExecutorId(this.mExecutorID).build();
@@ -379,67 +380,4 @@ private void sendRequest(MessageProcessor msgProcessor) throws InterruptedExcept
             }
         }
     }
-
-    /**
-     * Set text displayed on User Interface.
-     *
-     * @param text The associated TextView object.
-     * @param value The value of the updated text.
-     */
-    private void setText(final TextView text,final String value){
-        runOnUiThread(() -> text.setText(value));
-    }
-
-    @Override
-    protected void onCreate(Bundle savedInstanceState) {
-        Log.i("OnCreate", "");
-        super.onCreate(savedInstanceState);
-        getWindow().setFlags(WindowManager.LayoutParams.FLAG_KEEP_SCREEN_ON,
-                WindowManager.LayoutParams.FLAG_KEEP_SCREEN_ON);
-        getWindow().addFlags(WindowManager.LayoutParams.FLAG_FULLSCREEN);
-        setContentView(R.layout.activity_main);
-
-        this.mUserId = findViewById(R.id.userId);
-        this.mExecuteStatus = findViewById(R.id.executeStatus);
-        this.mExecuteResult = findViewById(R.id.executeResult);
-
-        mThread = new HandlerThread("MNNTrain");
-        mThread.start();
-        mHandle = new Handler(mThread.getLooper());
-
-        mHandle.post(() -> {
-            try {
-                runExecutor();
-            } catch (Exception e) {
-                e.printStackTrace();
-            }
-        });
-    }
-
-    @Override
-    protected void onPause() {
-        Log.i("OnPause", "");
-        super.onPause();
-    }
-
-    @Override
-    protected void onResume() {
-        Log.i("OnResume", "");
-        super.onResume();
-    }
-
-    @Override
-    public void onBackPressed() {
-        Log.i("onBackPressed", "");
-        Intent home = new Intent(Intent.ACTION_MAIN);
-        home.addCategory(Intent.CATEGORY_HOME);
-        startActivity(home);
-    }
-
-    @Override
-    protected void onDestroy() {
-        Log.i("OnDestroy", "");
-        mThread.quitSafely();
-        super.onDestroy();
-    }
 }
diff --git a/fedscale/edge/android/app/src/main/java/com/fedscale/android/utils/Common.java b/fedscale/edge/android/app/src/main/java/com/fedscale/android/utils/Common.java
index df1410be..b8642532 100644
--- a/fedscale/edge/android/app/src/main/java/com/fedscale/android/utils/Common.java
+++ b/fedscale/edge/android/app/src/main/java/com/fedscale/android/utils/Common.java
@@ -29,6 +29,8 @@ public class Common {
     public static String START_ROUND = "start_round";
     public static String CLIENT_CONNECT = "client_connect";
     public static String CLIENT_TRAIN = "client_train";
+    public static String CLIENT_TRAIN_LOCALLY = "client_train_locally";
+    public static String CLIENT_TRAIN_LOCALLY_FIN = "client_train_locally_finish";
     public static String DUMMY_EVENT = "dummy_event";
     public static String UPLOAD_MODEL = "upload_model";
 
diff --git a/fedscale/edge/android/app/src/main/jni/include/converter/include/cli.hpp b/fedscale/edge/android/app/src/main/jni/include/converter/include/cli.hpp
deleted file mode 100644
index 11ea6020..00000000
--- a/fedscale/edge/android/app/src/main/jni/include/converter/include/cli.hpp
+++ /dev/null
@@ -1,30 +0,0 @@
-//
-//  cli.hpp
-//  MNNConverter
-//
-//  Created by MNN on 2019/01/31.
-//  Copyright © 2018, Alibaba Group Holding Limited
-//
-
-#ifndef CLI_HPP
-#define CLI_HPP
-
-#include <iostream>
-#include "config.hpp"
-namespace MNN {
-class MNN_PUBLIC Cli {
-public:
-    static bool initializeMNNConvertArgs(modelConfig &modelPath, int argc, char **argv);
-    static bool convertModel(modelConfig& modelPath);
-    static int testconvert(const std::string& defaultCacheFile, const std::string& directName, float maxErrorRate);
-    static bool mnn2json(const char* modelFile, const char* jsonFile, int flag = 3);
-    static bool json2mnn(const char* jsonFile, const char* modelFile);
-};
-};
-
-class CommonKit {
-public:
-    static bool FileIsExist(std::string path);
-};
-
-#endif // CLI_HPP
diff --git a/fedscale/edge/android/app/src/main/jni/include/converter/include/config.hpp b/fedscale/edge/android/app/src/main/jni/include/converter/include/config.hpp
deleted file mode 100644
index 6c13c666..00000000
--- a/fedscale/edge/android/app/src/main/jni/include/converter/include/config.hpp
+++ /dev/null
@@ -1,63 +0,0 @@
-//
-//  config.hpp
-//  MNNConverter
-//
-//  Created by MNN on 2019/01/31.
-//  Copyright © 2018, Alibaba Group Holding Limited
-//
-
-#ifndef CONFIG_HPP
-#define CONFIG_HPP
-#include <string>
-#include <MNN/MNNDefine.h>
-
-class MNN_PUBLIC modelConfig {
-public:
-    modelConfig()
-        : MNNModel(),
-          prototxtFile(),
-          modelFile(),
-          bizCode("MNN"),
-          model(modelConfig::MAX_SOURCE),
-          benchmarkModel(false),
-          saveHalfFloat(false){
-    }
-    enum MODEL_SOURCE { TENSORFLOW = 0, CAFFE, ONNX, MNN, TFLITE, TORCH, JSON, MAX_SOURCE };
-
-    // MNN model path
-    std::string MNNModel;
-    // if model is tensorflow, this value is NULL;
-    std::string prototxtFile;
-    // tensorflow pb, or caffe model
-    std::string modelFile;
-    // bizCode
-    std::string bizCode;
-    // input config file
-    std::string inputConfigFile;
-    // model source
-    MODEL_SOURCE model;
-    bool benchmarkModel;
-    bool saveHalfFloat;
-    bool forTraining = false;
-    int weightQuantBits = 0;// If weightQuantBits > 0, it means the bit
-    bool weightQuantAsymmetric = false;
-    // The path of the model compression file that stores the int8 calibration table
-    // or sparse parameters.
-    std::string compressionParamsFile = "";
-    bool saveStaticModel = false;
-    int optimizePrefer = 0;
-    float targetVersion = 1.3;
-    int defaultBatchSize = 0;
-    int optimizeLevel = 1;
-    bool keepInputFormat = false;
-    bool alignDenormalizedValue = true;
-    bool detectSparseSpeedUp = true;
-    std::string customOpLibs = "";
-    std::string authCode = "";
-    std::string testDir = "";
-    float testThredhold = 0.01;
-    bool mnn2json = false;
-    bool dumpInfo = false;
-};
-
-#endif // CONFIG_HPP
diff --git a/fedscale/edge/android/app/src/main/jni/mnntrainnative.cpp b/fedscale/edge/android/app/src/main/jni/mnntrainnative.cpp
index 8187a152..fd0a2215 100644
--- a/fedscale/edge/android/app/src/main/jni/mnntrainnative.cpp
+++ b/fedscale/edge/android/app/src/main/jni/mnntrainnative.cpp
@@ -21,7 +21,6 @@
 #include "train/source/transformer/Transformer.hpp"
 #include "cpp/ConvertToFullQuant.hpp"
 #include "module/PipelineModule.hpp"
-#include "converter/include/cli.hpp"
 
 using namespace MNN;
 using namespace MNN::Express;
@@ -61,7 +60,6 @@ Java_com_fedscale_android_mnn_MNNNative_nativeTrain(
     std::string trainJsonPath       = directory + "/" + modelFilename;
     std::string trainModelPath      = directory + "/model.mnn";
     std::string newModelPath        = directory + "/temp_model.mnn";
-    std::string newJsonPath         = directory + "/temp_json.mnn";
 
     auto clientId = static_cast<jstring>(env->CallObjectMethod(trainingConf_, getString, env->NewStringUTF("client_id")));
     const int numClasses        = env->CallIntMethod(trainingConf_, getInt, env->NewStringUTF("num_classes"));
@@ -76,9 +74,6 @@ Java_com_fedscale_android_mnn_MNNNative_nativeTrain(
     const float momentum        = env->CallDoubleMethod(trainingConf_, getDouble, env->NewStringUTF("momentum"));
     const float weight_decay    = env->CallDoubleMethod(trainingConf_, getDouble, env->NewStringUTF("weight_decay"));
 
-    MNN::Cli::json2mnn(trainJsonPath.c_str(), trainModelPath.c_str());
-    MNN_PRINT("[JSON->MNN][JSON][%s][MNN][%s]", trainJsonPath.c_str(), trainModelPath.c_str());
-
     auto varMap = Variable::loadMap(trainModelPath.c_str());
     if (varMap.empty()) {
         MNN_ERROR("Can not load model %s\n", trainModelPath.c_str());
@@ -176,10 +171,7 @@ Java_com_fedscale_android_mnn_MNNNative_nativeTrain(
         exe->dumpProfile();
     }
 
-    MNN::Cli::mnn2json(newModelPath.c_str(), newJsonPath.c_str());
-    MNN_PRINT("[MNN->JSON][MNN][%s][JSON][%s]", newModelPath.c_str(), newJsonPath.c_str());
-
-    std::ifstream t(newJsonPath);
+    std::ifstream t(newModelPath);
     std::stringstream buffer;
     buffer << t.rdbuf();
 
@@ -234,10 +226,6 @@ Java_com_fedscale_android_mnn_MNNNative_nativeTest(
     const int testBatchSize     = env->CallIntMethod(testingConf_, getInt, env->NewStringUTF("test_bsz"));
     const int testNumWorkers    = env->CallIntMethod(testingConf_, getInt, env->NewStringUTF("num_workers"));
 
-    MNN_PRINT("[JSON->MNN][start][JSON][%s][MNN][%s]", testJsonPath.c_str(), testModelPath.c_str());
-    MNN::Cli::json2mnn(testJsonPath.c_str(), testModelPath.c_str());
-    MNN_PRINT("[JSON->MNN][end][JSON][%s][MNN][%s]", testJsonPath.c_str(), testModelPath.c_str());
-
     auto varMap = Variable::loadMap(testModelPath.c_str());
     if (varMap.empty()) {
         MNN_ERROR("Can not load model %s\n", testModelPath.c_str());
diff --git a/fedscale/edge/android/app/src/main/jniLibs/arm64-v8a/libMNNConvertDeps.so b/fedscale/edge/android/app/src/main/jniLibs/arm64-v8a/libMNNConvertDeps.so
deleted file mode 100644
index bf5822a0..00000000
Binary files a/fedscale/edge/android/app/src/main/jniLibs/arm64-v8a/libMNNConvertDeps.so and /dev/null differ
diff --git a/fedscale/edge/android/app/src/main/jniLibs/armeabi-v7a/libMNNConvertDeps.so b/fedscale/edge/android/app/src/main/jniLibs/armeabi-v7a/libMNNConvertDeps.so
deleted file mode 100644
index 509f00a9..00000000
Binary files a/fedscale/edge/android/app/src/main/jniLibs/armeabi-v7a/libMNNConvertDeps.so and /dev/null differ
diff --git a/fedscale/edge/android/app/src/main/res/layout/activity_main.xml b/fedscale/edge/android/app/src/main/res/layout/activity_main.xml
index edf638c9..d1b3bfbf 100644
--- a/fedscale/edge/android/app/src/main/res/layout/activity_main.xml
+++ b/fedscale/edge/android/app/src/main/res/layout/activity_main.xml
@@ -9,7 +9,6 @@
         android:layout_width="match_parent"
         android:layout_height="wrap_content"
         android:layout_alignParentBottom="true"
-        android:layout_marginBottom="7dp"
         android:background="#ffffff"
         android:orientation="vertical">
 
@@ -17,24 +16,46 @@
             android:id="@+id/userId"
             android:layout_width="match_parent"
             android:layout_height="wrap_content"
-            android:layout_marginLeft="10dp"
-            android:layout_marginTop="6dp"
             android:gravity="center_vertical"
-            android:textSize="18dp"
-            android:textStyle="bold" />
+            android:textSize="18sp" />
+
+        <TextView
+            android:id="@+id/executeRound"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:gravity="center_vertical"
+            android:textSize="18sp" />
 
         <TextView
             android:id="@+id/executeStatus"
             android:layout_width="match_parent"
             android:layout_height="wrap_content"
-            android:layout_marginLeft="10dp"
-            android:textSize="18dp" />
+            android:textSize="18sp" />
 
         <TextView
             android:id="@+id/executeResult"
             android:layout_width="match_parent"
             android:layout_height="wrap_content"
-            android:layout_marginLeft="10dp"
-            android:textSize="18dp" />
+            android:textSize="18sp" />
+
+        <TextView
+            android:id="@+id/executeMsg"
+            android:text="@string/no_model"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:textSize="18sp" />
+
+        <Button
+            android:id="@+id/flButton"
+            android:text="@string/start_fl"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content" />
+
+        <Button
+            android:id="@+id/finetuneButton"
+            android:text="@string/start_finetune"
+            android:enabled="false"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content" />
     </LinearLayout>
 </RelativeLayout>
\ No newline at end of file
diff --git a/fedscale/edge/android/app/src/main/res/values/strings.xml b/fedscale/edge/android/app/src/main/res/values/strings.xml
index 1423977b..8cb8e7a6 100644
--- a/fedscale/edge/android/app/src/main/res/values/strings.xml
+++ b/fedscale/edge/android/app/src/main/res/values/strings.xml
@@ -1,4 +1,10 @@
 <resources>
-    <string name="app_name">FedScale Demo MNN</string>
+    <string name="app_name">FedScale Demo</string>
     <string name="appbar_scrolling_view_behavior" translatable="false">android.support.design.widget.AppBarLayout$ScrollingViewBehavior</string>
+    <string name="start_fl">Start FL</string>
+    <string name="stop_fl">Stop FL</string>
+    <string name="start_finetune">Start Local Fine-tune</string>
+    <string name="stop_finetune">Stop Local Fine-tune</string>
+    <string name="no_model">You may not start local fine-tuning because there is no model locally. Please connect to the server first.</string>
+    <string name="has_model">You may now start local fine-tuning. If the FL executor is running, you should stop it first.</string>
 </resources>