# Notes on 1st project

The given data at [iui-datalrelease1-sose2021-readonly/\*](/opt/iui-datarelease1-sose2021) represents sensor data from a pen.

| Data   | MinVal | MaxVal | Description                                     |
| ------ |:------:|:------:| ----------------------------------------------- |
| Millis |   -    |   -    | Timestamp from tablet (Unix time)               |
| Acc1 X |   -    |  32768 | Front/Tip accelerometer (Direction: Left/Right) |
| Acc1 Y |   -    |  32768 | Front/Tip accelerometer (Direction: Up/Down)    |
| Acc1 Z |   -    |  32768 | Front/Tip accelerometer (Direction: Back/Front) |
| Acc2 X |   -    |   8192 | Back accelerometer (Direction: Left/Right)      |
| Acc2 Y |   -    |   8192 | Back accelerometer (Direction: Up/Down)         |
| Acc2 Z |   -    |   8192 | Back accelerometer (Direction: Back/Front)      |
| Gyro X |   -    |  32768 | Gyroscope sensor                                |
| Gyro Y |   -    |  32768 | Gyroscope sensor                                |
| Gyro Z |   -    |  32768 | Gyroscope sensor                                |
| Mag X  |   -    |   8192 | Magnetometer                                    |
| Mag Y  |   -    |   8192 | Magnetometer                                    |
| Mag Z  |   -    |   8192 | Magnetometer                                    |
| Force  |   -    |   4096 | Force applied                                   |
| Time   |   -    |   -    | Time from start of "recording"                  |

There were 100 participants.

The folder-structure is as follows:\
`/opt/iui-datarelease1-sose2021/{P}/split_letters_csv/{N}{A}.csv`

| Variable | Description                           | 
| -------- | ------------------------------------- |
|     P    | The ID of the participant             |
|     N    | The N-th letter the participant wrote |
|     A    | The letter that was written           |

Each participants folder contains a `calibration.txt`, which contains the calibration data of the pen for the participant.

Sensor data was recorded at 100hz (100 recordings/s => 1 recording/ms).

## Preprocessing

### General

Since information has different scale (i.e. Acc1: [-32768;32768] and Acc2 [-8192;8192]) the information has to be valued differently based on their importance.

### Millis

- Could be used for identifying each data entry -> needs to be normalized to the first entry of the data set to see the comlete timeline of the data

### Acc1/Acc2/Gyro/Mag

todo

### Force

- Sometimes sensor data was recorded even when there is no action -> we need to determine the area of interest 
  - maybe sliding window, where window avg has to be certain threshold
  - general threshold aproach (filter out data below threshold)
  - more ideas welcome
  
- Data could be normalized by each users relative strength or data entry 

### Time

## Neural Network
This segment are notes dedicated to the neural network itself.

### Ideas

- Don't use batch normalization but normalize by maxval of sensor

### Results

#### Test1 (72.99%)

```python
thresh      = 100
leeway      = 5

model = Sequential()

model.add(Flatten(input_shape=X_filter[0].shape))

model.add(BatchNormalization())

model.add(Dense(1560, activation='relu'))

model.add(Dense(750, activation='relu'))

model.add(Dense(300, activation='relu'))

model.add(Dense(156, activation='relu'))

model.add(Dense(26, activation='softmax'))
```

```
Evaluate on test data
82/82 [==============================] - 0s 3ms/step - loss: 1.4249 - acc: 0.7299
```

#### Test2 (75.21%)

```python
thresh      = 75
leeway      = 3

model = Sequential()

model.add(Flatten(input_shape=X_filter[0].shape))

model.add(BatchNormalization())

model.add(Dense(1560, activation='relu'))

model.add(Dense(750, activation='relu'))

model.add(Dense(300, activation='relu'))

model.add(Dense(156, activation='relu'))

model.add(Dense(26, activation='softmax'))
```

```
Evaluate on test data
82/82 [==============================] - 0s 2ms/step - loss: 1.5145 - acc: 0.7521
```

#### Test3 (36.96%)

```python
thresh      = 10
leeway      = 3

model = Sequential()

model.add(Flatten(input_shape=X_filter[0].shape))

model.add(BatchNormalization())

model.add(Dense(1560, activation='relu'))

model.add(Dense(750, activation='relu'))

model.add(Dense(300, activation='relu'))

model.add(Dense(156, activation='relu'))

model.add(Dense(26, activation='softmax'))
```

```
Evaluate on test data
82/82 [==============================] - 0s 2ms/step - loss: 2.5231 - acc: 0.3696
```

#### Test4 (71.51%)

```python
thresh      = 50
leeway      = 3

model = Sequential()

model.add(Flatten(input_shape=X_filter[0].shape))

model.add(BatchNormalization())

model.add(Dense(1560, activation='relu'))

model.add(Dense(750, activation='relu'))

model.add(Dense(300, activation='relu'))

model.add(Dense(156, activation='relu'))

model.add(Dense(26, activation='softmax'))
```

```
Evaluate on test data
82/82 [==============================] - 0s 2ms/step - loss: 1.5005 - acc: 0.7151
```

#### Test5 ()