Welcome to LetsLockIn. This website is a GREAT tool for the ones who want to lock-in and focus on their work.
If you are seeking a tool that helps you focus, you are in the right place.
Now, Let's Lock In!
For more information, explore our official website:
letslockin.dev
- Scroll down for features
Our mission is to create a healthier and more productive work environment by leveraging advanced AI models to monitor focus, posture, and fatigue in real time—all while respecting user privacy.
- Not compatible with old versions of Chromium-based browsers (Chrome, Edge, etc.).
- Only compatible with the latest versions.
PLEASE CLICK THE UNMUTE BUTTON AT THE BOTTOM RIGHT CORNER OF THE DEMO VIDEO.
- Reload the page if the video does not show up.
demo2.mp4
- Forces good postures ==> Makes user lock-in.
- Significantly boost productivity.
- Tracks real-time focus and productivity trends
- Prevents burnout through early detection of overworking
- Increases concentration with actionable feedback
- Encourages better posture to prevent back and neck strain
- Prevents overworking
- Promotes healthier work-rest balance
- Clean, simple, and modern design
- All processing happens locally on your device
- No sensitive data leaves your browser, ensuring privacy
- Framework: TensorFlow 2.15.0, tensorflow-decision-forests 1.8.0
⚠️ Important: TensorFlow.js conversion is only supported for models trained with TensorFlow 2.15.0 or lower (I trained the model using Tensorflow 2.15.0). Newer versions such as 2.17.0 are not compatible for deployment. - Architecture: ResNet-based deep neural network
- Dataset Details:
- Training: AffectNet, RAF-DB, CK+
- Verification: Filtered subsets of the same datasets
- Challenges Addressed:
- Overfitting mitigated using dropout and data augmentation.
- Gradient vanishing and exploding resolved with residual connections.
- Optimized for real-time web deployment by reducing model size and complexity.
- v2_afrafck_model:
-
Short for version 2_affectnet_Raf-DB_CK+_model
-
Includes:
- Training ipynb and validating ipynb files
- accuracy reports,
- normalized accuracy reports,
- confusion matrices,
- normalized confusion matrices,
- Paramerters report, log, final .h5 model,
- final model architecture (json)
-
IMPORTANT FILES (SHOULD READ):
letslockin_train.ipynb confusion_matrix.ipynb Norm_RAF_DB_accuracy_report.txt (normalized RAF-DB validation accuracy report) Norm_RAF_DB_confusion_matrix.png model_performance.png
-
- tf2js_model:
- Short for Tensorflow to JS model
- Includes:
- .Bin file
- .Json file
- HTML, CSS, SCRIPT
Our Complex Residual Network consists of 41 carefully designed layers:
flowchart TD
input1[Input Layer 160x160x1] --> conv2d[Conv2D 32]
conv2d --> bn1[BatchNorm]
bn1 --> act1[ReLU]
act1 --> maxpool[MaxPool2D]
maxpool --> drop1[Dropout 0.5]
%% First Block
drop1 --> conv2d_1[Conv2D_1 64]
drop1 --> conv2d_3[Conv2D_3 64]
conv2d_1 --> bn2[BatchNorm_1]
bn2 --> act2[ReLU_1]
act2 --> conv2d_2[Conv2D_2 64]
conv2d_2 --> bn3[BatchNorm_2]
conv2d_3 --> bn4[BatchNorm_3]
bn3 & bn4 --> add1[Add_1]
add1 --> act3[ReLU_2]
act3 --> drop2[Dropout 0.5]
%% Second Block
drop2 --> conv2d_4[Conv2D_4 128]
drop2 --> conv2d_6[Conv2D_6 128]
conv2d_4 --> bn5[BatchNorm_4]
bn5 --> act4[ReLU_3]
act4 --> conv2d_5[Conv2D_5 128]
conv2d_5 --> bn6[BatchNorm_5]
conv2d_6 --> bn7[BatchNorm_6]
bn6 & bn7 --> add2[Add_2]
add2 --> act5[ReLU_4]
act5 --> drop3[Dropout 0.5]
%% Third Block
drop3 --> conv2d_7[Conv2D_7 128]
drop3 --> conv2d_9[Conv2D_9 128]
conv2d_7 --> bn8[BatchNorm_7]
bn8 --> act6[ReLU_5]
act6 --> conv2d_8[Conv2D_8 128]
conv2d_8 --> bn9[BatchNorm_8]
conv2d_9 --> bn10[BatchNorm_9]
bn9 & bn10 --> add3[Add_3]
add3 --> act7[ReLU_6]
act7 --> drop4[Dropout 0.5]
%% Final Layers
drop4 --> globalpool[Global AvgPool2D]
globalpool --> dense1[Dense 3072]
dense1 --> act8[ReLU_7]
act8 --> drop5[Dropout 0.5]
drop5 --> dense2[Dense 7 classes]
dense2 --> output[Output Softmax 7 classes]
%% Modern Color Styling
classDef input fill:#4A90E2,stroke:#333,stroke-width:2px,color:white
classDef conv fill:#2C3E50,stroke:#333,stroke-width:2px,color:white
classDef bn fill:#E74C3C,stroke:#333,stroke-width:2px,color:white
classDef act fill:#3498DB,stroke:#333,stroke-width:2px,color:white
classDef pool fill:#F39C12,stroke:#333,stroke-width:2px,color:white
classDef drop fill:#1ABC9C,stroke:#333,stroke-width:2px,color:white
classDef add fill:#9B59B6,stroke:#333,stroke-width:2px,color:white
classDef dense fill:#E67E22,stroke:#333,stroke-width:2px,color:white
classDef output fill:#16A085,stroke:#333,stroke-width:2px,color:white
class input1 input
class conv2d,conv2d_1,conv2d_2,conv2d_3,conv2d_4,conv2d_5,conv2d_6,conv2d_7,conv2d_8,conv2d_9 conv
class bn1,bn2,bn3,bn4,bn5,bn6,bn7,bn8,bn9,bn10 bn
class act1,act2,act3,act4,act5,act6,act7,act8 act
class maxpool,globalpool pool
class drop1,drop2,drop3,drop4,drop5 drop
class add1,add2,add3 add
class dense1,dense2 dense
class output output
- Validation Accuracy: 72.852%
- Real-time Processing: Up to 60 FPS, but I set a limit of 10 FPS when on the tab, 8 FPS when using other tabs or when the browser is minimized to save battery life.
- Browser Compatibility: 98% (Safari is very strict in allowing js to run in the background, so even though I set the FPS to 8 FPS, it may drop under 8 FPS. For other browsers, it works decently without significant battery impact)
# Clone the repository
git clone https://github.com/letslockin/letslockin.github.io.git
# Navigate to project directory
cd letslockin.github.io
# Start local server (Python 3)
python -m http.server (Windows OS)
python3 -m http.server (MacOS)
# Go to your browser and open
http://localhost:8000
⚠️ Important Note: This model must be trained with TensorFlow 2.15.0 or older versions of Tensorflow (2+) as TensorFlow.js conversion is not compatible with newer versions of TensorFlow like version 2.17.0
- High Accuracy Emotion Recognition
- 72.852% accuracy achieved using our 41-layer ResNet architecture
- Validated across multiple datasets
- Extensive Training Data
- 297,074 filtered training images
- 6,568 validation images
- Real-time Performance
- 30 FPS processing rate, 8 FPS background processing rate (could be higher up to 50 FPS but I set 8 FPS to maximize battery life)
- 98% browser compatibility
 Bach Pham AI Engineer @2006coder |
This application is designed to assist with focus and posture improvement. It is not a medical device and should not be used as a substitute for professional medical advice.
Copyright © 2024 LetsLockIn. All rights reserved.
