[repo in process]
This project aims to detect and recognize MNIST digit. See the RoadMap notebook if needed.
I randomly pasted each 28x28 digit in a 75x75 black background and trained a Darknet-like model to reproduce a YoloV1 algorithm on MNIST. The model handle a unique bounding box and tries to localize and classify a digit with a 6x6 grid.
The model has been trained on 10 epochs, on the MNIST training set.
The mAP has not been produced since it's a first approach on YoloV1 but for your information, the model achieved a 97.23% validation class accuracy and a validation MSE of 0.14 on confidence score (confidence of detecting an object within a bounding box).
==========================================================================================
YoloMNIST [64, 6, 6, 5] --
├─Sequential: 1-1 [64, 128, 6, 6] --
│ └─CNNBlock: 2-1 [64, 32, 37, 37] 1,632
│ └─MaxPool2d: 2-2 [64, 32, 18, 18] --
│ └─CNNBlock: 2-3 [64, 128, 16, 16] 37,120
│ └─MaxPool2d: 2-4 [64, 128, 8, 8] --
│ └─CNNBlock: 2-5 [64, 64, 8, 8] 8,320
│ └─CNNBlock: 2-6 [64, 128, 6, 6] 73,984
│ └─CNNBlock: 2-7 [64, 128, 6, 6] 147,710
├─Sequential: 1-2 [64, 540] --
│ └─Linear: 2-9 [64, 4096] 18,878,464
│ └─LeakyReLU: 2-10 [64, 4096] --
│ └─Linear: 2-11 [64, 540] 2,212,380
==========================================================================================
Trainable params: 21,359,612
After understanding what's under the hood of YOLO, this project aims to detect and recognize various objects on meal trays from my former company restaurant. Make sure to see the [RoadMap] if needed. \
The dataset and its details are available in their own repository.
The goal is to localize and classify objects among 8 categories :
- Plate (the main dish)
- Starter
- Dessert
- Bread
- Yogurt
- Fruit
- Drink (soda can, juice can, water bottle...)
- Cheese
Since I do not have a large dataset, I took advantage of pretrained ResNet 152 to custruct a yoloResNet model.
===============================================================================================
YoloResNet [16, 7, 7, 18] --
├─Sequential: 1-1 [16, 512, 56, 56] --
├─Sequential: 1-2 [16, 512, 7, 7] --
│ └─MaxPool2d: 2-7 [16, 512, 28, 28] --
│ └─Conv2d: 2-8 [16, 1024, 28, 28] 4,719,616
│ └─Conv2d: 2-9 [16, 512, 28, 28] 524,800
│ └─MaxPool2d: 2-10 [16, 512, 14, 14] --
│ └─Conv2d: 2-11 [16, 1024, 14, 14] 4,719,616
│ └─Conv2d: 2-12 [16, 512, 14, 14] 524,800
│ └─MaxPool2d: 2-13 [16, 512, 7, 7] --
├─Sequential: 1-3 [16, 882] --
│ └─Flatten: 2-14 [16, 25088] --
│ └─Linear: 2-15 [16, 4096] 102,764,544
│ └─LeakyReLU: 2-16 [16, 4096] --
│ └─Dropout: 2-17 [16, 4096] --
│ └─Linear: 2-18 [16, 882] 3,613,554
│ └─Sigmoid: 2-19 [16, 882] --
===============================================================================================
Total params: 119,432,114
Trainable params: 116,866,930
Non-trainable params: 2,565,184