Winners of the Hackathon , I got 1st Rank
Public surfaces are a possible way for the transmission of COVID-19. NIT Warangal administration has decided to minimize the spread of COVID-19 to ensure the safety of its residents. Elevator switches are public surfaces. So the admin wants to implement a deep learning model for elevators on the campus. The goal is to determine the destination floor of a person standing inside the elevator using hand signs. Accuracy of the model should be at least 90%.
- Original dataset
- 10 classes, for 10 unique hand signs.
- Image size: 100x100
- Color space: RGB
-
Applied data preprocessing
- Reshaped to 64x64
- From RGB to 1 channel grayscale
- Normalization range: -1 to 1
-
Applied data augmentation
- Random horizontal flip (as the hand could be right or left)
- -20 to 20 degrees random rotation with bicubic resample
- Filling with 0.85 * mean of the image (to prevent strong dark background after the random rotation)
-
Training and Validation data split
- About 20% of the data is used for validation.
- For each class in validation, there's total 40 images.
- So in total I've used 400 images for validation.
- And the remaining 1,662 images for training.
-
Training data after preprocessing and augmentation
-
I've followed simple DenseNet architecture, and replaced ReLU to LeakyReLU with alpha of 0.2
-
Here's the model summary:
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 32, 32] 3,200
BatchNorm2d-2 [-1, 64, 32, 32] 128
LeakyReLU-3 [-1, 64, 32, 32] 0
MaxPool2d-4 [-1, 64, 16, 16] 0
BatchNorm2d-5 [-1, 64, 16, 16] 128
LeakyReLU-6 [-1, 64, 16, 16] 0
Conv2d-7 [-1, 32, 16, 16] 18,464
BatchNorm2d-8 [-1, 96, 16, 16] 192
LeakyReLU-9 [-1, 96, 16, 16] 0
Conv2d-10 [-1, 32, 16, 16] 27,680
BatchNorm2d-11 [-1, 128, 16, 16] 256
LeakyReLU-12 [-1, 128, 16, 16] 0
Conv2d-13 [-1, 32, 16, 16] 36,896
BatchNorm2d-14 [-1, 160, 16, 16] 320
LeakyReLU-15 [-1, 160, 16, 16] 0
Conv2d-16 [-1, 32, 16, 16] 46,112
DenseBlock-17 [-1, 192, 16, 16] 0
BatchNorm2d-18 [-1, 192, 16, 16] 384
LeakyReLU-19 [-1, 192, 16, 16] 0
Conv2d-20 [-1, 96, 16, 16] 18,528
AvgPool2d-21 [-1, 96, 8, 8] 0
BatchNorm2d-22 [-1, 96, 8, 8] 192
LeakyReLU-23 [-1, 96, 8, 8] 0
Conv2d-24 [-1, 32, 8, 8] 27,680
BatchNorm2d-25 [-1, 128, 8, 8] 256
LeakyReLU-26 [-1, 128, 8, 8] 0
Conv2d-27 [-1, 32, 8, 8] 36,896
BatchNorm2d-28 [-1, 160, 8, 8] 320
LeakyReLU-29 [-1, 160, 8, 8] 0
Conv2d-30 [-1, 32, 8, 8] 46,112
BatchNorm2d-31 [-1, 192, 8, 8] 384
LeakyReLU-32 [-1, 192, 8, 8] 0
Conv2d-33 [-1, 32, 8, 8] 55,328
DenseBlock-34 [-1, 224, 8, 8] 0
BatchNorm2d-35 [-1, 224, 8, 8] 448
LeakyReLU-36 [-1, 224, 8, 8] 0
Conv2d-37 [-1, 112, 8, 8] 25,200
AvgPool2d-38 [-1, 112, 4, 4] 0
BatchNorm2d-39 [-1, 112, 4, 4] 224
LeakyReLU-40 [-1, 112, 4, 4] 0
Conv2d-41 [-1, 32, 4, 4] 32,288
BatchNorm2d-42 [-1, 144, 4, 4] 288
LeakyReLU-43 [-1, 144, 4, 4] 0
Conv2d-44 [-1, 32, 4, 4] 41,504
BatchNorm2d-45 [-1, 176, 4, 4] 352
LeakyReLU-46 [-1, 176, 4, 4] 0
Conv2d-47 [-1, 32, 4, 4] 50,720
BatchNorm2d-48 [-1, 208, 4, 4] 416
LeakyReLU-49 [-1, 208, 4, 4] 0
Conv2d-50 [-1, 32, 4, 4] 59,936
DenseBlock-51 [-1, 240, 4, 4] 0
BatchNorm2d-52 [-1, 240, 4, 4] 480
LeakyReLU-53 [-1, 240, 4, 4] 0
Conv2d-54 [-1, 120, 4, 4] 28,920
AvgPool2d-55 [-1, 120, 2, 2] 0
BatchNorm2d-56 [-1, 120, 2, 2] 240
LeakyReLU-57 [-1, 120, 2, 2] 0
Conv2d-58 [-1, 32, 2, 2] 34,592
BatchNorm2d-59 [-1, 152, 2, 2] 304
LeakyReLU-60 [-1, 152, 2, 2] 0
Conv2d-61 [-1, 32, 2, 2] 43,808
BatchNorm2d-62 [-1, 184, 2, 2] 368
LeakyReLU-63 [-1, 184, 2, 2] 0
Conv2d-64 [-1, 32, 2, 2] 53,024
BatchNorm2d-65 [-1, 216, 2, 2] 432
LeakyReLU-66 [-1, 216, 2, 2] 0
Conv2d-67 [-1, 32, 2, 2] 62,240
DenseBlock-68 [-1, 248, 2, 2] 0
BatchNorm2d-69 [-1, 248, 2, 2] 496
LeakyReLU-70 [-1, 248, 2, 2] 0
AdaptiveMaxPool2d-71 [-1, 248, 1, 1] 0
Flatten-72 [-1, 248] 0
Linear-73 [-1, 10] 2,490
================================================================
Total params: 758,226
Trainable params: 758,226
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.02
Forward/backward pass size (MB): 6.36
Params size (MB): 2.89
Estimated Total Size (MB): 9.26
----------------------------------------------------------------
- And here's the model visualization:
-
Pytorch:
pip install torch==1.5.1+cu101 torchvision==0.6.1+cu101 -f https://download.pytorch.org/whl/torch_stable.html- Replace
cu101with your cuda version.
-
Pytorch Summary:
pip install torchsummary -
TQDM:
pip install tqdm -
Pillow:
pip install Pillow
-
Training Accuracy : 99.819712
-
Validation Accuracy : 99.609375
-
Validation F1 Score : 0.9989696
-
Training Loss vs Epoch :-
- Training Accuracy vs Validation Accuracy :-
- Per Class Accuracy for Validation:-
- Confusion Matrix for Validation:-
So in validation (out of 400 images), only two images are predicted incorrectly, impressive!
Replace the test_root variable value to the test data directory.