File size: 5,118 Bytes
a849f56 5ef6acf 064ce34 5ef6acf 3251ca5 93164d1 3251ca5 064ce34 3251ca5 064ce34 93164d1 3251ca5 93164d1 3251ca5 93164d1 3251ca5 93164d1 3251ca5 93164d1 3251ca5 93164d1 3251ca5 064ce34 93164d1 ed1bf7e 064ce34 4bb8874 9c192dd 4bb8874 ed1bf7e 93164d1 064ce34 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 |
---
license: mit
language:
- en
metrics:
- accuracy
results:
- task:
type: image-classification
metrics:
- name: Accuracy
type: VGG
value: 0.725
- name: Accuracy
type: Resnet18
value: 0.901
pipeline_tag: image-classification
tags:
- tanks
---
# Russian Tank Destroyed or Not Classifier
Binary Image Classifier for whether a tank is destroyed or not using images from <a href = https://www.oryxspioenkop.com/2022/02/attack-on-europe-documenting-equipment.html>Oryx's</a> collection of confirmed Russian armor losses.
Two models were trained:
1. VGG
2. Finetuned Resnet18
## Model Description and Architecture
Models were trained in Pytorch. VGG was built from scratch while resnet18 was finetuned on pytorch's base resnet18 model with weights from IMAGENET1K_V1. Documentation for pytorch's resnet18 can be found <a href = https://pytorch.org/vision/main/models/generated/torchvision.models.resnet18.html>here</a>.
### VGG Model Architecture:
```
VGG(
(features): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
(3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(4): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(6): ReLU(inplace=True)
(7): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(8): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(9): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(10): ReLU(inplace=True)
(11): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(12): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(13): ReLU(inplace=True)
(14): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(avgpool): AdaptiveAvgPool2d(output_size=(5, 5))
(classifier): Sequential(
(0): Linear(in_features=3200, out_features=512, bias=True)
(1): ReLU(inplace=True)
(2): Dropout(p=0.3, inplace=False)
(3): Linear(in_features=512, out_features=256, bias=True)
(4): ReLU(inplace=True)
(5): Dropout(p=0.3, inplace=False)
(6): Linear(in_features=256, out_features=1, bias=True)
(7): Sigmoid()
)
)
```
## Training Details
### Dataset:
Training data was sourced from a <a href = https://www.kaggle.com/datasets/piterfm/2022-ukraine-russia-war-equipment-losses-oryx>kaggle </a>dataset compiling images from Oryx's site. Only Russian tank images were used for training.
Image filenames in the original dataset contain a list of tags: ['destroyed', 'captured', 'abandoned']. Some images have multiple tags as a result of a vehicle being for example both abandoned then captured, or as a result of an image having multiple vehicles. Images with multiple vehicles or no tags were excluded. Single vehicle images with multiple tags were considered not destroyed if the destroyed tag was not present in the image filename.
After data cleaning and labeling, there were 487 destroyed tanks and 228 not destroyed tanks. These images were split 85/15 into a training and validation set. Imbalanced class representation was by using a weighted random sampler that overweighted the minority "not destroyed" class.
Images were resized and manipulated using the following transformation:
```
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])
composed_transform = transforms.Compose([transforms.Resize((256, 256)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean, std)])
```
### Model Hyperparameters:
VGG:
* Batch Size = 16
* Epochs = 20
* lr = 1e-3
* dropout_rate = 0.3
* Optimizer = Adam
* Loss = BCELoss
Resnet18:
* Batch Size = 16
* Epochs = 46
* lr = 1e-3
* Optimizer = Adam
* Loss = CrossEntropyLoss
### Results:
VGG: 73% Accuracy Validation set
Resnet18: 90% Accuracy Validaton set
Has limited ability to correctly recognize the states of vehicles outside of the training scope.
### Limitations:
Doesn't have the ability to recognize whether the provided image is a tank or not.
<img src="https://huggingface.co./Dingaling01/russian_tank_destroyed_or_not_cnn/resolve/main/brownie.png" alt="drawing" width="250"/>
|