README.md

metadata

title: Cifar10
emoji: 🔥
colorFrom: purple
colorTo: indigo
sdk: gradio
sdk_version: 4.38.1
app_file: app.py
fullWidth: true
models:
  - >-
    https://huggingface.co/spaces/Shilpaj/cifar10/blob/main/epoch%3D23-step%3D2112.ckpt
datasets:
  - CIFAR10
pinned: false
license: mit

Model Trained for CIFAR10

• This Application demonstrate the inference side of the model trained on the CIFAR dataset
• David C's model architecture is recreated and trained on CIFAR10 dataset to achieve the accuracy of 90+% within 24 epochs
• Once Cycle Policy is used to speed up the training process
• The model is coded using PyTorch Lightning. Mentioned below is the link for Training Repository where you can check the code for model training and tracked metrics

Training Repo Link

• After the Training, model checkpoint is stored on the system and uploaded to Gradio Spaces. Attached below is the link to download model file

  Download Model File

• This app has four features in four tabs:

• GradCam:

  • To visualize which portion of the image model is actually looking at while inferencing on the image
  • Using this, we can come up with an augmentation strategy that can improve the model accuracy

• Misclassified Image:

  • While training the model, though the test accuracy was 90+%, there were still 10% images which were misclassified
  • The feature helps to visualize those images with their respective correct and incorrect labels
  • This can be used to come up with a strategy to improve accuracy for a particular class

• Feature Map Visualization:

  • There are 6 block in this model
  • Each block has 2 or 3 convolutional layers
  • The output of specific kernel are visualized for first convolutional layer of all 6 blocks

• Kernel Visualization:

  • In the first layer of each of the six blocks, there are kernels which are creating the feature maps
  • Some of those kernels are visualized in this section

Dependencies:

• Python Version: 3.x
• PyTorch Lightning: 2.0.6

Usage:

GradCam

• Upload an image or select from example
• Select how much percentage of original image should be overlapped on what actually model is looking at in the image
• Select the number of top classes you want to see
• Select the block number for which first convolutional layer's activation you want to see
• Click on the Submit button to see the results

Misclassified Images

• Select the number of misclassified images you want to see
• Click on Display Misclassified Images to show the images in the center and their respective correct and misclassified labels in the sequence

Feature Map Visualization

• Upload an image
• Select the kernel number
• Click on Visualize FeatureMaps to see the feature maps created by the kernel number for first convolutional layer in each of the six blocks

Kernel Visualization

• Select the block number from which first convolutional layer's kernel to be visualized
• Click on Visualize Kernels to see the kernels