Cifar10 – Resnet

Resnet18 model with CIFAR-10 dataset

This project utilizes the Resnet18 model to perform image classification on the CIFAR-10 dataset. The CIFAR-10 dataset comprises 60,000 32×32 color images distributed across 10 classes, with each class containing 6,000 images.

Using Tensorleap we can save time while exploring and detecting high loss and or unlabeled clusters and samples.

Latent Space Exploration

The following plot illustrates a population exploration map, depicting the similarity among samples based on the latent space of a trained model. This map is constructed using the extracted features of the model.

The visualization shows all the data points, revealing two distinct clusters. These clusters represent images labeled as animals or vehicles according to their ground truth labels. The noticeable separation between the clusters indicates a significant difference between these two categories.

Detecting & Handling High Loss Clusters

After conducting further analysis, it has been observed that the ‘cat’ cluster consists of several samples with higher loss, indicated by larger dot sizes on the plot. A closer examination reveals that a significant number of these samples were incorrectly predicted as ‘dog’ by the model. This suggests that there is a need to include more images of cats and possibly dogs in the training dataset to improve the model’s ability to recognize them accurately.

Detecting & Handling High Loss Unlabeled Clusters

In the k-means clusters, clusters 5 and 2 are observed to be in proximity to each other. Cluster 5 predominantly consists of images with a light background and objects exhibiting an orange shade. On the other hand, cluster number 2 also has a light background, but the objects within it appear in a darker shade.

– cluster 5:

– cluster 2:

fetching similar
An alternative method for identifying clusters in the model’s latent space is to retrieve similar samples based on a selected sample. This approach allows you to pinpoint a cluster that exhibits a specific intrinsic property of interest. The figure presented below illustrates such a cluster, comprising images that feature animal faces captured from a profile perspective.

Sample Loss Analysis

Within this section, we delve into the examination of features that impact the model’s predictions. Tensorleap automatically generates a heatmap that quantifies the effect different features have on the loss.

In the image depicted below, it is evident that the presence of background features leads to a higher loss function. Conversely, the inclusion of features in the nose area contributes to a lower loss function, which signifies the image classification as a cat.

Getting Started with Tensorleap Project

This quick start guide will walk you through the steps to get started with this example repository project.

Prerequisites

Before you begin, ensure that you have the following prerequisites installed:

• Python (version 3.7 or higher).
• Poetry.
• Tensorleap platform access. To request a free trial click here.
• Tensorleap CLI.

Tensorleap CLI Installation

with curl:

curl -s https://raw.githubusercontent.com/tensorleap/leap-cli/master/install.sh | bash

Tensorleap CLI Usage

Tensorleap Login

tensorleap auth login [api key] [api url].

leap project push

include:
    - leap_binder.py
    - cifar10_resnet/configs.py
    - [...]

poetry run test