A CLI tool that utilizes a ResNet convolutional neural network to recognize content in images and sort them into classes.

Usage: tensort <target_dir> <class_count> [-n | --no-names]

Arguments:s
<target_dir>    : Path to the target directory
<class_count>   : Number of classes
-n, --no-names  : Do not generate class names (optional)

Example:
tensort /path/to/images_dir 5 -n

The recognized image formats consist of: jpg, jpeg, png

The names of the files do describe what the images actually are in this example.

:~$ ls
cbum.jpg                jay1.jpg  jordan_barrett.jpg
cut_off_sleeve_guy.jpg  jay2.jpg  some_bird1.jpeg
golden_retriever.jpg    jay3.jpg  some_bird2.jpg

:~$ ./path/to/bin/tensort /path/to/images 5
Running tensort with options:
<target_dir>        : /path/to/images
<class_count>       : 5
<no_class_names>    : false
            
Neural network running on device: Cuda(0)

Generating image embeddings...
Computing similarities and clustering embeddings...
Averaging tensors and deriving class names...
Moving files...

Results:
Windsor tie (1):
	=> /path/to/images/jordan_barrett.jpg

dumbbell (2):
	=> /path/to/images/cbum.jpg
	=> /path/to/images/cut_off_sleeve_guy.jpg

bulbul (3):
	=> /path/to/images/some_bird2.jpg
	=> /path/to/images/some_bird1.jpeg

jay (4):
	=> /path/to/images/jay3.jpg
	=> /path/to/images/jay2.jpg
	=> /path/to/images/jay1.jpg

golden retriever (5):
	=> /path/to/images/golden_retriever.jpg

:~$ ls
'bulbul (3)'    'golden retriever (5)'  'Windsor tie (1)'
'dumbbell (2)'  'jay (4)'

It is worth noting that any classification problems that do occur are a limitation of the model and could be improved with more training and adding more possible output classifications.

1. Read every image from target_dir and generate an embedding of each image. In the case of the pretrained model used in this application, the embedding is a 1000 dimensional vector representing a probabability distribution of likely classifications.

2. Compute pairwise cosine similarities for each embedding. This was done with the following formula. Let $t_0, t_1$ be vectors in the same dimensional space, then, $$cs(t_0, t_1) = \frac{t_0 \cdot t_1}{||t_0|| \times ||t_1||}$$ which produces a similarity value, $-1 \leq cs(t_0, t_1) \leq 1$. Then, using this formula, pairwise cosine similarities are easily computed to produce the following cartesian relation, $$\langle cs(t_0, t_1), cs(t_0, t_2), ..., cs(t_0, t_n), ..., cs(t_k, t_0), ..., cs(t_k, t_n) \rangle$$ as a vector of similarity values. When extracting the cosine similarity between any two tensors in this vector, the following formula can be used, $$k = i + (j \times c)$$ where $k$ is the index of the target cosine similarity, $i$ is the index of the first tensor, $j$ is the index of the second tensor, and $c$ is the total number of tensors that embed images.

3. Generate similarity thresholds. I am using a heuristic algorithm which uses the calculated pairwise similarity vector and class_count to generate a similarity threshold which will be used to determine if an image belongs in one classification or should be in a new one. This similarity threshold is used to conduct initial class assignments in part 4.

4. Cluster image embeddings. Overall, this algorithm performs an initial assignment of embedding indices to clusters and then optimizes the assignment by finding the best fit for overflowed embedding indices based on cosine similarity.

5. Generate class names. This part can be opted out with the -n | --no-names flags mentioned above. In this part, a tensor averaged along each dimension is generated for each classification and then the classification with the highest probability is selected as the class name.

6. Finally, since everything has now been computed, moving the files into a directory tree that corresponds to the generated classifications is straightforward.

The pretrained model used for this project can be found here.

Setting up the development environment can be done by following the README from the tch-rs repository here

When installing libtorch, ensure that the version that supports CUDA is used.

On Linux, .bashrc should contain the following (at least for my setup)

export LIBTORCH_BYPASS_VERSION_CHECK=1
export LIBTORCH=/path/to/libtorch
export LD_LIBRARY_PATH=/path/to/libtorch:$LD_LIBRARY_PATH