This repository compares the performace of two deep learning architectures, DistilBert transformer and LSTM, for classification on the Food 101-Dataset. The Food-101 dataset comprises 250 images and the corresponding captions of each of the 101 food dishes. For the experiemnts here, we do not use the images, only text captions of the images. The text files, train_titles.csv and test_titles.csv, are provided in the repository.
There are two python code files in this repository, which are explained below:
1. distilBertClassifier.ipynb: python notebook that imports the pre-trained distilbert transfomer from huggingface, fine-tunes the weights, and tests the fine-tuned model.
2. lstmClassifier.ipynb: python notebok that trains and tests the lstm + feedfoward network using pytorch.
The test and train dataset, train_titles.csv and test_tiles.csv, are also included in the repsitory.
Please email any questions to aabbasi1@iastae.edu
- The distilbert clasifier achieves an expected 86 percent classification accuracy on the text Food 101 Dataset.
-
A good sanity check is to inspect a single data point output from the data loaders by decoding the output token ids and reading the decoded sentence to see if it makes sense. For example,
Encoded Text = tensor([ 101, 3313, 6207, 11345, 2007, 9781, 4168, 2389, 19116, 17974, 1064, 2026, 2890, 6895, 10374, 1012, 4012, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]), torch.LongTensor, torch.Size([128])Decoded tokens from encoded ids: '[CLS] double apple pie with cornmeal crust recipe | myrecipes. com [SEP] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD]'Notice how the decoded sentence has special tokens like [CLS], [SEP], [PAD] that indicate the Start-of-Sentece/Classification token, the Separator/End-of-sentence token and the padding token [PAD].