Awesome
VAuLT: Vision-and-Augmented-Language Transformer
Code for VAuLT: Augmenting the Vision-and-Language Transformer For Sentiment Classification on Social Media.
Installation
Experiments were conducted with Python 3.7.4. To install the necessary packages (preferably in a virtual environment):
pip install -e .
If you want to alter the implementation, we suggest you install with:
pip install -e .[dev]
(Note to zsh users: escape [ and ]).
Entity Linking
While our experiments showed no improvement from the implemented form of entity linking, you can use it for the TWITTER-1# datasets by following these instructions (we install their package using pip in setup.py, aka the above commands) and enabling it when running the scripts.
Using VAuLT
from vault.models.vault import VaultModel, VaultProcessor
import requests
from PIL import Image
vilt_model_name_or_path = "dandelin/vilt-b32-mlm"
bert_model_name_or_path = "vinai/bertweet-base"
processor = VaultProcessor.from_pretrained(
vilt_model_name_or_path, bert_model_name_or_path
)
model = VaultModel.from_pretrained(
vilt_model_name_or_path,
bert_model_name_or_path
use_vilt_position_embeddings=False,
)
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
text = "a bunch of [MASK] laying on a [MASK]."
encoding = processor(image, text, return_tensors="pt")
outputs = model(**encoding)
outputs.keys() # odict_keys(['last_hidden_state', 'pooler_output'])
Datasets
TWITTER-1#
Follow the instructions here. We expect the final structure of the dataset to be:
/data/twitter-tmsc/
├──twitter2015
│ ├── dev.tsv
│ ├── dev.txt
│ ├── test.tsv
│ ├── test.txt
│ ├── train.tsv
│ └── train.txt
├──twitter2015_images
│ ├── 996.jpg
│ ├── 997.jpg
│ ├── 998.jpg
│ ├── ...
│ └── 123456789.jpg
├──twitter
│ ├── dev.tsv
. .
. .
. .
but the dataset configuration is configurable within the code (not the given scripts).
MVSA
Find the links here. We expect the final structure of the dataset to be:
/data/mvsa/
├── MVSA
│ ├── corrupt_ids.txt
│ ├── data
│ │ ├── 10000.jpg
│ │ ├── 10000.txt
│ │ ├── 10001.jpg
│ │ ├── 10001.txt
│ │ ├── 10002.jpg
│ │ ├── 10002.txt
│ │ ├── ...
│ │ └── 9998.txt
│ └── labelResultAll.txt
└── MVSA_Single
├── data
│ ├── 1000.jpg
│ ├── 1000.txt
│ ├── 1001.jpg
│ ├── 1001.txt
│ ├── 1002.jpg
│ ├── 1002.txt
│ ├── ...
│ └── 9.txt
└── labelResultAll.txt
where corrupt_ids.txt contains 3151, 3910 and 5995 in separate lines.
Bloomberg
Download annotations and text here and the images here (based on this). We expect the final structure of the dataset to be:
/data/bloomberg-twitter-text-image/
├── bloomberg-textimage.csv
└── Twitter_images
├── T1011681837276639236.jpg
├── T1011685052982415360.jpg
├── T1011695871761833987.jpg
├── T1011697874470653952.jpg
├── T1011698048647360512.jpg
├── T1011699257290739713.jpg
├── T1011700728883220482.jpg
├── ...
Running the scripts
To replicate all existing results, run:
chmod +x /scripts/*
./scripts/test-results.sh -t /data/twitter-tmsc -b /data/bloomberg-twitter-text-image -m /data/mvsa -c 0 -r 5
./scripts/frozen-lms.sh -t /data/twitter-tmsc -b /data/bloomberg-twitter-text-image -c 0 -r 5
./scripts/toms.sh -t /data/twitter-tmsc -c 0 -r 5
The dataset directories are given based on the directory structures above, and you can change the device (-c) or the number of repetitions (-r) to your liking.
The final metrics will be available at ./experiment_logs under subfolders with descriptive names, e.g.
experiment_logs/
├── TomBERTTMSC
│ ├── None,twitter2015(train),bert-base-uncased,False_0
│ └── None,twitter(train),bert-base-uncased,False_0
├── TomViLTTMSC
│ ├── twitter2015(train),bert-base-uncased,vilt-b32-mlm,False_0
│ ├── twitter2015(train),bert-base-uncased,vilt-b32-mlm,False_1
│ ├── twitter(train),bert-base-uncased,vilt-b32-mlm,False_0
│ └── twitter(train),bert-base-uncased,vilt-b32-mlm,False_1
├── VaultTMSCBloomberg
│ ├── bert-base-uncased,bloomberg-twitter-text-image(train;dev),vilt-b32-mlm_0
│ ├── bert-base-uncased,bloomberg-twitter-text-image(train;dev),vilt-b32-mlm_1
│ ├── bertweet-base,bloomberg-twitter-text-image(train;dev),vilt-b32-mlm_0
│ ├── bertweet-base,bloomberg-twitter-text-image(train;dev),vilt-b32-mlm_1
│ └── None,bloomberg-twitter-text-image(train;dev),vilt-b32-mlm_1
├── VaultTMSCMVSA
│ ├── bert-base-uncased,MVSA_Single(train;dev),vilt-b32-mlm_0
│ ├── bert-base-uncased,MVSA(train;dev),vilt-b32-mlm_0
│ ├── bertweet-base,MVSA_Single(train;dev),vilt-b32-mlm_0
│ ├── bertweet-base,MVSA(train;dev),vilt-b32-mlm_0
│ ├── None,MVSA_Single(train;dev),vilt-b32-mlm_0
│ └── None,MVSA(train;dev),vilt-b32-mlm_0
└── VaultTMSCTwitter201X
├── bert-base-uncased,twitter2015(train;dev),vilt-b32-mlm,False_0
├── bert-base-uncased,twitter2017(train;dev),vilt-b32-mlm,False_0
├── bertweet-base,twitter2015(train;dev),vilt-b32-mlm,False_0
├── bertweet-base,twitter2015(train;dev),vilt-b32-mlm,False_1
├── bertweet-base,twitter(train;dev),vilt-b32-mlm,False_0
├── bertweet-base,twitter(train;dev),vilt-b32-mlm,False_1
├── None,twitter2015(train;dev),vilt-b32-mlm,False_0
└── None,twitter(train;dev),vilt-b32-mlm,False_0
The name of each specific experiment subfolder contains some important information like which models were used, which splits, etc. The final integer is used to differentiate runs with different hyperparameters if they do not appear in the main name (the same experiment ran twice will log additional experiments in the same subfolder). Each of the above "leaf" subfolders will look like:
experiment_logs/VaultTMSCTwitter201X/bert-base-uncased,twitter2015(train;dev),vilt-b32-mlm,False_0/
├── aggregated_metrics.yml
├── metrics.yml
├── obj.pkl
├── params.yml
└── plots
└── train_loss.png
aggregated_metrics.yml contains the mean and the standard deviation of all logged metrics across multiple runs for the epoch designated as the final one (e.g. simple training will use the last epoch, early stopping will use the corresponding step, see example below), metrics.yml contains all the logged metrics, params.yml contains the hyperparameter configuration, plots contains plots of the metrics that are logged across all training (not final metrics, e.g. those of the test set), and obj.pkl contains the logger object, which you can re-load by using the load_existent method of vault.logging_utils.ExperimentHandler.
? ''
: best_train_loss: 0.0016+-0.0002
test_eval_accuracy: 0.7563+-0.0084
test_eval_loss: 1.5356+-0.0650
test_macro_f1_score: 0.6997+-0.0171