Twitter Sentiment Analysis
Team name: AntipathyGlance
Team members: Rolando, Klim, Mengtao, Fabian
How to use
- Clone this repo
- Download the twitter Datasets and extract everything in this folder.
- Install the python dependencies, You can install them with
pip install -r pip-freeze.txt
- Run the HF_dataset.py
- Now you can run HF_transformer.py for the best models.
- You can also run the file baselines.py for the baselines. (you will need to create an empty 'images' folder)
(Quick Note: When running locally or on the cluster the environment and the paths are usually different. For this reason there exists an --on_cluster parameter that allows to switch between environments. You will need to set up some environment variables on the cluster:
CIL_PROJECT_PATH
CIL_EXPERIMENTS_PATH → it is recommended to put this one on the scratch partition.
)
Branches
There are 3 main branches for the final project:
-
master: This is where most of the code is, in its final state. It can run all the HuggingFace models using the HF_dataset.py file. The required parameters must be provided for it to run.
-
CA-KL: You will need to switch to this branch, check that you still have the correct HF_data.txt file. Then you will be able to run this model with
python SCAL_transformer.py --cfg configs/default_gpu.yaml
- bertTweetCustom: You will need to switch to this branch, check that you still have the correct HF_data.txt file. You will be able to run this model in the same way as master. For example:
python HF_transformer.py --tag "Test Transformer" --on_cluster --train --test --amount_per_it 50000 --amount_of_data 2000000 --full_data --bs_train 32 --bs_eval 16 --n_epochs 3
Best Model command
To reproduce the best run we had, you will need to run this:
python HF_transformer.py --tag "Test BERTTweet" --on_cluster --train --test --n_epochs 3 --lr 1e-5 --fp16 --model_name vinai/bertweet-base --bs_train 32 --bs_eval 32 --tokenizer_max_length 128
It would be advised to run it on the cluster as it can take quite a lot of time... If you run it on the cluster you should set the correct environment variables, if not, please remove the "--on_cluster" parameter.
HF_transformer.py Usage
usage: HF_transformer.py [-h] [--tag TAG] [--seed SEED] [--on_cluster] [--autosubmit] [--discord] [--load_model LOAD_MODEL] [--test] [--train] [--num_labels NUM_LABELS] [--full_data] [--amount_of_data AMOUNT_OF_DATA] [--amount_per_it AMOUNT_PER_IT] [--start_at_it START_AT_IT] [--use_HF_dataset_format] [--freq_words] [--tokenizer_max_length TOKENIZER_MAX_LENGTH] [--model_name MODEL_NAME] [--train_val_ratio TRAIN_VAL_RATIO] [--lr LR] [--n_epochs N_EPOCHS] [--weight_decay WEIGHT_DECAY] [--bs_train BS_TRAIN] [--bs_eval BS_EVAL] [--fp16]
options: -h, --help show this help message and exit --tag TAG A custom tag for this experiment --seed SEED Random number generator seed --on_cluster --autosubmit --discord --load_model LOAD_MODEL Checkpoint ID of a pretrained model in the experiment folder --test Load a model and generate output file --train Train a model --num_labels NUM_LABELS How many different classes there are to predict --full_data Use the full dataset --amount_of_data AMOUNT_OF_DATA Amount of Data --amount_per_it AMOUNT_PER_IT Amount of data to load on each iteration --start_at_it START_AT_IT Start at a certain iteration (usefull for resuming a training) --use_HF_dataset_format Use the cleaned HF dataset we created. --freq_words Use the most frequent words. --tokenizer_max_length TOKENIZER_MAX_LENGTH Set the tokenizer max_length parameter. --model_name MODEL_NAME Default model name to load --train_val_ratio TRAIN_VAL_RATIO The training/validation ratio to use for the given dataset --lr LR Learning rate --n_epochs N_EPOCHS Number of train epochs --weight_decay WEIGHT_DECAY Weight decay --bs_train BS_TRAIN Batch size for the training set --bs_eval BS_EVAL Batch size for validation/test set --fp16 Uses fp16 for training (Not always supported)
Twitter Datasets
Download the tweet datasets from here: http://www.da.inf.ethz.ch/teaching/2018/CIL/material/exercise/twitter-datasets.zip
The dataset should have the following files:
- sample_submission.csv
- train_neg.txt : a subset of negative training samples
- train_pos.txt: a subset of positive training samples
- test_data.txt:
- train_neg_full.txt: the full negative training samples
- train_pos_full.txt: the full positive training samples