{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "5338af6d",
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import AutoModel, AutoTokenizer\n",
    "import numpy as np"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "3e5e27fb",
   "metadata": {},
   "outputs": [],
   "source": [
    "model = AutoModel.from_pretrained(\"UF-NLPC-Lab/bart-stance-mixed\", trust_remote_code=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "eccdbd4e",
   "metadata": {},
   "outputs": [],
   "source": [
    "tokenizer = AutoTokenizer.from_pretrained(\"UF-NLPC-Lab/bart-stance-mixed\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "86cf1631",
   "metadata": {},
   "outputs": [],
   "source": [
    "# One claim target, and one noun-phrase target\n",
    "# First prediction should be FAVOR, second should be NONE as the author says nothing about his opinion on salad itself.\n",
    "samples = [\"Carrots are the superior vegetable.\", \"I don't like ranch on my salad.\"]\n",
    "targets = [\"Brocoli is inferior to carrots.\", \"salad\"]\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f6965a16",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Model was trained with target-then-sample. \n",
    "# You can probably get away with tokenizing like this...\n",
    "encoding = tokenizer(text=targets, text_pair=samples, return_tensors='pt', padding=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "82673e84",
   "metadata": {},
   "outputs": [],
   "source": [
    "# ... but this is what we actually did during our own training and evaluation\n",
    "max_context_length = 256\n",
    "max_target_length = 64\n",
    "context_trunc = tokenizer.decode(tokenizer.encode(samples, max_length=max_context_length, add_special_tokens=False), is_split_into_words=False)\n",
    "target_trunc = tokenizer.decode(tokenizer.encode(targets, max_length=max_target_length, add_special_tokens=False), is_split_into_words=False)\n",
    "combined = tokenizer(text=context_trunc, text_pair=target_trunc, return_tensors='pt', padding=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "5d93f032",
   "metadata": {},
   "outputs": [],
   "source": [
    "output = model(**encoding)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "6a83f378",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['FAVOR', 'NONE']"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "logits = output.logits.detach().cpu().numpy()\n",
    "preds = np.argmax(logits, axis=-1)\n",
    "str_preds = [model.config.id2label[p] for p in preds]\n",
    "str_preds"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "26496ec3",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "min_transformers",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.9.22"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}