{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from datasets import load_dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_files = {\n",
    "    \"train\": \"./train.txt\",\n",
    "    \"val\": \"./val.txt\",\n",
    "    \"test\": \"./test.txt\",\n",
    "}\n",
    "ds = load_dataset(\"text\", data_files=data_files)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "ds['train'] = ds['train'].rename_column('text', 'SMILE')\n",
    "ds['val'] = ds['val'].rename_column('text', 'SMILE')\n",
    "ds['test'] = ds['test'].rename_column('text', 'SMILE')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import selfies as sf\n",
    "\n",
    "def try_convert(row):\n",
    "    selfie = None\n",
    "    try:\n",
    "        selfie = sf.encoder(row['SMILE'])\n",
    "    except:\n",
    "        pass\n",
    "\n",
    "    return {'SELFIE': selfie}\n",
    "\n",
    "# Alongside the SMILES, we also need to convert them to SELFIES\n",
    "# ds['train'] = ds['train'].add_column('SELFIE', ds['train'].map(try_convert, num_proc=8))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "ds['train'] = ds['train'].map(try_convert, num_proc=8)\n",
    "ds['val'] = ds['val'].map(try_convert, num_proc=8)\n",
    "ds['test'] = ds['test'].map(try_convert, num_proc=8)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Drop the rows where the conversion failed\n",
    "ds['train'] = ds['train'].filter(lambda row: row['SELFIE'] is not None)\n",
    "ds['val'] = ds['val'].filter(lambda row: row['SELFIE'] is not None)\n",
    "ds['test'] = ds['test'].filter(lambda row: row['SELFIE'] is not None)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from tokenizers import Tokenizer\n",
    "\n",
    "tokenizer = Tokenizer.from_pretrained(\"haydn-jones/GuacamolSELFIETokenizer\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "unk_id = tokenizer.token_to_id('<UNK>')\n",
    "\n",
    "# Drop any rows where the tokenization has an <UNK> token\n",
    "ds['train'] = ds['train'].filter(lambda row: unk_id not in tokenizer.encode(row['SELFIE']).ids, num_proc=8)\n",
    "ds['val'] = ds['val'].filter(lambda row: unk_id not in tokenizer.encode(row['SELFIE']).ids, num_proc=8)\n",
    "ds['test'] = ds['test'].filter(lambda row: unk_id not in tokenizer.encode(row['SELFIE']).ids, num_proc=8)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "ds.save_to_disk('./guacamol')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "repo_id = \"haydn-jones/Guacamol\"\n",
    "\n",
    "# Push the dataset to the repo\n",
    "ds.push_to_hub(repo_id)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "ddpm",
   "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": "3.11.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}