MBart-Translation.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import AutoModelForSeq2SeqLM, AutoTokenizer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "#model_name = \"facebook/mbart-large-en-ro\"\n",
    "model_name = \"cahya/mbart-large-en-de\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "device = \"cuda\"\n",
    "task = \"translation\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "## Helper functions from transformers/examples/seq2seq/utils.py\n",
    "def use_task_specific_params(model, task):\n",
    "    \"\"\"Update config with summarization specific params.\"\"\"\n",
    "    task_specific_params = model.config.task_specific_params\n",
    "\n",
    "    if task_specific_params is not None:\n",
    "        pars = task_specific_params.get(task, {})\n",
    "        logger.info(f\"using task specific params for {task}: {pars}\")\n",
    "        model.config.update(pars)\n",
    "\n",
    "def trim_batch(\n",
    "    input_ids, pad_token_id, attention_mask=None,\n",
    "):\n",
    "    \"\"\"Remove columns that are populated exclusively by pad_token_id\"\"\"\n",
    "    keep_column_mask = input_ids.ne(pad_token_id).any(dim=0)\n",
    "    if attention_mask is None:\n",
    "        return input_ids[:, keep_column_mask]\n",
    "    else:\n",
    "        return (input_ids[:, keep_column_mask], attention_mask[:, keep_column_mask])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device)\n",
    "tokenizer = AutoTokenizer.from_pretrained(model_name)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "decoder_start_token_id = tokenizer.bos_token_id"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "use_task_specific_params(model, task)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "texts = [\"I am very hungry\", \"The weather is very hot today\"]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.\n"
     ]
    }
   ],
   "source": [
    "#batch = tokenizer(texts, return_tensors=\"pt\", truncation=True, padding=\"max_length\").to(device)\n",
    "batch = tokenizer(texts, return_tensors=\"pt\", truncation=True, padding='longest').to(device)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'input_ids': tensor([[    87,    444,   4552,   1926,  47285,      2, 250004,      1],\n",
       "        [   581,  92949,     83,   4552,   8010,  18925,      2, 250004]],\n",
       "       device='cuda:0'), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 0],\n",
       "        [1, 1, 1, 1, 1, 1, 1, 1]], device='cuda:0')}"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "batch"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "input_ids, attention_mask = trim_batch(**batch, pad_token_id=tokenizer.pad_token_id)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "summaries = model.generate(\n",
    "            input_ids=input_ids,\n",
    "            attention_mask=attention_mask,\n",
    "            decoder_start_token_id=decoder_start_token_id\n",
    "        )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([[0, 0, 6, 5],\n",
       "        [0, 0, 6, 5]], device='cuda:0')"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "summaries"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "dec = tokenizer.batch_decode(summaries, skip_special_tokens=True, clean_up_tokenization_spaces=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['.', '.']"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "dec"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.7.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}