main-text-gen-rnn.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "import os, time, random, math, re, json, string, sys, datetime, textwrap\n",
    "import matplotlib.pyplot as plt\n",
    "from tqdm import tqdm\n",
    "from tokenizer import tokenizer, detokenizer\n",
    "from MyModel import MyModel, OneStep\n",
    "import keras.backend as K\n",
    "import tensorflow_model_optimization as tfmot"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Compute dtype: float32\n",
      "Variable dtype: float32\n"
     ]
    }
   ],
   "source": [
    "# policy = tf.keras.mixed_precision.Policy('mixed_float16')\n",
    "# tf.keras.mixed_precision.set_global_policy(policy)\n",
    "\n",
    "policy = tf.keras.mixed_precision.global_policy()\n",
    "\n",
    "print('Compute dtype: %s' % policy.compute_dtype)\n",
    "print('Variable dtype: %s' % policy.variable_dtype)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# gpu growth\n",
    "gpus = tf.config.experimental.list_physical_devices('GPU')\n",
    "for gpu in gpus:\n",
    "    tf.config.experimental.set_memory_growth(gpu, True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(10000000,\n",
       " \"I am a third year (starting fourth year in the fall) PhD student in mathematics. I've passed all qua\",\n",
       " 9134)"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Dataset\n",
    "all_text = open(\"tmp/posts.txt\", 'rb').read().decode(encoding='utf-8')\n",
    "all_text = all_text[:10_000_000]\n",
    "\n",
    "len(all_text), all_text[:100], len(all_text.split(\"<SPECIAL:END>\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "5002\n"
     ]
    }
   ],
   "source": [
    "all_tokens = tokenizer(all_text)\n",
    "alltokens100k = all_tokens[:500_000]\n",
    "# vocab = list(set(all_tokens[:500_000]))\n",
    "# vocab = list(sorted(vocab, key=alltokens100k.count, reverse=True))\n",
    "vocab = json.load(open(\"models/vocab.json\", 'r'))\n",
    "print(len(vocab))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "# create tokens to ids and ids to tokens, vocab is sorted by frequency\n",
    "text_vectorization = tf.keras.layers.TextVectorization(\n",
    "    standardize=None,\n",
    "    split=None,\n",
    "    vocabulary=vocab[:5000]\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "vocab size 5000\n"
     ]
    }
   ],
   "source": [
    "print(\"vocab size\", len(text_vectorization.get_vocabulary()))\n",
    "# The maximum length sentence we want for a single input in characters\n",
    "seq_length = 150\n",
    "all_ids = text_vectorization(all_tokens)\n",
    "ds = tf.data.Dataset.from_tensor_slices(all_ids)\n",
    "sequences = ds.batch(seq_length + 1, drop_remainder=True)\n",
    "\n",
    "def split_input_target(chunk):\n",
    "    input_text = chunk[:-1]\n",
    "    target_text = chunk[1:]\n",
    "    return input_text, target_text\n",
    "\n",
    "ds = sequences.map(split_input_target)\n",
    "\n",
    "# Batch size\n",
    "BATCH_SIZE = 256\n",
    "ds = ds.repeat(-1).shuffle(10000).batch(BATCH_SIZE, drop_remainder=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "I'm looking for a lot in an answer. It should:\n",
      "\n",
      "be short\n",
      "open the [UNK] person's eyes\n",
      "allow for a [UNK] exit by both parties\n",
      "\n",
      "Here is an example of the kind of answer I'm looking for:\n",
      "\n",
      "Question posed by a white person to a [UNK] black [UNK]:\" What country were you born in?\"\n",
      "Answer:\" Long [unk].\"\n",
      "\n",
      "Edit: Please assume there was no hidden [UNK]. Just some combination of [UNK] variety [UNK] and [UNK] thinking.\n",
      "Typical scenarios: [UNK] [UNK]... parent meeting at your [UNK] school... standing in line at the [UNK] store....\n",
      "Edit 9 / 18:\n"
     ]
    }
   ],
   "source": [
    "for example in ds.take(1):\n",
    "    print(detokenizer(np.array(text_vectorization.get_vocabulary())[example[0][0].numpy()]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"my_model\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " embedding (Embedding)       multiple                  1280000   \n",
      "                                                                 \n",
      " lstm (LSTM)                 multiple                  18882560  \n",
      "                                                                 \n",
      " dense (Dense)               multiple                  10245000  \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 30,407,560\n",
      "Trainable params: 30,407,560\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "# Length of the vocabulary in StringLookup Layer\n",
    "vocab_size = len(text_vectorization.get_vocabulary())\n",
    "\n",
    "# The embedding dimension\n",
    "embedding_dim = 256\n",
    "\n",
    "# Number of RNN units\n",
    "rnn_units = 2048\n",
    "\n",
    "model = MyModel(vocab_size=vocab_size, embedding_dim=embedding_dim, rnn_units=rnn_units)\n",
    "model.build(tf.TensorShape([None, None]))\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<tf.Tensor: shape=(1, 3, 5000), dtype=float32, numpy=\n",
       "array([[[-1.2481312e-03,  5.1734853e-05,  7.7220460e-04, ...,\n",
       "          1.3093359e-03,  7.3006219e-04,  2.5928987e-03],\n",
       "        [ 2.8894821e-04, -6.5484259e-04, -7.7146059e-04, ...,\n",
       "         -1.1503412e-03,  4.4443540e-04,  1.9332578e-03],\n",
       "        [-9.7052840e-04, -8.2985597e-04,  2.5322498e-04, ...,\n",
       "         -1.0170377e-03, -4.9060740e-04,  2.7268527e-03]]], dtype=float32)>"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model(tf.constant([[1, 2, 3]]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(256, 150, 5000) # (batch_size, sequence_length, vocab_size)\n",
      "Model: \"my_model\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " embedding (Embedding)       multiple                  1280000   \n",
      "                                                                 \n",
      " lstm (LSTM)                 multiple                  18882560  \n",
      "                                                                 \n",
      " dense (Dense)               multiple                  10245000  \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 30,407,560\n",
      "Trainable params: 30,407,560\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "\n",
    "for input_example_batch, target_example_batch in ds.take(1):\n",
    "    example_batch_predictions = model(input_example_batch)\n",
    "    print(example_batch_predictions.shape, \"# (batch_size, sequence_length, vocab_size)\")\n",
    "\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# does not work so this is a hack\n",
    "def plot_model_works(model):\n",
    "    # inp is cop of input_example_batch\n",
    "    inp = tf.keras.layers.Input(shape=[None], dtype=tf.int64)\n",
    "    model1 = tf.keras.Model(inp, model.call(inp))\n",
    "    return tf.keras.utils.plot_model(model1, show_shapes=True)\n",
    "\n",
    "\n",
    "plot_model_works(model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "('float32', 'float32', 'float32')"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.layers[0].dtype, model.layers[1].dtype, model.layers[2].dtype"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Input:\n",
      " Done with my current field? If not, how can one deal with the impression of failure and regret?\n",
      "\n",
      "(English is not my first language, any edits are [UNK] welcomed) <SPECIAL:END> There is an [UNK] sharing the office with me. Once day his Pi came to talk to him. Since there were only 3 of us in the office, although I didn't pay attention, their conversation just automatically went into my mind.\n",
      "\n",
      "The story is like this: the [UNK] had worked for 2 months, and he was going to leave. However, he would not let the university know that he was leaving. Instead, the Pi would tell the university that the [UNK] would work for another month.\n",
      "\n",
      "Next Char Predictions:\n",
      " Subfield 25 push refrain bound bound bound itself entirely everyone direction water departmental youtube min collaborating sexual meetings content handwritten was ask course mentioning course reviewer's “ “ “ serving tax researcher r luck i've i've scientists amazon count wait undergrads exception non seemingly india note failed failed consistent consistent art art activities magnitude syllabus entirely entirely capacity employee pure gaps political political political political suitable collaborations collaborations should've atmosphere industry holding race defenses it'll disclaimer it'll it'll “ allowing groups groups algebra prices capacity sum sum follow materials neighboring workload passages passages samples watching watching host frankly itself frustrated frustrated consider consider ground suspect abc “ state still is refrain databases discovered guidelines believe quantity quantity discord retraction external independent careers november thesis recruitment used gauge gauge copy-pasted used legally participated respect next respected case positive common least develop teaches gauge indication penalties choose 8-9 mathematica considering unexpectedly limited\n"
     ]
    }
   ],
   "source": [
    "print(\"Input:\\n\", detokenizer(np.array(text_vectorization.get_vocabulary())[input_example_batch[0].numpy()]))\n",
    "print()\n",
    "print(\"Next Char Predictions:\\n\", detokenizer(np.array(text_vectorization.get_vocabulary())[np.argmax(example_batch_predictions[0], axis=-1)]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Prediction shape:  (256, 150, 5000)  # (batch_size, sequence_length, vocab_size)\n",
      "Mean loss:         tf.Tensor(8.517181, shape=(), dtype=float32)\n",
      "Exp of mean loss:  4999.941\n",
      "Vocab size:        5000\n"
     ]
    }
   ],
   "source": [
    "loss = tf.losses.SparseCategoricalCrossentropy(from_logits=True)\n",
    "\n",
    "def loss(labels, logits):\n",
    "    # scce but capped at 2000.0 to avoid exploding gradients\n",
    "    loss_val = tf.keras.losses.sparse_categorical_crossentropy(labels, logits, from_logits=True)\n",
    "    loss_val = tf.clip_by_value(loss_val, clip_value_min=-1000.0, clip_value_max=2000.0)\n",
    "    return tf.reduce_mean(loss_val)\n",
    "\n",
    "\n",
    "example_batch_mean_loss = loss(target_example_batch, example_batch_predictions)\n",
    "print(\"Prediction shape: \", example_batch_predictions.shape, \" # (batch_size, sequence_length, vocab_size)\")\n",
    "print(\"Mean loss:        \", example_batch_mean_loss)\n",
    "print(\"Exp of mean loss: \", np.exp(example_batch_mean_loss))\n",
    "print(\"Vocab size:       \", vocab_size)\n",
    "\n",
    "optimizer = tf.optimizers.Adam(learning_rate=tf.keras.optimizers.schedules.ExponentialDecay(0.003, 300, 0.95))\n",
    "\n",
    "model.compile(optimizer=optimizer, loss=loss)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "10/10 [==============================] - 17s 2s/step - loss: 0.6322\n"
     ]
    }
   ],
   "source": [
    "# Directory where the checkpoints will be saved\n",
    "checkpoint_dir = './tmp/training_checkpoints'\n",
    "# Name of the checkpoint files\n",
    "checkpoint_prefix = os.path.join(checkpoint_dir, \"ckpt_{epoch}\")\n",
    "\n",
    "checkpoint_callback=tf.keras.callbacks.ModelCheckpoint(\n",
    "    filepath=checkpoint_prefix,\n",
    "    save_weights_only=True)\n",
    "\n",
    "\n",
    "# Set epochs to 50 and steps_per_epoch to 600 to replicate the results\n",
    "history = model.fit(ds, epochs=1, steps_per_epoch=10, callbacks=[checkpoint_callback])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [],
   "source": [
    "# model.save_weights(\"tmp/weights3.h5\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [],
   "source": [
    "# model.load_weights(\"tmp/weights2.h5\")\n",
    "# model.load_weights(\"models/model1-test2-backup/weights.h5\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# plot loss\n",
    "plt.plot(history.history['loss'])\n",
    "plt.title('model loss')\n",
    "plt.ylabel('loss')\n",
    "plt.xlabel('epoch')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "20/20 [==============================] - 12s 511ms/step - loss: 0.8702\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "0.8701754808425903"
      ]
     },
     "execution_count": 38,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.evaluate(ds, steps=20)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "OneStep.__init__\n"
     ]
    }
   ],
   "source": [
    "one_step_model = OneStep(model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [],
   "source": [
    "MODEL_NAME = \"models/model1-test2\"\n",
    "\n",
    "# make sure the directory exists\n",
    "os.makedirs(MODEL_NAME, exist_ok=True)\n",
    "\n",
    "model.save_weights(f\"{MODEL_NAME}/weights.h5\")\n",
    "\n",
    "with open(f\"{MODEL_NAME}/vocab.json\", \"w\") as f:\n",
    "    f.write(json.dumps(text_vectorization.get_vocabulary()))\n",
    "\n",
    "with open(f\"{MODEL_NAME}/meta.json\", \"w\") as f:\n",
    "    f.write(json.dumps({\n",
    "        \"embedding_dim\": embedding_dim,\n",
    "        \"rnn_units\": rnn_units,\n",
    "        \"vocab_size\": vocab_size,\n",
    "    }))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 100/100 [00:01<00:00, 68.66it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "I recently published my first research paper and it was accepted in a peer-reviewed journal in the field of Ai. I worked in private engineering in Europe (mostly in area, linear algebra, coding, and field topics that I have never seen on a topic that is required in the field on research and also engineering). I am worried about the quality of the world at the moment, but I know that this source of my ability of doing research is for all of those fields, being a big\"\n",
      "--------------------------------------------------------------------------------\n",
      "I recently published my first research paper and it was accepted in a peer-reviewed\n",
      "journal in the field of Ai. I worked in private engineering in Europe (mostly in area,\n",
      "linear algebra, coding, and field topics that I have never seen on a topic that is\n",
      "required in the field on research and also engineering). I am worried about the quality of\n",
      "the world at the moment, but I know that this source of my ability of doing research is\n",
      "for all of those fields, being a big\"\n",
      "\n",
      "Run time: 2.3109986782073975\n"
     ]
    }
   ],
   "source": [
    "def generate_text(starting, num_generate):\n",
    "    start = time.time()\n",
    "    states = None\n",
    "    next_char = text_vectorization(tokenizer(starting))\n",
    "    result = next_char.numpy()\n",
    "\n",
    "    for n in tqdm(range(num_generate)):\n",
    "        next_char = tf.expand_dims(next_char, 0)\n",
    "        next_char, states = one_step_model.generate_one_step(next_char, states=states, temperature=0.7)\n",
    "        result = np.append(result, next_char.numpy(), axis=0)\n",
    "    result = detokenizer([text_vectorization.get_vocabulary()[r] for r in result])\n",
    "    end = time.time()\n",
    "    print()\n",
    "    print(result)\n",
    "    print(\"-\" * 80)\n",
    "    w = textwrap.TextWrapper(width=90,break_long_words=True,replace_whitespace=False)\n",
    "    print(w.fill(result))\n",
    "    print(\"\\nRun time:\", end - start)\n",
    "\n",
    "generate_text(\"I recently published my\", 100)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "5000"
      ]
     },
     "execution_count": 42,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(text_vectorization.get_vocabulary())"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "tf",
   "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": "3.9.13"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}