ocroline3-test.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Populating the interactive namespace from numpy and matplotlib\n"
     ]
    }
   ],
   "source": [
    "%pylab inline\n",
    "#sys.path.append('../torchmore')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import sys\n",
    "import argparse\n",
    "import itertools\n",
    "from itertools import islice"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import pylab\n",
    "import psutil\n",
    "import editdistance\n",
    "import torch\n",
    "import torch.optim as optim\n",
    "from torch import nn"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torchmore as tm\n",
    "import torchtrainers as tt\n",
    "import webdataset as wds\n",
    "import tarproclib as tpl"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "from ocroline import lineest\n",
    "from ocroline import sequence\n",
    "from torchmore import helpers\n",
    "codec = sequence.ascii_codec"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Continuum Mechanics\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "ds = wds.WebDataset(\"/mdata/tmb-ocr/uw3-dew-training.tgz\", extensions=\"png txt\")\n",
    "for image, txt in ds:\n",
    "    break\n",
    "imshow(image)\n",
    "print(txt)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "import math\n",
    "def batchedbuckets(bs=5, scale=1.5):\n",
    "    buckets = {}\n",
    "    def f(data):\n",
    "        for image, text in data:\n",
    "            l = image.shape[1]\n",
    "            r = int(math.floor(math.log(l) / math.log(scale)))\n",
    "            batched = buckets.get(r, [[], []])\n",
    "            batched[0].append(image)\n",
    "            batched[1].append(text)\n",
    "            #print(l, r, len(batched[0]))\n",
    "            buckets[r] = batched\n",
    "            if len(batched[0])>bs:\n",
    "                yield tuple(batched)\n",
    "                batched[0] = []\n",
    "                batched[1] = []\n",
    "        for r, batched in list(buckets.items()):\n",
    "            if len(batched[0])==0: continue\n",
    "            yield tuple(batched)\n",
    "    return f"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(6, 48, 1295) (6, 159, 97)\n",
      "(6, 48, 1345) (6, 161, 97)\n",
      "(6, 48, 1405) (6, 163, 97)\n",
      "(6, 48, 1429) (6, 167, 97)\n",
      "(6, 48, 1680) (6, 185, 97)\n",
      "(6, 48, 278) (6, 39, 97)\n",
      "(6, 48, 1682) (6, 199, 97)\n",
      "(6, 48, 654) (6, 77, 97)\n",
      "(6, 48, 1683) (6, 193, 97)\n",
      "(6, 48, 1445) (6, 173, 97)\n",
      "(6, 48, 1682) (6, 187, 97)\n",
      "(6, 48, 433) (6, 53, 97)\n",
      "(6, 48, 930) (6, 113, 97)\n",
      "(6, 48, 1446) (6, 165, 97)\n",
      "(6, 48, 1037) (6, 121, 97)\n",
      "(6, 48, 951) (6, 119, 97)\n",
      "(6, 48, 1261) (6, 137, 97)\n",
      "(6, 48, 1262) (6, 133, 97)\n",
      "(6, 48, 994) (6, 125, 97)\n",
      "(6, 48, 951) (6, 123, 97)\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "ds = wds.WebDataset(\"/mdata/tmb-ocr/uw3-dew-training.tgz\",\n",
    "                    pipeline=batchedbuckets(),\n",
    "                    extensions=\"png txt\")\n",
    "for image, txt in islice(ds, 0, 20):\n",
    "    image = sequence.makebatch(image)\n",
    "    image = np.mean(image, axis=3)\n",
    "    txt = codec.encode_batch(txt)\n",
    "    print(image.shape, txt.shape)\n",
    "subplot(211); imshow(image[0])\n",
    "subplot(212); imshow(txt[0])\n",
    "image = torch.Tensor(image).unsqueeze(3)\n",
    "txt = torch.Tensor(txt)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<module 'torchmore.layers' from '/usr/local/lib/python3.7/dist-packages/torchmore/layers.py'>"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from torchmore import flex\n",
    "from torchmore import layers\n",
    "from importlib import reload\n",
    "reload(layers)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "torch.Size([6, 951, 97])\n"
     ]
    }
   ],
   "source": [
    "reload(layers)\n",
    "reload(flex)\n",
    "#flex.verbose = True\n",
    "\n",
    "ninput = 48\n",
    "noutput = 97\n",
    "\n",
    "B, W, H, D = (0, 900), (0, 9000), ninput, (0, 5000)\n",
    "model = nn.Sequential(\n",
    "    # reorder to Torch conventions\n",
    "    layers.Reorder(\"BHWD\", \"BDHW\"),\n",
    "    layers.CheckSizes(B, 1, H, W, name=\"input\"),\n",
    "    # convolutional layers\n",
    "    flex.Conv2d(100, 3, padding=(1, 1)),  # BDWH\n",
    "    nn.ReLU(),\n",
    "    # turn image into sequence\n",
    "    layers.Reshape(0, [1, 2], 3),\n",
    "    layers.CheckSizes(B, W, D),\n",
    "    # run 1D LSTM\n",
    "    flex.Lstm1(100),\n",
    "    flex.Conv1d(noutput, 1),\n",
    "    # reorder\n",
    "    layers.Reorder(\"BDW\", \"BWD\"),\n",
    "    layers.CheckSizes(B, W, noutput, name=\"output\"))\n",
    "\n",
    "print(model.forward(image).shape)\n",
    "model.cuda()\n",
    "\n",
    "optimizer = optim.SGD(model.parameters(),\n",
    "                      lr=1e-4,\n",
    "                      momentum=0.9)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Help on built-in function softmax:\n",
      "\n",
      "softmax(...)\n",
      "\n"
     ]
    }
   ],
   "source": [
    "help(torch.softmax)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "torch.Size([6, 48, 951, 1])\n",
      "torch.Size([6, 123, 97])\n",
      "cuda:0 cuda:0\n",
      "torch.Size([6, 951, 97])\n",
      "torch.Size([6, 951, 97])\n"
     ]
    }
   ],
   "source": [
    "reload(helpers)\n",
    "\n",
    "model.cuda()\n",
    "\n",
    "def train_batch(model, input, target):\n",
    "    \"\"\"Train a BHWD input batch against a BWD target batch.\"\"\"\n",
    "    assert input.size(0) == target.size(0)\n",
    "    b, h, w, d = input.size()\n",
    "    assert d == 1\n",
    "    input = input.cuda()\n",
    "    target = target.cuda()\n",
    "    logits = model.forward(input)\n",
    "    probs = torch.softmax(logits, 2)\n",
    "    optimizer.zero_grad()\n",
    "    deltas, aligned = helpers.ctc_loss(probs, target)\n",
    "    print(deltas.device, aligned.device)\n",
    "    optimizer.step()\n",
    "    return probs, aligned\n",
    "\n",
    "print(image.shape)\n",
    "print(txt.shape)\n",
    "probs, aligned = train_batch(model, image, txt)\n",
    "print(probs.shape)\n",
    "print(aligned.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x7fa15afa6f28>"
      ]
     },
     "execution_count": 49,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "imshow(aligned[0].detach().cpu().numpy().T[:,:100])"
   ]
  },
  {
   "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": "3.7.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}