how-to-use-scivision.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# How to use `scivision`\n",
    "\n",
    "In this notebook, we will:\n",
    "\n",
    "1. Demonstrate using the scivision [Python API](https://scivision.readthedocs.io/en/latest/api.html) to load several pretrained image classification models\n",
    "2. Use the scivision catalog to find a matching dataset, which the model can be run on\n",
    "3. Run the model on the data, performing simple model inference\n",
    "4. Use the scivision catalog to find another model that can be run on the same dataset"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "First let's import some things from scivision: `default_catalog` is a scivision **catalog** that will let us discover models and datasets, and `load_pretrained_model` provides a convenient way to load and run a model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "from scivision import default_catalog, load_pretrained_model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Inspecting our model in the scivision catalog\n",
    "\n",
    "A scivision catalog is a collection of **models** and **datasources**.\n",
    "\n",
    "For this example, we want to find datasources compatible with the model catalog entry \"image-classifiers\".  But first, let's use the catalog to retrive the \"image-classifiers\" repository url and take a look at the data contained in the *default catalog* (the built-in catalog, distributed as part of scivision) and see how this is structured."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "AnyUrl('https://github.com/alan-turing-institute/scivision_classifier', scheme='https', host='github.com', tld='com', host_type='domain', path='/alan-turing-institute/scivision_classifier')"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Get the model repo url\n",
    "models_catalog = default_catalog.models.to_dataframe()\n",
    "model_repo = models_catalog[models_catalog.name == \"image-classifiers\"].url.item()\n",
    "model_repo # Why not paste the repo link into your browser and see how it looks?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>description</th>\n",
       "      <th>tasks</th>\n",
       "      <th>url</th>\n",
       "      <th>pkg_url</th>\n",
       "      <th>format</th>\n",
       "      <th>pretrained</th>\n",
       "      <th>labels_required</th>\n",
       "      <th>institution</th>\n",
       "      <th>tags</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>image-classifiers</td>\n",
       "      <td>None</td>\n",
       "      <td>(TaskEnum.classificiation,)</td>\n",
       "      <td>https://github.com/alan-turing-institute/scivi...</td>\n",
       "      <td>git+https://github.com/alan-turing-institute/s...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>False</td>\n",
       "      <td>(alan-turing-institute,)</td>\n",
       "      <td>(classification, 2D, image)</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                name description                        tasks  \\\n",
       "8  image-classifiers        None  (TaskEnum.classificiation,)   \n",
       "\n",
       "                                                 url  \\\n",
       "8  https://github.com/alan-turing-institute/scivi...   \n",
       "\n",
       "                                             pkg_url format  pretrained  \\\n",
       "8  git+https://github.com/alan-turing-institute/s...  image        True   \n",
       "\n",
       "   labels_required               institution                         tags  \n",
       "8            False  (alan-turing-institute,)  (classification, 2D, image)  "
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Inspecting model entry and its metadata in the default catalog\n",
    "models_catalog[models_catalog.name == \"image-classifiers\"]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Loading the model\n",
    "\n",
    "Some model entries in the scivision catalog contain a single loadable model. Here, let's load the \"image-classifiers\" entry, which contains each of the models in the [image-classifiers](https://pypi.org/project/image-classifiers/) package. We can do this with the `load_pretrained_model` function. We'll load 2 models (`resnet18` and `densenet121`).\n",
    "\n",
    "Note: the \"model\" arg is not mandatory and will have a default value for any model catalog entries that have more than one model - for model catalog entries with a single model, this need not be set:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Note: when loading this model repo for the first time you'll need to add the allow_install=True flag\n",
    "resnet_model = load_pretrained_model(model_repo, model='resnet18', allow_install=True)\n",
    "densenet_model = load_pretrained_model(model_repo, model='densenet121', allow_install=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "scivision.PretrainedModel( \n",
       "  module='scivision_classifier', \n",
       "  model='resnet18', \n",
       "  source='https://github.com/alan-turing-institute/scivision_classifier' \n",
       "  pipe='DataPipe(input=<Parameter \"X: numpy.ndarray\">, output=<Parameter \"image: numpy.ndarray\">)' \n",
       ")"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# let's explore one of the model objects\n",
    "resnet_model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Later, we'll use these models to make predictions on image data found in the scivision catalog."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Query the default scivision data catalog"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now let's use the `default_catalog` to identify datasources in the catalog that are compatible with our models (based on `tasks`, `format` and `labels_provided`/`labels_required`)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>description</th>\n",
       "      <th>tasks</th>\n",
       "      <th>domains</th>\n",
       "      <th>url</th>\n",
       "      <th>format</th>\n",
       "      <th>labels_provided</th>\n",
       "      <th>institution</th>\n",
       "      <th>tags</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>data-002</td>\n",
       "      <td>None</td>\n",
       "      <td>(TaskEnum.classificiation,)</td>\n",
       "      <td>(oceanography,)</td>\n",
       "      <td>https://github.com/scotthosking/intake-plankto...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>(CEFAS,)</td>\n",
       "      <td>(help-needed, 2D)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>data-003</td>\n",
       "      <td>Koala</td>\n",
       "      <td>(TaskEnum.object_detection, TaskEnum.classific...</td>\n",
       "      <td>(computer-vision,)</td>\n",
       "      <td>https://github.com/alan-turing-institute/scivi...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>(Wikipedia by Erik Veland,)</td>\n",
       "      <td>(demo,)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>data-004</td>\n",
       "      <td>Sample dataset of Plankton required to demonst...</td>\n",
       "      <td>(TaskEnum.classificiation,)</td>\n",
       "      <td>(computer-vision,)</td>\n",
       "      <td>https://raw.githubusercontent.com/alan-turing-...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>(Centre for Environment, Fisheries and Aquacul...</td>\n",
       "      <td>(2D, plankton, ecology, environmental-science)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>data-005</td>\n",
       "      <td>Sample dataset of satellite imagery required t...</td>\n",
       "      <td>(TaskEnum.classificiation, TaskEnum.segmentation)</td>\n",
       "      <td>(computer-vision, earth-observation, ecology)</td>\n",
       "      <td>https://github.com/MartinSJRogers/VEdge_Detect...</td>\n",
       "      <td>image</td>\n",
       "      <td>False</td>\n",
       "      <td>(Cambridge University,)</td>\n",
       "      <td>(2D, satellite, remote-sensing, ecology, envir...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>data-007</td>\n",
       "      <td>A labelled dataset of honeybee images, from Ka...</td>\n",
       "      <td>(TaskEnum.classificiation,)</td>\n",
       "      <td>(nature, insects, bee-keeping)</td>\n",
       "      <td>https://github.com/pwochner/BeeImage_test</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>(alan-turing-institute,)</td>\n",
       "      <td>(bees, biology, classification)</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "       name                                        description  \\\n",
       "2  data-002                                               None   \n",
       "3  data-003                                              Koala   \n",
       "4  data-004  Sample dataset of Plankton required to demonst...   \n",
       "5  data-005  Sample dataset of satellite imagery required t...   \n",
       "7  data-007  A labelled dataset of honeybee images, from Ka...   \n",
       "\n",
       "                                               tasks  \\\n",
       "2                        (TaskEnum.classificiation,)   \n",
       "3  (TaskEnum.object_detection, TaskEnum.classific...   \n",
       "4                        (TaskEnum.classificiation,)   \n",
       "5  (TaskEnum.classificiation, TaskEnum.segmentation)   \n",
       "7                        (TaskEnum.classificiation,)   \n",
       "\n",
       "                                         domains  \\\n",
       "2                                (oceanography,)   \n",
       "3                             (computer-vision,)   \n",
       "4                             (computer-vision,)   \n",
       "5  (computer-vision, earth-observation, ecology)   \n",
       "7                 (nature, insects, bee-keeping)   \n",
       "\n",
       "                                                 url format  labels_provided  \\\n",
       "2  https://github.com/scotthosking/intake-plankto...  image             True   \n",
       "3  https://github.com/alan-turing-institute/scivi...  image             True   \n",
       "4  https://raw.githubusercontent.com/alan-turing-...  image             True   \n",
       "5  https://github.com/MartinSJRogers/VEdge_Detect...  image            False   \n",
       "7          https://github.com/pwochner/BeeImage_test  image             True   \n",
       "\n",
       "                                         institution  \\\n",
       "2                                           (CEFAS,)   \n",
       "3                        (Wikipedia by Erik Veland,)   \n",
       "4  (Centre for Environment, Fisheries and Aquacul...   \n",
       "5                            (Cambridge University,)   \n",
       "7                           (alan-turing-institute,)   \n",
       "\n",
       "                                                tags  \n",
       "2                                  (help-needed, 2D)  \n",
       "3                                            (demo,)  \n",
       "4     (2D, plankton, ecology, environmental-science)  \n",
       "5  (2D, satellite, remote-sensing, ecology, envir...  \n",
       "7                    (bees, biology, classification)  "
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "compatible_datasources = default_catalog.compatible_datasources(\"image-classifiers\").to_dataframe()\n",
    "compatible_datasources"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's use `data-003`, an image dataset containing a single image of a Koala."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<div class=\"alert alert-block alert-info\">\n",
    "We use the honey bee data set data-007 that we uploaded to Github instead of the koala (although the koala is definitely cuter)\n",
    "</div> "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>description</th>\n",
       "      <th>tasks</th>\n",
       "      <th>domains</th>\n",
       "      <th>url</th>\n",
       "      <th>format</th>\n",
       "      <th>labels_provided</th>\n",
       "      <th>institution</th>\n",
       "      <th>tags</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>data-007</td>\n",
       "      <td>A labelled dataset of honeybee images, from Ka...</td>\n",
       "      <td>(TaskEnum.classificiation,)</td>\n",
       "      <td>(nature, insects, bee-keeping)</td>\n",
       "      <td>https://github.com/pwochner/BeeImage_test</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>(alan-turing-institute,)</td>\n",
       "      <td>(bees, biology, classification)</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "       name                                        description  \\\n",
       "7  data-007  A labelled dataset of honeybee images, from Ka...   \n",
       "\n",
       "                         tasks                         domains  \\\n",
       "7  (TaskEnum.classificiation,)  (nature, insects, bee-keeping)   \n",
       "\n",
       "                                         url format  labels_provided  \\\n",
       "7  https://github.com/pwochner/BeeImage_test  image             True   \n",
       "\n",
       "                institution                             tags  \n",
       "7  (alan-turing-institute,)  (bees, biology, classification)  "
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "target_datasource = compatible_datasources.loc[compatible_datasources['name'] == 'data-007']\n",
    "target_datasource"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Load the dataset\n",
    "\n",
    "Now let's load the dataset using the scivision python API, specifically the [load_dataset](https://scivision.readthedocs.io/en/latest/api.html#scivision.io.reader.load_dataset) function, which takes as input the url of the data repository (structured as per [this template](https://scivision.readthedocs.io/en/latest/data_repository_template.html)), which we can get from the target datasource:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "from scivision import load_dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "AnyUrl('https://github.com/pwochner/BeeImage_test', scheme='https', host='github.com', tld='com', host_type='domain', path='/pwochner/BeeImage_test')"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data_url = target_datasource['url'].item()\n",
    "data_url"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The returned data config object of the`load_dataset` function is an \"intake catalog\". You can read our [documentation](https://scivision.readthedocs.io/en/latest/data_repository_template.html#data-config-file) to understand this better, but for now, let's inspect this config:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/yaml": "data:\n  args:\n    path: https://raw.githubusercontent.com/pwochner/BeeImage_test/main/.scivision/data.yml\n  description: ''\n  driver: intake.catalog.local.YAMLFileCatalog\n  metadata: {}\n",
      "text/plain": [
       "data:\n",
       "  args:\n",
       "    path: https://raw.githubusercontent.com/pwochner/BeeImage_test/main/.scivision/data.yml\n",
       "  description: ''\n",
       "  driver: intake.catalog.local.YAMLFileCatalog\n",
       "  metadata: {}\n"
      ]
     },
     "metadata": {
      "application/json": {
       "root": "data"
      }
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "data_config = load_dataset(data_url)\n",
    "data_config"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/pwochner/opt/miniconda3/envs/scivision/lib/python3.9/site-packages/xarray/core/dataarray.py:1195: FutureWarning: None value for 'chunks' is deprecated. It will raise an error in the future. Use instead '{}'\n",
      "  warnings.warn(\n"
     ]
    }
   ],
   "source": [
    "images = data_config.honey_bees().to_dask()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Clicking the `path` link to the location of this data config file online (in the dataset repo) reveals that there is one data source called `test_image`, and that the `intake_xarray.image.ImageSource` \"intake driver\" is being used. We can retrive the test image data in an image format which the model will accept, like so:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "test_image = images[0]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's take a look at the image with `matplotlib`:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x1760675e0>"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(test_image)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Model predictions\n",
    "\n",
    "Now let's use the loaded model on the test image data we found in the via catalog."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1/1 [==============================] - 0s 295ms/step\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "'isopod : 25.87%'"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "resnet_model.predict(test_image)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1/1 [==============================] - 1s 1s/step\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "'revolver : 9.33%'"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "densenet_model.predict(test_image)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "As you can see, the models have given predictions for the test image, with a confidence score. Check out the code in the [model repo](https://github.com/alan-turing-institute/scivision_classifier) to see how this was determined!"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Query the default scivision model catalog\n",
    "\n",
    "Using our test Koala image dataset, let's search the scivision default catalog for other models that can be used with it:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>description</th>\n",
       "      <th>tasks</th>\n",
       "      <th>url</th>\n",
       "      <th>pkg_url</th>\n",
       "      <th>format</th>\n",
       "      <th>pretrained</th>\n",
       "      <th>labels_required</th>\n",
       "      <th>institution</th>\n",
       "      <th>tags</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>model-000</td>\n",
       "      <td>None</td>\n",
       "      <td>(TaskEnum.object_detection, TaskEnum.segmentat...</td>\n",
       "      <td>https://github.com/stardist/stardist</td>\n",
       "      <td>git+https://github.com/stardist/stardist.git@main</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>True</td>\n",
       "      <td>(epfl,)</td>\n",
       "      <td>(2D, 3D, optical-microscopy, xray, microtomogr...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>model-001</td>\n",
       "      <td>None</td>\n",
       "      <td>(TaskEnum.segmentation, TaskEnum.thresholding,...</td>\n",
       "      <td>https://github.com/danforthcenter/plantcv</td>\n",
       "      <td>git+https://github.com/danforthcenter/plantcv@...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>True</td>\n",
       "      <td>(danforthcenter,)</td>\n",
       "      <td>(2D, hyperspectral, multispectral, near-infrar...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>scivision-test-plugin</td>\n",
       "      <td>None</td>\n",
       "      <td>(TaskEnum.object_detection,)</td>\n",
       "      <td>https://github.com/alan-turing-institute/scivi...</td>\n",
       "      <td>git+https://github.com/alan-turing-institute/s...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>False</td>\n",
       "      <td>(alan-turing-institute,)</td>\n",
       "      <td>(dummy,)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>mapreader-plant</td>\n",
       "      <td>automated detection of plant patches in images...</td>\n",
       "      <td>(TaskEnum.classificiation, TaskEnum.object_det...</td>\n",
       "      <td>https://github.com/alan-turing-institute/mapre...</td>\n",
       "      <td>git+https://github.com/alan-turing-institute/m...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>False</td>\n",
       "      <td>(alan-turing-institute,)</td>\n",
       "      <td>(2D, plant, phenotype, rgb, biology, agriculture)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>resnet50-plantkton</td>\n",
       "      <td>automated classification of plankton images co...</td>\n",
       "      <td>(TaskEnum.classificiation,)</td>\n",
       "      <td>https://github.com/alan-turing-institute/plank...</td>\n",
       "      <td>git+https://github.com/alan-turing-institute/p...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>False</td>\n",
       "      <td>(alan-turing-institute, cefas, plankton-analyt...</td>\n",
       "      <td>(2D, plankton, ecology, environmental-science)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>image-classifiers</td>\n",
       "      <td>None</td>\n",
       "      <td>(TaskEnum.classificiation,)</td>\n",
       "      <td>https://github.com/alan-turing-institute/scivi...</td>\n",
       "      <td>git+https://github.com/alan-turing-institute/s...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>False</td>\n",
       "      <td>(alan-turing-institute,)</td>\n",
       "      <td>(classification, 2D, image)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>9</th>\n",
       "      <td>huggingface-classifiers</td>\n",
       "      <td>None</td>\n",
       "      <td>(TaskEnum.classificiation,)</td>\n",
       "      <td>https://github.com/alan-turing-institute/scivi...</td>\n",
       "      <td>git+https://github.com/alan-turing-institute/s...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>False</td>\n",
       "      <td>(alan-turing-institute, huggingface)</td>\n",
       "      <td>(classification, 2D, image)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>11</th>\n",
       "      <td>huggingface-object-detection</td>\n",
       "      <td>None</td>\n",
       "      <td>(TaskEnum.object_detection,)</td>\n",
       "      <td>https://github.com/alan-turing-institute/scivi...</td>\n",
       "      <td>git+https://github.com/alan-turing-institute/s...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>False</td>\n",
       "      <td>(alan-turing-institute, huggingface)</td>\n",
       "      <td>()</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>12</th>\n",
       "      <td>detectreeRGB-forest</td>\n",
       "      <td>None</td>\n",
       "      <td>(TaskEnum.object_detection, TaskEnum.segmentat...</td>\n",
       "      <td>https://github.com/shmh40/detectreeRGB-treecro...</td>\n",
       "      <td>git+https://github.com/shmh40/detectreeRGB-tre...</td>\n",
       "      <td>image</td>\n",
       "      <td>True</td>\n",
       "      <td>False</td>\n",
       "      <td>(cambridge-university,)</td>\n",
       "      <td>(2D, satellite, remote-sensing, ecology, envir...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                            name  \\\n",
       "0                      model-000   \n",
       "1                      model-001   \n",
       "4          scivision-test-plugin   \n",
       "5                mapreader-plant   \n",
       "6             resnet50-plantkton   \n",
       "8              image-classifiers   \n",
       "9        huggingface-classifiers   \n",
       "11  huggingface-object-detection   \n",
       "12           detectreeRGB-forest   \n",
       "\n",
       "                                          description  \\\n",
       "0                                                None   \n",
       "1                                                None   \n",
       "4                                                None   \n",
       "5   automated detection of plant patches in images...   \n",
       "6   automated classification of plankton images co...   \n",
       "8                                                None   \n",
       "9                                                None   \n",
       "11                                               None   \n",
       "12                                               None   \n",
       "\n",
       "                                                tasks  \\\n",
       "0   (TaskEnum.object_detection, TaskEnum.segmentat...   \n",
       "1   (TaskEnum.segmentation, TaskEnum.thresholding,...   \n",
       "4                        (TaskEnum.object_detection,)   \n",
       "5   (TaskEnum.classificiation, TaskEnum.object_det...   \n",
       "6                         (TaskEnum.classificiation,)   \n",
       "8                         (TaskEnum.classificiation,)   \n",
       "9                         (TaskEnum.classificiation,)   \n",
       "11                       (TaskEnum.object_detection,)   \n",
       "12  (TaskEnum.object_detection, TaskEnum.segmentat...   \n",
       "\n",
       "                                                  url  \\\n",
       "0                https://github.com/stardist/stardist   \n",
       "1           https://github.com/danforthcenter/plantcv   \n",
       "4   https://github.com/alan-turing-institute/scivi...   \n",
       "5   https://github.com/alan-turing-institute/mapre...   \n",
       "6   https://github.com/alan-turing-institute/plank...   \n",
       "8   https://github.com/alan-turing-institute/scivi...   \n",
       "9   https://github.com/alan-turing-institute/scivi...   \n",
       "11  https://github.com/alan-turing-institute/scivi...   \n",
       "12  https://github.com/shmh40/detectreeRGB-treecro...   \n",
       "\n",
       "                                              pkg_url format  pretrained  \\\n",
       "0   git+https://github.com/stardist/stardist.git@main  image        True   \n",
       "1   git+https://github.com/danforthcenter/plantcv@...  image        True   \n",
       "4   git+https://github.com/alan-turing-institute/s...  image        True   \n",
       "5   git+https://github.com/alan-turing-institute/m...  image        True   \n",
       "6   git+https://github.com/alan-turing-institute/p...  image        True   \n",
       "8   git+https://github.com/alan-turing-institute/s...  image        True   \n",
       "9   git+https://github.com/alan-turing-institute/s...  image        True   \n",
       "11  git+https://github.com/alan-turing-institute/s...  image        True   \n",
       "12  git+https://github.com/shmh40/detectreeRGB-tre...  image        True   \n",
       "\n",
       "    labels_required                                        institution  \\\n",
       "0              True                                            (epfl,)   \n",
       "1              True                                  (danforthcenter,)   \n",
       "4             False                           (alan-turing-institute,)   \n",
       "5             False                           (alan-turing-institute,)   \n",
       "6             False  (alan-turing-institute, cefas, plankton-analyt...   \n",
       "8             False                           (alan-turing-institute,)   \n",
       "9             False               (alan-turing-institute, huggingface)   \n",
       "11            False               (alan-turing-institute, huggingface)   \n",
       "12            False                            (cambridge-university,)   \n",
       "\n",
       "                                                 tags  \n",
       "0   (2D, 3D, optical-microscopy, xray, microtomogr...  \n",
       "1   (2D, hyperspectral, multispectral, near-infrar...  \n",
       "4                                            (dummy,)  \n",
       "5   (2D, plant, phenotype, rgb, biology, agriculture)  \n",
       "6      (2D, plankton, ecology, environmental-science)  \n",
       "8                         (classification, 2D, image)  \n",
       "9                         (classification, 2D, image)  \n",
       "11                                                 ()  \n",
       "12  (2D, satellite, remote-sensing, ecology, envir...  "
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "compatible_models = default_catalog.compatible_models(\"data-003\").to_dataframe()\n",
    "compatible_models"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The `huggingface-classifiers` catalog entry can be used to load some of the most popular image classification models from [Hugging Face](https://huggingface.co/models?pipeline_tag=image-classification&sort=downloads). See the list of included models in the [model repo](https://github.com/alan-turing-institute/scivision_huggingface). As before, let's load one of the named models and run it on our test image:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/pwochner/opt/miniconda3/envs/scivision/lib/python3.9/site-packages/torch/functional.py:478: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at  /Users/runner/work/pytorch/pytorch/pytorch/aten/src/ATen/native/TensorShape.cpp:2895.)\n",
      "  return _VF.meshgrid(tensors, **kwargs)  # type: ignore[attr-defined]\n"
     ]
    }
   ],
   "source": [
    "huggingface_repo = models_catalog[models_catalog.name == \"huggingface-classifiers\"].url.item()\n",
    "microsoft_model = load_pretrained_model(huggingface_repo, model='microsoft_swin_tiny_patch4_window7_224', allow_install=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'Predicted class: isopod'"
      ]
     },
     "execution_count": 19,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "microsoft_model.predict(test_image)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}