cascaded CNN training

train_cascadedCNN.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "import logging\n",
    "\n",
    "from hudl_beatrix.models import CascadedCNNModelConfig, cascadedcnn_model_factory\n",
    "from hudl_beatrix.optimizers import RmsPropOptimizerConfig, AdamOptimizerConfig, LearningRateConfig\n",
    "from hudl_beatrix.extractors import AlexNetLiteExtractorConfig, BranchExtractorConfig\n",
    "\n",
    "from hudl_beatrix.datasets.patch import PatchDataset\n",
    "\n",
    "from hudl_beatrix.datasets.downloaded import WatsonDatasetConfig\n",
    "from hudl_beatrix.datasets.patch import PatchDatasetConfig, patch_dataset_factory\n",
    "\n",
    "logging.info('Test')\n",
    "logger = logging.getLogger()\n",
    "logger.setLevel(logging.INFO)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "cascaded_cnn_config = CascadedCNNModelConfig(\n",
    "    base_extractor=AlexNetLiteExtractorConfig(),\n",
    "    branch_extractors=[\n",
    "         BranchExtractorConfig(branch_id=1, pool=True),\n",
    "#          BranchExtractorConfig(branch_id=2, pool=True),\n",
    "#          BranchExtractorConfig(branch_id=3, pool=False),\n",
    "#          BranchExtractorConfig(branch_id=4, pool=False)\n",
    "    ],\n",
    "    optimizer=AdamOptimizerConfig(learning_rate=LearningRateConfig(learning_rate=0.001))\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = cascadedcnn_model_factory(cascaded_cnn_config)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "estimator = tf.estimator.Estimator(model.__call__, model_dir='./train_cascade_dir/')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset_config =  [WatsonDatasetConfig(root='patch01', image_method='imageio')]\n",
    "patch_dataset_config = PatchDatasetConfig(datasets=dataset_config, \n",
    "                                         name='patch01', image_method='imageio', overwrite=False,\n",
    "                                         negative_ratio=1, batch_size=32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset = patch_dataset_factory(patch_dataset_config)\n",
    "dataset.extract_patches()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "input_fn_train = dataset.train_input_fn\n",
    "input_fn_dev = dataset.dev_input_fn"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "for epoch in range(2000):\n",
    "    estimator.train(input_fn=input_fn_train, steps = None)\n",
    "    estimator.evaluate(input_fn=input_fn_dev)\n",
    "    print (epoch)"
   ]
  }
 ],
 "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": "3.6.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}