empty_archs_seg_mn.ipynb

{"nbformat":4,"nbformat_minor":0,"metadata":{"accelerator":"GPU","colab":{"name":"dbunet_BN_segmentation_median_nerve.ipynb","provenance":[{"file_id":"1cD8zNIINsbHSqw8GOLkOO4cU44THdv6r","timestamp":1625762126160},{"file_id":"14PaQSWctrfqCgMbV1EdCNv-1tU0e9Ex5","timestamp":1625500295169}],"collapsed_sections":[]},"kernelspec":{"display_name":"Python 3","name":"python3"},"language_info":{"name":"python"}},"cells":[{"cell_type":"code","metadata":{"id":"i_yB1dFDXxBU"},"source":["from google.colab import drive\n","drive.mount('/content/drive')\n","\n","root_path = '/content/drive/MyDrive/CVDL'\n","dumps_path =  f\"{root_path}/dumps/\"\n","dataset  = f\"{root_path}/dataset/\"\n","\n","!cp \"$root_path/modules/lib.py\" .\n","!cp \"$root_path/modules/plots.py\" .\n","!cp \"$root_path/modules/metrics.py\" ."],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"WfhZ4xmMPKCe"},"source":["# Preprocessing"]},{"cell_type":"markdown","metadata":{"id":"0Lwyml4qW7ts"},"source":["## Imports"]},{"cell_type":"code","metadata":{"id":"_DKfoow7sh3u"},"source":["import cv2\n","import pickle\n","import matplotlib.pyplot as plt\n","import numpy as np\n","import cv2 as cv\n","import os\n","import tensorflow as tf\n","import importlib\n","from datetime import datetime\n","\n","from keras.initializers import glorot_uniform\n","\n","import lib\n","import plots\n","import metrics\n","\n","tf.keras.backend.clear_session()"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"Yh36iM9VW7uC"},"source":["## Parameters"]},{"cell_type":"code","metadata":{"id":"IYOYkQy_aNSD"},"source":["# Deafault Res: 606x468 (ar 101 : 78)\n","\n","# Tested Res's:\n","# 224x160 (ar 7 : 5)\n","# 384x288\n","# 576x432\n","# 512x384\n","input_res = (384, 288)\n","\n","seed = 42\n","\n","params = {\n","          'x':input_res[0],                                \n","          'y':input_res[1],                                \n","          'batch_size': 8,                        \n","          'n_channels_mask': 1,                            \n","          'n_channels': 1,                        \n","          'shuffle': True,              \n","          'learningRate': 1e-3,                      \n","          'momentum': 0.98,                        \n","          'nEpoches': 100,                        \n","          }\n","\n","ratio_test = 0.85\n","ratio_val  = 0.8\n","\n","lib.create_dump([params, seed], dumps_path + \"params_du\")\n","\n","type_ds  = ['Images','Masks']\n","modality = ['train','validation','test']\n","\n","# get path Img and Mask\n","path_img  = os.path.join(dataset, type_ds[0])\n","path_msk  = os.path.join(dataset, type_ds[1])\n","path_     = [path_img, path_msk]\n","\n","# get length imgs and proof that is equal with masks\n","length_imgs  = len(os.listdir(os.path.join(dataset,type_ds[0] + \"/\"))) \n","length_masks = len(os.listdir(os.path.join(dataset,type_ds[1] + \"/\")))\n","params['length_data'] = length_imgs,\n","\n","# get name of file imgs\n","imgs_list_path = os.listdir(path_img)"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"VKlQcWFDW7uE"},"source":["## Data loading "]},{"cell_type":"code","metadata":{"id":"Vmt_YNsFPssY"},"source":["# LOAD & SPLIT DATA ============================================================\n","mode = 1  # mode = 0: Random Split\n","          # mode = 1: Split for Patients\n","          \n","dataset_list = lib.get_dataset_list(\n","                                mode, \n","                                path_, \n","                                imgs_list_path, \n","                                params, \n","                                ratio_test, \n","                                ratio_val, \n","                                seed, \n","                                modality)\n","\n","# Dumping splitted data\n","lib.create_dump(dataset_list, f\"{dumps_path}data_rt{ratio_test}_rv{ratio_val}_{input_res[0]}x{input_res[1]}\")"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"9VS9PNzw6Pp2"},"source":["##Loading Data from Pickle"]},{"cell_type":"code","metadata":{"id":"OF4QPe3Rv57S"},"source":["dataset_list = pickle.load(open( f\"{dumps_path}data_rt{ratio_test}_rv{ratio_val}_{input_res[0]}x{input_res[1]}\", \"rb\" ))"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"TDi2fSk96V14"},"source":["##Data Augmentation"]},{"cell_type":"code","metadata":{"id":"_u3XQo-vW7uJ"},"source":["# DATA AUGMENTATION ============================================================\n","train_generator, val_generator = lib.data_augmentation(seed, dataset_list, params)\n","\n","params[\"length_training\"]   = len(dataset_list[0]['train'])\n","params['length_validation'] = len(dataset_list[0]['validation'])\n","\n","augmented_batch_plot = plots.plot_augmented_batch(train_generator)\n","plt.show()"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"eZMPcQNURjPd"},"source":["print(params[\"length_training\"])\n","print(params['length_validation'])\n","print(len(dataset_list[0]['test']))"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"AJRSxu-3W7uP"},"source":["# Model definition & training"]},{"cell_type":"markdown","metadata":{"id":"qaDOvDLTYOTH"},"source":["## Training"]},{"cell_type":"code","metadata":{"id":"pfFpxzd269xL"},"source":["params['learningRate'] = 1e-3\n","params['nEpoches'] = 100\n","\n","\n","models_name = ['unet', 'lightweight_unet', 'dbunet', 'dbunet_BN']\n","arch_name = models_name[4]\n","model_weights = f\"{root_path}/dataset/results/models/{arch_name}_Glorot_e{params['nEpoches']}_lr{params['learningRate']}\"\n","log_path = f\"{model_weights}/logs/fit/\" + datetime.now().strftime(\"%Y%m%d-%H%M%S\")\n","!rm -rf $log_path\n","!cp \"$root_path/modules/models/{arch_name}.py\" .\n","arch = importlib.import_module(arch_name)\n","\n","\n","input_size = (input_res[1], input_res[0], 1)\n","initializer = tf.keras.initializers.GlorotUniform()\n","model = arch.build_model(initializer, input_size)\n","model.summary()"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"WfKA5g_BtmcE"},"source":["loss      = metrics.bce_dice()\n","recall    = tf.keras.metrics.Recall()\n","precision = tf.keras.metrics.Precision()\n","metric    = [metrics.dice, metrics.iou, recall, precision]\n","\n","print(f'Model wil be saved in {model_weights}')\n","\n","train_history, model = lib.train(model_weights, log_path, params, model, loss, metric, train_generator, val_generator)"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"VPk54x9RYT4A"},"source":["## Evaluating"]},{"cell_type":"code","metadata":{"id":"7LY5fghx609Z"},"source":["results_tf = model.evaluate(dataset_list[0]['test'], dataset_list[1]['test'], return_dict = True)\n","train_figure = plots.plot_train_metric(train_history.history, params['nEpoches'], \"dice\")\n","plt.show()"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"p_hF8IosBm0R"},"source":["plt.imshow(dataset_list[1]['validation'][22, :,:, 0], cmap='gray')"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"-D-3kJvbaL9h"},"source":["threshold = 0.5\n","min_dice = 0.60\n","results_df, avg = plots.get_results(dataset_list, params, model, threshold)\n","print(results_df)\n","print('\\n')\n","print(f\"Threshold value: {threshold}\")\n","print(f\"Best dice: {max(results_df['dice'])}\")\n","print(f\"Average dice: {avg}\")\n","print(f\"Minimum requested dice value (bad samples): {min_dice}\")\n","good_res = results_df.loc[(results_df['dice'] >= min_dice)]\n","print(f\"Average dice without bad samples: {good_res['dice'].mean()}\")"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"XVaYmcK2bwVz"},"source":["bad_res = results_df.loc[(results_df['dice'] < min_dice)]\n","print(bad_res)\n","\n","num = bad_res['dice'].count()\n","den = results_df['dice'].count()\n","\n","print(num/den)\n","print('\\n')\n","print(\"Loading bad samples... :\")\n","print('\\n')\n","for idx in bad_res.index:\n","  print('\\n')\n","  print('\\n')\n","  print(f\"Image index {idx} - dice: {bad_res.loc[idx]}\")\n","  threshold = 0.5\n","  test_image = dataset_list[0]['test'][idx,:,:,0]\n","  test_mask = dataset_list[1]['test'][idx,:,:,0]\n","  test_results = lib.get_preditcions(test_image, test_mask, params, model, threshold)\n","  results_figure = plots.plot_test_results(test_results)\n","\n","  plt.show()"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"c0Dn6v6UcKum"},"source":["model.summary()"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"w2Pjy39gYnI5"},"source":["## Saving"]},{"cell_type":"code","metadata":{"id":"eYmRJJ9LCkRn"},"source":["name = 'doubleunet_d0.9027'\n","model.save(f'/content/drive/MyDrive/CVDL/trained_models/{name}/model.h5', save_format='h5')\n","lib.create_dump(train_history.history, f'/content/drive/MyDrive/CVDL/trained_models/{name}/train_history')"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"_C_dy_XSOFG9"},"source":["from tensorflow import keras\n","\n","c_objects = {\n","    \"bce_dice_loss\": metrics.bce_dice_loss,\n","    \"dice\": metrics.dice,\n","    \"iou\": metrics.iou,\n","    \"precision\": tf.keras.metrics.Precision(),\n","    \"recall\": tf.keras.metrics.Recall(),\n","}\n","\n","loaded_model = keras.models.load_model(f'/content/drive/MyDrive/CVDL/trained_models/{name}/model.h5', custom_objects=c_objects)\n","\n","loaded_model.compile(loss=metrics.bce_dice_loss, optimizer='adam', metrics=metric)\n","\n","results_tf = loaded_model.evaluate(dataset_list[0]['test'], dataset_list[1]['test'], return_dict = True)"],"execution_count":null,"outputs":[]}]}