{
"cells": [
{
"cell_type": "markdown",
"id": "412f12eb",
"metadata": {},
"source": [
"# Cityscapes PSPNet-style — загрузка готовых весов и быстрая оценка\n",
"\n",
"Этот ноутбук **не обучает** модель. Он:\n",
"1. загружает датасет,\n",
"2. строит ту же архитектуру,\n",
"3. загружает checkpoint из первого ноутбука,\n",
"4. делает инференс и считает метрики,\n",
"5. измеряет **полное время выполнения ноутбука** от начала до конца."
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "d8ec3c27",
"metadata": {},
"outputs": [],
"source": [
"import time\n",
"NOTEBOOK_T0 = time.perf_counter()\n",
"\n",
"import os\n",
"from pathlib import Path\n",
"import random\n",
"\n",
"import cv2\n",
"import numpy as np\n",
"import pandas as pd\n",
"import matplotlib.pyplot as plt\n",
"import tensorflow as tf\n",
"from tensorflow.keras import layers, models\n",
"\n",
"SEED = 42\n",
"random.seed(SEED)\n",
"np.random.seed(SEED)\n",
"tf.random.set_seed(SEED)"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "bb1ab3e5",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"DATA_ROOT = /home/konnilol/Downloads/cityscapes_data/cityscapes_data\n",
"VAL_DIR = /home/konnilol/Downloads/cityscapes_data/cityscapes_data/val\n",
"WEIGHTS = /home/konnilol/Downloads/work_cityscapes_pspnet/best_model.weights.h5\n"
]
}
],
"source": [
"# ------------------------------\n",
"# 1. Пути\n",
"# ------------------------------\n",
"MANUAL_DATA_ROOT = None\n",
"MANUAL_WEIGHTS_PATH = None # например: Path('./work_cityscapes_pspnet/best_model.weights.h5')\n",
"\n",
"DATA_CANDIDATES = [\n",
" Path('./cityscapes_data/cityscapes_data'),\n",
" Path('./cityscapes-image-pairs/cityscapes_data/cityscapes_data'),\n",
" Path('./data/cityscapes_data/cityscapes_data'),\n",
" Path('./dataset/cityscapes_data/cityscapes_data'),\n",
"]\n",
"\n",
"WEIGHT_CANDIDATES = [\n",
" Path('./work_cityscapes_pspnet/best_model.weights.h5'),\n",
" Path('./best_model.weights.h5'),\n",
"]\n",
"\n",
"DATA_ROOT = Path(MANUAL_DATA_ROOT) if MANUAL_DATA_ROOT is not None else None\n",
"if DATA_ROOT is None:\n",
" for p in DATA_CANDIDATES:\n",
" if p.exists():\n",
" DATA_ROOT = p\n",
" break\n",
"if DATA_ROOT is None:\n",
" raise FileNotFoundError('Не найдена папка cityscapes_data/cityscapes_data.')\n",
"\n",
"WEIGHTS_PATH = Path(MANUAL_WEIGHTS_PATH) if MANUAL_WEIGHTS_PATH is not None else None\n",
"if WEIGHTS_PATH is None:\n",
" for p in WEIGHT_CANDIDATES:\n",
" if p.exists():\n",
" WEIGHTS_PATH = p\n",
" break\n",
"if WEIGHTS_PATH is None:\n",
" raise FileNotFoundError(\n",
" 'Не найден файл best_model.weights.h5. Сначала запустите training-ноутбук.'\n",
" )\n",
"\n",
"VAL_DIR = DATA_ROOT / 'val'\n",
"print('DATA_ROOT =', DATA_ROOT.resolve())\n",
"print('VAL_DIR =', VAL_DIR.resolve())\n",
"print('WEIGHTS =', WEIGHTS_PATH.resolve())"
]
},
{
"cell_type": "code",
"execution_count": 10,
"id": "ba6f6264",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"val_imgs : (500, 128, 128, 3)\n",
"val_masks: (500, 128, 128, 3)\n"
]
}
],
"source": [
"# ------------------------------\n",
"# 2. Чтение данных\n",
"# ------------------------------\n",
"IMG_SIZE = 128\n",
"MAX_VAL = 512\n",
"BATCH_SIZE = 16\n",
"\n",
"def list_images(folder: Path):\n",
" files = sorted([p for p in folder.iterdir() if p.suffix.lower() in {'.png', '.jpg', '.jpeg'}])\n",
" return files\n",
"\n",
"def read_pair_image(path, out_size=IMG_SIZE):\n",
" img = cv2.imread(str(path))\n",
" if img is None:\n",
" raise ValueError(f'Не удалось прочитать файл: {path}')\n",
" img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0\n",
" h, w, _ = img.shape\n",
" mid = w // 2\n",
" image = img[:, :mid]\n",
" mask = img[:, mid:]\n",
" image = cv2.resize(image, (out_size, out_size), interpolation=cv2.INTER_LINEAR)\n",
" mask = cv2.resize(mask, (out_size, out_size), interpolation=cv2.INTER_NEAREST)\n",
" return image.astype(np.float32), mask.astype(np.float32)\n",
"\n",
"val_files = list_images(VAL_DIR)\n",
"if MAX_VAL:\n",
" val_files = val_files[:MAX_VAL]\n",
"\n",
"val_imgs, val_masks = [], []\n",
"for p in val_files:\n",
" x, y = read_pair_image(p, IMG_SIZE)\n",
" val_imgs.append(x)\n",
" val_masks.append(y)\n",
"\n",
"val_imgs = np.array(val_imgs, dtype=np.float32)\n",
"val_masks = np.array(val_masks, dtype=np.float32)\n",
"\n",
"print('val_imgs :', val_imgs.shape)\n",
"print('val_masks:', val_masks.shape)"
]
},
{
"cell_type": "code",
"execution_count": 11,
"id": "2aa81358",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Веса успешно загружены.\n",
"Архитектура совпадает с training-ноутбуком.\n"
]
}
],
"source": [
"# ------------------------------\n",
"# 3. Архитектура (ИДЕНТИЧНА training-ноутбуку)\n",
"# ------------------------------\n",
"def conv_block(x, filters, block_name):\n",
" f1, f2, f3 = filters\n",
" skip = x\n",
"\n",
" x = layers.Conv2D(f1, 1, padding='same', kernel_initializer='he_normal', name=f'{block_name}_a')(x)\n",
" x = layers.BatchNormalization(name=f'{block_name}_bn_a')(x)\n",
" x = layers.LeakyReLU(0.1, name=f'{block_name}_act_a')(x)\n",
"\n",
" x = layers.Conv2D(f2, 3, padding='same', dilation_rate=2, kernel_initializer='he_normal', name=f'{block_name}_b')(x)\n",
" x = layers.BatchNormalization(name=f'{block_name}_bn_b')(x)\n",
" x = layers.LeakyReLU(0.1, name=f'{block_name}_act_b')(x)\n",
"\n",
" x = layers.Conv2D(f3, 1, padding='same', kernel_initializer='he_normal', name=f'{block_name}_c')(x)\n",
" x = layers.BatchNormalization(name=f'{block_name}_bn_c')(x)\n",
"\n",
" skip = layers.Conv2D(f3, 1, padding='same', kernel_initializer='he_normal', name=f'{block_name}_skip')(skip)\n",
" skip = layers.BatchNormalization(name=f'{block_name}_bn_skip')(skip)\n",
"\n",
" x = layers.Add(name=f'{block_name}_add')([x, skip])\n",
" x = layers.ReLU(name=f'{block_name}_relu')(x)\n",
" return x\n",
"\n",
"\n",
"def pyramid_pooling(x, out_channels=32):\n",
" h = x.shape[1]\n",
" w = x.shape[2]\n",
"\n",
" b1 = layers.GlobalAveragePooling2D(name='ppm_gap')(x)\n",
" b1 = layers.Reshape((1, 1, x.shape[-1]))(b1)\n",
" b1 = layers.Conv2D(out_channels, 1, padding='same', activation='relu', name='ppm_red_1x1')(b1)\n",
" b1 = layers.UpSampling2D(size=(h, w), interpolation='bilinear', name='ppm_red_up')(b1)\n",
"\n",
" b2 = layers.AveragePooling2D(pool_size=(2, 2), name='ppm_pool_2')(x)\n",
" b2 = layers.Conv2D(out_channels, 1, padding='same', activation='relu', name='ppm_yellow_1x1')(b2)\n",
" b2 = layers.UpSampling2D(size=(2, 2), interpolation='bilinear', name='ppm_yellow_up')(b2)\n",
"\n",
" b3 = layers.AveragePooling2D(pool_size=(4, 4), name='ppm_pool_4')(x)\n",
" b3 = layers.Conv2D(out_channels, 1, padding='same', activation='relu', name='ppm_blue_1x1')(b3)\n",
" b3 = layers.UpSampling2D(size=(4, 4), interpolation='bilinear', name='ppm_blue_up')(b3)\n",
"\n",
" b4 = layers.AveragePooling2D(pool_size=(8, 8), name='ppm_pool_8')(x)\n",
" b4 = layers.Conv2D(out_channels, 1, padding='same', activation='relu', name='ppm_green_1x1')(b4)\n",
" b4 = layers.UpSampling2D(size=(8, 8), interpolation='bilinear', name='ppm_green_up')(b4)\n",
"\n",
" return layers.Concatenate(name='ppm_concat')([x, b1, b2, b3, b4])\n",
"\n",
"\n",
"def build_model():\n",
" inputs = layers.Input(shape=(IMG_SIZE, IMG_SIZE, 3), name='input')\n",
" x = layers.Conv2D(32, 3, padding='same', activation='relu', name='stem_conv')(inputs)\n",
" x = conv_block(x, [32, 32, 64], 'enc1')\n",
" x = layers.MaxPooling2D(name='pool1')(x)\n",
" x = conv_block(x, [64, 64, 128], 'enc2')\n",
" x = layers.MaxPooling2D(name='pool2')(x)\n",
" x = conv_block(x, [128, 128, 256], 'enc3')\n",
" x = pyramid_pooling(x, out_channels=32)\n",
" x = layers.UpSampling2D((2, 2), interpolation='bilinear', name='up1')(x)\n",
" x = layers.Conv2D(128, 3, padding='same', activation='relu', name='dec_conv1')(x)\n",
" x = layers.UpSampling2D((2, 2), interpolation='bilinear', name='up2')(x)\n",
" x = layers.Conv2D(64, 3, padding='same', activation='relu', name='dec_conv2')(x)\n",
" outputs = layers.Conv2D(3, 1, activation='sigmoid', name='rgb_mask')(x)\n",
" return models.Model(inputs, outputs, name='pspnet_compact')\n",
"\n",
"model = build_model()\n",
"model.load_weights(WEIGHTS_PATH)\n",
"print('Веса успешно загружены.')\n",
"print('Архитектура совпадает с training-ноутбуком.')"
]
},
{
"cell_type": "code",
"execution_count": 12,
"id": "5dc45a45",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\u001b[1m32/32\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m3s\u001b[0m 50ms/step\n"
]
},
{
"data": {
"application/vnd.positron.dataexplorer+json": {
"comm_id": "0e1e9671-61d2-4ddc-9f18-40caadd5c07c",
"shape": {
"columns": 2,
"rows": 5
},
"source": "pandas",
"title": "pandas",
"version": 1
},
"text/html": [
"\n",
"\n",
"
\n",
" \n",
" \n",
" | \n",
" metric | \n",
" value | \n",
"
\n",
" \n",
" \n",
" \n",
" | 0 | \n",
" mse | \n",
" 0.012189 | \n",
"
\n",
" \n",
" | 1 | \n",
" mae | \n",
" 0.053986 | \n",
"
\n",
" \n",
" | 2 | \n",
" ssim | \n",
" 0.685340 | \n",
"
\n",
" \n",
" | 3 | \n",
" pred_time_sec | \n",
" 2.852485 | \n",
"
\n",
" \n",
" | 4 | \n",
" fps | \n",
" 175.285746 | \n",
"
\n",
" \n",
"
\n",
"
"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# ------------------------------\n",
"# 5. Примеры\n",
"# ------------------------------\n",
"n_show = min(6, len(val_imgs))\n",
"idxs = np.linspace(0, len(val_imgs)-1, n_show, dtype=int)\n",
"\n",
"fig, axes = plt.subplots(n_show, 3, figsize=(10, 3 * n_show))\n",
"for row, idx in enumerate(idxs):\n",
" axes[row, 0].imshow(val_imgs[idx]); axes[row, 0].set_title('Image'); axes[row, 0].axis('off')\n",
" axes[row, 1].imshow(val_masks[idx]); axes[row, 1].set_title('GT'); axes[row, 1].axis('off')\n",
" axes[row, 2].imshow(np.clip(pred_masks[idx], 0, 1)); axes[row, 2].set_title('Pred'); axes[row, 2].axis('off')\n",
"plt.tight_layout()\n",
"plt.show()"
]
},
{
"cell_type": "code",
"execution_count": 14,
"id": "5034dd92",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Полное время выполнения ноутбука: 11.24 сек\n"
]
},
{
"data": {
"application/vnd.positron.dataexplorer+json": {
"comm_id": "47a9d844-ff04-4ad0-b71d-6df712a7dd3d",
"shape": {
"columns": 2,
"rows": 1
},
"source": "pandas",
"title": "pandas",
"version": 1
},
"text/html": [
"\n",
"\n",
"
\n",
" \n",
" \n",
" | \n",
" metric | \n",
" value | \n",
"
\n",
" \n",
" \n",
" \n",
" | 0 | \n",
" full_notebook_time_sec | \n",
" 11.242676 | \n",
"
\n",
" \n",
"
\n",
"