Detectron2 – hra na špióny

15. 11. 2024 0:00 Jiří Raška

V předchozích dvou článcích jsem z knihovny Detectron2 využíval modely trénované na konkrétní datové sadě, což byla v mém případě Common Objects in Context. To vám může postačovat v těch případech, kdy chcete vyhledávat objekty, které tato datová sada zná. Co když ale chci hledat něco, co v té datové sadě není? Můžu existující modely naučit „novým kouskům“? Na to se zkusím podívat dnes.

Tak jako dříve budu potřebovat dostat knihovnu Detectron2 do svého prostředí, tak jen rychle:

!git clone 'https://github.com/facebookresearch/detectron2'
import sys, os, distutils.core

dist = distutils.core.run_setup("./detectron2/setup.py")
!python -m pip install {' '.join([f"'{x}'" for x in dist.install_requires])}
sys.path.insert(0, os.path.abspath('./detectron2'))

import torch, detectron2
!nvcc --version
TORCH_VERSION = ".".join(torch.__version__.split(".")[:2])
CUDA_VERSION = torch.__version__.split("+")[-1]
print("torch: ", TORCH_VERSION, "; cuda: ", CUDA_VERSION)
print("detectron2:", detectron2.__version__)

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2023 NVIDIA Corporation
Built on Wed_Nov_22_10:17:15_PST_2023
Cuda compilation tools, release 12.3, V12.3.107
Build cuda_12.3.r12.3/compiler.33567101_0
torch:  2.4 ; cuda:  2.4.0
detectron2: 0.6

Datová sada – Satellite Small Objects Computer Vision Project

Pro dnešní pokusy jsem si vybral datovou sadu Satellite Small Objects Computer Vision Project. Jedná se o sadu leteckých snímků s lokalizovanými objekty jako jsou auta, nákladní vozy, letadla, lodě a podobně. Níže je uvedeno několik příkladů, takže si můžete udělat představu, co v těch obrázcích asi je.

Tuto datovou sadu jsem si vybral také proto, že jí autoři poskytují také ve formátu COCO JSON. To mně výrazně ušetří práci, neboť takovou sadu mohu jednoduše zaregistrovat v knihovně Detectron2 a rovnou s ní pracovat. Pokud bych takto formátovanou a anotovanou datovou sadu neměl, pak bych si musel napsat vlastní funkci podle návodu Use Custom Datasets.

Nejdříve několik potřebných závislostí:

import detectron2
from detectron2.utils.logger import setup_logger
setup_logger()

import numpy as np
import cv2
import matplotlib.pyplot as plt

from detectron2 import model_zoo
from detectron2.engine import DefaultPredictor
from detectron2.config import get_cfg
from detectron2.utils.visualizer import Visualizer
from detectron2.data import MetadataCatalog, DatasetCatalog

Takto vypadá cesta k mé datové sadě:

DATASET_BASE = '/kaggle/input/satellite-small-objects-dataset-coco-json-format'

Díky tomu, že mám datovou sadu v COCO JSON formátu, mohu jí rovnou zaregistrovat pod názvem ‚satellite_train‘, ‚satellite_valid‘ a ‚satellite_test‘:

detectron2.data.datasets.register_coco_instances('satellite_train', {}, os.path.join(DATASET_BASE, 'train', '_annotations.coco.json'), os.path.join(DATASET_BASE, 'train'))
detectron2.data.datasets.register_coco_instances('satellite_valid', {}, os.path.join(DATASET_BASE, 'valid', '_annotations.coco.json'), os.path.join(DATASET_BASE, 'valid'))
detectron2.data.datasets.register_coco_instances('satellite_test', {}, os.path.join(DATASET_BASE, 'test', '_annotations.coco.json'), os.path.join(DATASET_BASE, 'test'))

Jak jsem slíbil dříve, zde je ukázka několika obrázků společně s jejich anotací (v tomto případě z datové množiny určené pro trénování):

import random

metadata = detectron2.data.MetadataCatalog.get('satellite_train')
dataset = detectron2.data.DatasetCatalog.get('satellite_train')

rows, cols = 2, 2

for row in range(rows):
    fig=plt.figure(figsize=(16, 8))
    for i, d in enumerate(random.sample(dataset, cols), start=1):
        fig.add_subplot(1, cols, i)

        image = cv2.imread(d['file_name'])

        visualizer = Visualizer(image, metadata, scale=1.0)
        out = visualizer.draw_dataset_dict(d)

        plt.imshow(out.get_image()[:,:,::-1])
    plt.show()
WARNING [10/28 15:38:29 d2.data.datasets.coco]: 
Category ids in annotations are not in [1, #categories]! We'll apply a mapping for you.

[10/28 15:38:29 d2.data.datasets.coco]: Loaded 4414 images in COCO format from /kaggle/input/satellite-small-objects-dataset-coco-json-format/train/_annotations.coco.json

cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_R_101_FPN_3x.yaml"))
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-Detection/faster_rcnn_R_101_FPN_3x.yaml")

cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 128

cfg.MODEL.ANCHOR_GENERATOR.SIZES = [[8, 16, 32, 64, 128]]
cfg.MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS = [[0.25, 0.5, 1.0, 2.0]]

cfg.MODEL.RPN.IN_FEATURES = ['p2', 'p3', 'p4', 'p5', 'p6']
cfg.INPUT.MIN_SIZE_TRAIN = 0
cfg.INPUT.MAX_SIZE_TRAIN = 99999
cfg.INPUT.RANDOM_FLIP = "horizontal"

cfg.DATASETS.TRAIN = ("satellite_train",)
cfg.DATASETS.TEST = ("satellite_valid",)

cfg.DATALOADER.NUM_WORKERS = 2

cfg.SOLVER.IMS_PER_BATCH = 2
cfg.SOLVER.BASE_LR = 0.00125
cfg.SOLVER.MAX_ITER = 1000

os.makedirs(cfg.OUTPUT_DIR, exist_ok=True)

Takže v tomto okamžiku již konfigurační objekt mám. Na jeho základě musím vytvořit objekt DefaultTrainer a s jeho pomocí spustit proces trénování:

from detectron2.engine import DefaultTrainer

trainer = DefaultTrainer(cfg)
trainer.resume_or_load(resume=False)
trainer.train()
[10/28 15:38:33 d2.engine.defaults]: Model:
GeneralizedRCNN(
  (backbone): FPN(
    (fpn_lateral2): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1))
    (fpn_output2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (fpn_lateral3): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1))
    (fpn_output3): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (fpn_lateral4): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1))
    (fpn_output4): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (fpn_lateral5): Conv2d(2048, 256, kernel_size=(1, 1), stride=(1, 1))
    (fpn_output5): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (top_block): LastLevelMaxPool()
    (bottom_up): ResNet(
      (stem): BasicStem(
        (conv1): Conv2d(
          3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False
          (norm): FrozenBatchNorm2d(num_features=64, eps=1e-05)
        )
      )
      (res2): Sequential(
        (0): BottleneckBlock(
          (shortcut): Conv2d(
            64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv1): Conv2d(
            64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=64, eps=1e-05)
          )
          (conv2): Conv2d(
            64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=64, eps=1e-05)
          )
          (conv3): Conv2d(
            64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
        )
        (1): BottleneckBlock(
          (conv1): Conv2d(
            256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=64, eps=1e-05)
          )
          (conv2): Conv2d(
            64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=64, eps=1e-05)
          )
          (conv3): Conv2d(
            64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
        )
        (2): BottleneckBlock(
          (conv1): Conv2d(
            256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=64, eps=1e-05)
          )
          (conv2): Conv2d(
            64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=64, eps=1e-05)
          )
          (conv3): Conv2d(
            64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
        )
      )
      (res3): Sequential(
        (0): BottleneckBlock(
          (shortcut): Conv2d(
            256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False
            (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
          )
          (conv1): Conv2d(
            256, 128, kernel_size=(1, 1), stride=(2, 2), bias=False
            (norm): FrozenBatchNorm2d(num_features=128, eps=1e-05)
          )
          (conv2): Conv2d(
            128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=128, eps=1e-05)
          )
          (conv3): Conv2d(
            128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
          )
        )
        (1): BottleneckBlock(
          (conv1): Conv2d(
            512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=128, eps=1e-05)
          )
          (conv2): Conv2d(
            128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=128, eps=1e-05)
          )
          (conv3): Conv2d(
            128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
          )
        )
        (2): BottleneckBlock(
          (conv1): Conv2d(
            512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=128, eps=1e-05)
          )
          (conv2): Conv2d(
            128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=128, eps=1e-05)
          )
          (conv3): Conv2d(
            128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
          )
        )
        (3): BottleneckBlock(
          (conv1): Conv2d(
            512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=128, eps=1e-05)
          )
          (conv2): Conv2d(
            128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=128, eps=1e-05)
          )
          (conv3): Conv2d(
            128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
          )
        )
      )
      (res4): Sequential(
        (0): BottleneckBlock(
          (shortcut): Conv2d(
            512, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
          (conv1): Conv2d(
            512, 256, kernel_size=(1, 1), stride=(2, 2), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (1): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (2): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (3): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (4): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (5): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (6): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (7): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (8): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (9): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (10): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (11): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (12): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (13): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (14): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (15): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (16): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (17): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (18): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (19): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (20): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (21): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
        (22): BottleneckBlock(
          (conv1): Conv2d(
            1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv2): Conv2d(
            256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=256, eps=1e-05)
          )
          (conv3): Conv2d(
            256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=1024, eps=1e-05)
          )
        )
      )
      (res5): Sequential(
        (0): BottleneckBlock(
          (shortcut): Conv2d(
            1024, 2048, kernel_size=(1, 1), stride=(2, 2), bias=False
            (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
          )
          (conv1): Conv2d(
            1024, 512, kernel_size=(1, 1), stride=(2, 2), bias=False
            (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
          )
          (conv2): Conv2d(
            512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
          )
          (conv3): Conv2d(
            512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
          )
        )
        (1): BottleneckBlock(
          (conv1): Conv2d(
            2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
          )
          (conv2): Conv2d(
            512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
          )
          (conv3): Conv2d(
            512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
          )
        )
        (2): BottleneckBlock(
          (conv1): Conv2d(
            2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
          )
          (conv2): Conv2d(
            512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=512, eps=1e-05)
          )
          (conv3): Conv2d(
            512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False
            (norm): FrozenBatchNorm2d(num_features=2048, eps=1e-05)
          )
        )
      )
    )
  )
  (proposal_generator): RPN(
    (rpn_head): StandardRPNHead(
      (conv): Conv2d(
        256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)
        (activation): ReLU()
      )
      (objectness_logits): Conv2d(256, 20, kernel_size=(1, 1), stride=(1, 1))
      (anchor_deltas): Conv2d(256, 80, kernel_size=(1, 1), stride=(1, 1))
    )
    (anchor_generator): DefaultAnchorGenerator(
      (cell_anchors): BufferList()
    )
  )
  (roi_heads): StandardROIHeads(
    (box_pooler): ROIPooler(
      (level_poolers): ModuleList(
        (0): ROIAlign(output_size=(7, 7), spatial_scale=0.25, sampling_ratio=0, aligned=True)
        (1): ROIAlign(output_size=(7, 7), spatial_scale=0.125, sampling_ratio=0, aligned=True)
        (2): ROIAlign(output_size=(7, 7), spatial_scale=0.0625, sampling_ratio=0, aligned=True)
        (3): ROIAlign(output_size=(7, 7), spatial_scale=0.03125, sampling_ratio=0, aligned=True)
      )
    )
    (box_head): FastRCNNConvFCHead(
      (flatten): Flatten(start_dim=1, end_dim=-1)
      (fc1): Linear(in_features=12544, out_features=1024, bias=True)
      (fc_relu1): ReLU()
      (fc2): Linear(in_features=1024, out_features=1024, bias=True)
      (fc_relu2): ReLU()
    )
    (box_predictor): FastRCNNOutputLayers(
      (cls_score): Linear(in_features=1024, out_features=81, bias=True)
      (bbox_pred): Linear(in_features=1024, out_features=320, bias=True)
    )
  )
)
WARNING [10/28 15:38:33 d2.data.datasets.coco]: 
Category ids in annotations are not in [1, #categories]! We'll apply a mapping for you.

[10/28 15:38:33 d2.data.datasets.coco]: Loaded 4414 images in COCO format from /kaggle/input/satellite-small-objects-dataset-coco-json-format/train/_annotations.coco.json
[10/28 15:38:34 d2.data.build]: Removed 3 images with no usable annotations. 4411 images left.
[10/28 15:38:34 d2.data.build]: Distribution of instances among all 12 categories:
|   category    | #instances   |  category  | #instances   |   category    | #instances   |
|:-------------:|:-------------|:----------:|:-------------|:-------------:|:-------------|
| car-plane-b.. | 0            |  airplane  | 2207         |     boat      | 7043         |
|      car      | 19151        | cargo ship | 396          |  fighter jet  | 2090         |
|  helicopter   | 1559         |    jet     | 2788         | military ai.. | 684          |
| propeller p.. | 2616         |   truck    | 6271         |     yacht     | 468          |
|               |              |            |              |               |              |
|     total     | 45273        |            |              |               |              |
[10/28 15:38:34 d2.data.dataset_mapper]: [DatasetMapper] Augmentations used in training: [ResizeShortestEdge(short_edge_length=(0, 0), max_size=99999, sample_style='choice'), RandomFlip()]
[10/28 15:38:34 d2.data.build]: Using training sampler TrainingSampler
[10/28 15:38:34 d2.data.common]: Serializing the dataset using: <class 'detectron2.data.common._TorchSerializedList'>
[10/28 15:38:34 d2.data.common]: Serializing 4411 elements to byte tensors and concatenating them all ...
[10/28 15:38:34 d2.data.common]: Serialized dataset takes 2.64 MiB
[10/28 15:38:34 d2.data.build]: Making batched data loader with batch_size=2
WARNING [10/28 15:38:34 d2.solver.build]: SOLVER.STEPS contains values larger than SOLVER.MAX_ITER. These values will be ignored.
[10/28 15:38:34 d2.checkpoint.detection_checkpoint]: [DetectionCheckpointer] Loading from https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_101_FPN_3x/137851257/model_final_f6e8b1.pkl ...
model_final_f6e8b1.pkl: 243MB [00:02, 105MB/s]                           
[10/28 15:38:36 d2.engine.train_loop]: Starting training from iteration 0
/opt/conda/lib/python3.10/site-packages/torch/functional.py:513: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at /usr/local/src/pytorch/aten/src/ATen/native/TensorShape.cpp:3609.)
  return _VF.meshgrid(tensors, **kwargs)  # type: ignore[attr-defined]
[10/28 15:38:43 d2.utils.events]:  eta: 0:02:37  iter: 19  total_loss: 1.825  loss_cls: 0.7122  loss_box_reg: 0.09764  loss_rpn_cls: 0.6996  loss_rpn_loc: 0.2088    time: 0.1660  last_time: 0.1597  data_time: 0.0112  last_data_time: 0.0050   lr: 2.4976e-05  max_mem: 1866M
[10/28 15:38:59 d2.utils.events]:  eta: 0:02:37  iter: 39  total_loss: 1.721  loss_cls: 0.7451  loss_box_reg: 0.1015  loss_rpn_cls: 0.6872  loss_rpn_loc: 0.1941    time: 0.1669  last_time: 0.1666  data_time: 0.0058  last_data_time: 0.0056   lr: 4.9951e-05  max_mem: 1866M
[10/28 15:39:03 d2.utils.events]:  eta: 0:02:32  iter: 59  total_loss: 1.472  loss_cls: 0.5056  loss_box_reg: 0.09716  loss_rpn_cls: 0.6772  loss_rpn_loc: 0.2121    time: 0.1654  last_time: 0.1566  data_time: 0.0058  last_data_time: 0.0054   lr: 7.4926e-05  max_mem: 1866M
[10/28 15:39:06 d2.utils.events]:  eta: 0:02:28  iter: 79  total_loss: 1.204  loss_cls: 0.3156  loss_box_reg: 0.0735  loss_rpn_cls: 0.6643  loss_rpn_loc: 0.1577    time: 0.1642  last_time: 0.1593  data_time: 0.0058  last_data_time: 0.0055   lr: 9.9901e-05  max_mem: 1925M
[10/28 15:39:09 d2.utils.events]:  eta: 0:02:24  iter: 99  total_loss: 1.254  loss_cls: 0.2477  loss_box_reg: 0.06834  loss_rpn_cls: 0.6454  loss_rpn_loc: 0.1483    time: 0.1634  last_time: 0.1634  data_time: 0.0056  last_data_time: 0.0058   lr: 0.00012488  max_mem: 1925M
[10/28 15:39:12 d2.utils.events]:  eta: 0:02:21  iter: 119  total_loss: 1.106  loss_cls: 0.2737  loss_box_reg: 0.05979  loss_rpn_cls: 0.6255  loss_rpn_loc: 0.1491    time: 0.1631  last_time: 0.1580  data_time: 0.0057  last_data_time: 0.0050   lr: 0.00014985  max_mem: 1925M
[10/28 15:39:16 d2.utils.events]:  eta: 0:02:18  iter: 139  total_loss: 1.153  loss_cls: 0.2629  loss_box_reg: 0.0927  loss_rpn_cls: 0.5686  loss_rpn_loc: 0.1345    time: 0.1628  last_time: 0.1574  data_time: 0.0051  last_data_time: 0.0044   lr: 0.00017483  max_mem: 1925M
[10/28 15:39:19 d2.utils.events]:  eta: 0:02:15  iter: 159  total_loss: 1.251  loss_cls: 0.388  loss_box_reg: 0.1793  loss_rpn_cls: 0.5367  loss_rpn_loc: 0.1631    time: 0.1628  last_time: 0.1673  data_time: 0.0053  last_data_time: 0.0046   lr: 0.0001998  max_mem: 1956M
[10/28 15:39:22 d2.utils.events]:  eta: 0:02:12  iter: 179  total_loss: 1.502  loss_cls: 0.5512  loss_box_reg: 0.31  loss_rpn_cls: 0.4673  loss_rpn_loc: 0.1734    time: 0.1628  last_time: 0.1609  data_time: 0.0049  last_data_time: 0.0049   lr: 0.00022478  max_mem: 2086M
[10/28 15:39:25 d2.utils.events]:  eta: 0:02:08  iter: 199  total_loss: 1.605  loss_cls: 0.6488  loss_box_reg: 0.355  loss_rpn_cls: 0.3793  loss_rpn_loc: 0.1327    time: 0.1626  last_time: 0.1589  data_time: 0.0050  last_data_time: 0.0046   lr: 0.00024975  max_mem: 2086M
[10/28 15:39:29 d2.utils.events]:  eta: 0:02:05  iter: 219  total_loss: 1.531  loss_cls: 0.643  loss_box_reg: 0.3996  loss_rpn_cls: 0.3507  loss_rpn_loc: 0.1723    time: 0.1625  last_time: 0.1636  data_time: 0.0053  last_data_time: 0.0055   lr: 0.00027473  max_mem: 2086M
[10/28 15:39:32 d2.utils.events]:  eta: 0:02:02  iter: 239  total_loss: 1.6  loss_cls: 0.5032  loss_box_reg: 0.3765  loss_rpn_cls: 0.2588  loss_rpn_loc: 0.1068    time: 0.1627  last_time: 0.1607  data_time: 0.0053  last_data_time: 0.0045   lr: 0.0002997  max_mem: 2086M
[10/28 15:39:35 d2.utils.events]:  eta: 0:01:59  iter: 259  total_loss: 1.322  loss_cls: 0.5066  loss_box_reg: 0.3371  loss_rpn_cls: 0.2791  loss_rpn_loc: 0.1465    time: 0.1625  last_time: 0.1600  data_time: 0.0049  last_data_time: 0.0047   lr: 0.00032468  max_mem: 2086M
[10/28 15:39:38 d2.utils.events]:  eta: 0:01:55  iter: 279  total_loss: 1.115  loss_cls: 0.4296  loss_box_reg: 0.3921  loss_rpn_cls: 0.2088  loss_rpn_loc: 0.07798    time: 0.1623  last_time: 0.1560  data_time: 0.0049  last_data_time: 0.0047   lr: 0.00034965  max_mem: 2086M
[10/28 15:39:42 d2.utils.events]:  eta: 0:01:52  iter: 299  total_loss: 1.451  loss_cls: 0.4969  loss_box_reg: 0.4802  loss_rpn_cls: 0.2417  loss_rpn_loc: 0.1484    time: 0.1622  last_time: 0.1589  data_time: 0.0051  last_data_time: 0.0047   lr: 0.00037463  max_mem: 2086M
[10/28 15:39:45 d2.utils.events]:  eta: 0:01:49  iter: 319  total_loss: 1.529  loss_cls: 0.6149  loss_box_reg: 0.4817  loss_rpn_cls: 0.2397  loss_rpn_loc: 0.1353    time: 0.1622  last_time: 0.1617  data_time: 0.0053  last_data_time: 0.0061   lr: 0.0003996  max_mem: 2086M
[10/28 15:39:48 d2.utils.events]:  eta: 0:01:46  iter: 339  total_loss: 1.314  loss_cls: 0.5094  loss_box_reg: 0.3127  loss_rpn_cls: 0.2408  loss_rpn_loc: 0.1668    time: 0.1622  last_time: 0.1646  data_time: 0.0054  last_data_time: 0.0056   lr: 0.00042458  max_mem: 2086M
[10/28 15:39:51 d2.utils.events]:  eta: 0:01:42  iter: 359  total_loss: 1.264  loss_cls: 0.4988  loss_box_reg: 0.4391  loss_rpn_cls: 0.212  loss_rpn_loc: 0.1118    time: 0.1622  last_time: 0.1619  data_time: 0.0054  last_data_time: 0.0054   lr: 0.00044955  max_mem: 2086M
[10/28 15:39:55 d2.utils.events]:  eta: 0:01:39  iter: 379  total_loss: 1.273  loss_cls: 0.4243  loss_box_reg: 0.4456  loss_rpn_cls: 0.1561  loss_rpn_loc: 0.1081    time: 0.1622  last_time: 0.1565  data_time: 0.0052  last_data_time: 0.0050   lr: 0.00047453  max_mem: 2086M
[10/28 15:39:58 d2.utils.events]:  eta: 0:01:36  iter: 399  total_loss: 1.378  loss_cls: 0.5405  loss_box_reg: 0.4047  loss_rpn_cls: 0.1609  loss_rpn_loc: 0.1498    time: 0.1621  last_time: 0.1574  data_time: 0.0051  last_data_time: 0.0046   lr: 0.0004995  max_mem: 2086M
[10/28 15:40:01 d2.utils.events]:  eta: 0:01:33  iter: 419  total_loss: 1.282  loss_cls: 0.526  loss_box_reg: 0.424  loss_rpn_cls: 0.196  loss_rpn_loc: 0.1336    time: 0.1623  last_time: 0.1944  data_time: 0.0057  last_data_time: 0.0060   lr: 0.00052448  max_mem: 2086M
[10/28 15:40:04 d2.utils.events]:  eta: 0:01:30  iter: 439  total_loss: 1.035  loss_cls: 0.3585  loss_box_reg: 0.4496  loss_rpn_cls: 0.1151  loss_rpn_loc: 0.09004    time: 0.1624  last_time: 0.1592  data_time: 0.0054  last_data_time: 0.0046   lr: 0.00054945  max_mem: 2086M
[10/28 15:40:08 d2.utils.events]:  eta: 0:01:26  iter: 459  total_loss: 1.464  loss_cls: 0.5679  loss_box_reg: 0.5142  loss_rpn_cls: 0.1452  loss_rpn_loc: 0.1443    time: 0.1623  last_time: 0.1730  data_time: 0.0051  last_data_time: 0.0052   lr: 0.00057443  max_mem: 2086M
[10/28 15:40:11 d2.utils.events]:  eta: 0:01:23  iter: 479  total_loss: 1.309  loss_cls: 0.4826  loss_box_reg: 0.4617  loss_rpn_cls: 0.1588  loss_rpn_loc: 0.1444    time: 0.1623  last_time: 0.1648  data_time: 0.0056  last_data_time: 0.0056   lr: 0.0005994  max_mem: 2086M
[10/28 15:40:14 d2.utils.events]:  eta: 0:01:20  iter: 499  total_loss: 1.148  loss_cls: 0.4078  loss_box_reg: 0.4352  loss_rpn_cls: 0.1354  loss_rpn_loc: 0.1462    time: 0.1623  last_time: 0.1558  data_time: 0.0054  last_data_time: 0.0047   lr: 0.00062438  max_mem: 2123M
[10/28 15:40:17 d2.utils.events]:  eta: 0:01:17  iter: 519  total_loss: 1.164  loss_cls: 0.4361  loss_box_reg: 0.4807  loss_rpn_cls: 0.1038  loss_rpn_loc: 0.1119    time: 0.1622  last_time: 0.1608  data_time: 0.0052  last_data_time: 0.0049   lr: 0.00064935  max_mem: 2123M
[10/28 15:40:21 d2.utils.events]:  eta: 0:01:13  iter: 539  total_loss: 1.286  loss_cls: 0.418  loss_box_reg: 0.566  loss_rpn_cls: 0.1381  loss_rpn_loc: 0.1263    time: 0.1621  last_time: 0.1613  data_time: 0.0050  last_data_time: 0.0056   lr: 0.00067433  max_mem: 2123M
[10/28 15:40:24 d2.utils.events]:  eta: 0:01:10  iter: 559  total_loss: 1.287  loss_cls: 0.4525  loss_box_reg: 0.5101  loss_rpn_cls: 0.1441  loss_rpn_loc: 0.1599    time: 0.1622  last_time: 0.1578  data_time: 0.0056  last_data_time: 0.0055   lr: 0.0006993  max_mem: 2123M
[10/28 15:40:27 d2.utils.events]:  eta: 0:01:07  iter: 579  total_loss: 1.219  loss_cls: 0.4701  loss_box_reg: 0.4964  loss_rpn_cls: 0.1535  loss_rpn_loc: 0.1094    time: 0.1622  last_time: 0.1700  data_time: 0.0054  last_data_time: 0.0056   lr: 0.00072428  max_mem: 2123M
[10/28 15:40:30 d2.utils.events]:  eta: 0:01:04  iter: 599  total_loss: 1.412  loss_cls: 0.4726  loss_box_reg: 0.4421  loss_rpn_cls: 0.1535  loss_rpn_loc: 0.16    time: 0.1622  last_time: 0.1614  data_time: 0.0055  last_data_time: 0.0053   lr: 0.00074925  max_mem: 2123M
[10/28 15:40:34 d2.utils.events]:  eta: 0:01:01  iter: 619  total_loss: 1.179  loss_cls: 0.4391  loss_box_reg: 0.4381  loss_rpn_cls: 0.1475  loss_rpn_loc: 0.1388    time: 0.1624  last_time: 0.1759  data_time: 0.0062  last_data_time: 0.0072   lr: 0.00077423  max_mem: 2123M
[10/28 15:40:37 d2.utils.events]:  eta: 0:00:57  iter: 639  total_loss: 1.246  loss_cls: 0.5142  loss_box_reg: 0.4569  loss_rpn_cls: 0.146  loss_rpn_loc: 0.1369    time: 0.1624  last_time: 0.1584  data_time: 0.0053  last_data_time: 0.0054   lr: 0.0007992  max_mem: 2123M
[10/28 15:40:40 d2.utils.events]:  eta: 0:00:54  iter: 659  total_loss: 1.251  loss_cls: 0.4604  loss_box_reg: 0.4832  loss_rpn_cls: 0.1025  loss_rpn_loc: 0.09686    time: 0.1623  last_time: 0.1597  data_time: 0.0053  last_data_time: 0.0053   lr: 0.00082418  max_mem: 2123M
[10/28 15:40:44 d2.utils.events]:  eta: 0:00:51  iter: 679  total_loss: 1.446  loss_cls: 0.5317  loss_box_reg: 0.5111  loss_rpn_cls: 0.1562  loss_rpn_loc: 0.1607    time: 0.1623  last_time: 0.1577  data_time: 0.0052  last_data_time: 0.0049   lr: 0.00084915  max_mem: 2123M
[10/28 15:40:47 d2.utils.events]:  eta: 0:00:48  iter: 699  total_loss: 1.299  loss_cls: 0.4848  loss_box_reg: 0.5071  loss_rpn_cls: 0.1127  loss_rpn_loc: 0.1603    time: 0.1623  last_time: 0.1588  data_time: 0.0050  last_data_time: 0.0052   lr: 0.00087413  max_mem: 2123M
[10/28 15:40:50 d2.utils.events]:  eta: 0:00:45  iter: 719  total_loss: 1.192  loss_cls: 0.493  loss_box_reg: 0.4684  loss_rpn_cls: 0.1286  loss_rpn_loc: 0.172    time: 0.1622  last_time: 0.1589  data_time: 0.0052  last_data_time: 0.0055   lr: 0.0008991  max_mem: 2123M
[10/28 15:40:53 d2.utils.events]:  eta: 0:00:41  iter: 739  total_loss: 1.123  loss_cls: 0.4149  loss_box_reg: 0.4803  loss_rpn_cls: 0.1217  loss_rpn_loc: 0.1482    time: 0.1622  last_time: 0.1583  data_time: 0.0053  last_data_time: 0.0059   lr: 0.00092408  max_mem: 2123M
[10/28 15:40:56 d2.utils.events]:  eta: 0:00:38  iter: 759  total_loss: 1.068  loss_cls: 0.3793  loss_box_reg: 0.4372  loss_rpn_cls: 0.1  loss_rpn_loc: 0.1231    time: 0.1621  last_time: 0.1550  data_time: 0.0055  last_data_time: 0.0044   lr: 0.00094905  max_mem: 2123M
[10/28 15:41:00 d2.utils.events]:  eta: 0:00:35  iter: 779  total_loss: 1.191  loss_cls: 0.3968  loss_box_reg: 0.4612  loss_rpn_cls: 0.1292  loss_rpn_loc: 0.1381    time: 0.1620  last_time: 0.1595  data_time: 0.0048  last_data_time: 0.0047   lr: 0.00097403  max_mem: 2123M
[10/28 15:41:03 d2.utils.events]:  eta: 0:00:32  iter: 799  total_loss: 1.227  loss_cls: 0.4524  loss_box_reg: 0.4939  loss_rpn_cls: 0.1087  loss_rpn_loc: 0.1238    time: 0.1620  last_time: 0.1556  data_time: 0.0055  last_data_time: 0.0051   lr: 0.000999  max_mem: 2123M
[10/28 15:41:06 d2.utils.events]:  eta: 0:00:28  iter: 819  total_loss: 1.228  loss_cls: 0.4673  loss_box_reg: 0.503  loss_rpn_cls: 0.09038  loss_rpn_loc: 0.1221    time: 0.1620  last_time: 0.1580  data_time: 0.0060  last_data_time: 0.0052   lr: 0.001024  max_mem: 2123M
[10/28 15:41:09 d2.utils.events]:  eta: 0:00:25  iter: 839  total_loss: 1.257  loss_cls: 0.4763  loss_box_reg: 0.5219  loss_rpn_cls: 0.1317  loss_rpn_loc: 0.1374    time: 0.1620  last_time: 0.1568  data_time: 0.0053  last_data_time: 0.0060   lr: 0.001049  max_mem: 2123M
[10/28 15:41:13 d2.utils.events]:  eta: 0:00:22  iter: 859  total_loss: 1.251  loss_cls: 0.4692  loss_box_reg: 0.4617  loss_rpn_cls: 0.09215  loss_rpn_loc: 0.1187    time: 0.1620  last_time: 0.1581  data_time: 0.0047  last_data_time: 0.0047   lr: 0.0010739  max_mem: 2123M
[10/28 15:41:16 d2.utils.events]:  eta: 0:00:19  iter: 879  total_loss: 1.212  loss_cls: 0.4833  loss_box_reg: 0.4908  loss_rpn_cls: 0.09467  loss_rpn_loc: 0.1313    time: 0.1619  last_time: 0.1583  data_time: 0.0050  last_data_time: 0.0048   lr: 0.0010989  max_mem: 2125M
[10/28 15:41:19 d2.utils.events]:  eta: 0:00:16  iter: 899  total_loss: 1.213  loss_cls: 0.4254  loss_box_reg: 0.4659  loss_rpn_cls: 0.09542  loss_rpn_loc: 0.107    time: 0.1619  last_time: 0.1677  data_time: 0.0050  last_data_time: 0.0060   lr: 0.0011239  max_mem: 2125M
[10/28 15:41:22 d2.utils.events]:  eta: 0:00:12  iter: 919  total_loss: 1.105  loss_cls: 0.3572  loss_box_reg: 0.4378  loss_rpn_cls: 0.07979  loss_rpn_loc: 0.1088    time: 0.1619  last_time: 0.1618  data_time: 0.0052  last_data_time: 0.0045   lr: 0.0011489  max_mem: 2125M
[10/28 15:41:26 d2.utils.events]:  eta: 0:00:09  iter: 939  total_loss: 1.282  loss_cls: 0.4933  loss_box_reg: 0.5322  loss_rpn_cls: 0.1185  loss_rpn_loc: 0.1234    time: 0.1619  last_time: 0.1608  data_time: 0.0051  last_data_time: 0.0045   lr: 0.0011738  max_mem: 2125M
[10/28 15:41:29 d2.utils.events]:  eta: 0:00:06  iter: 959  total_loss: 1.159  loss_cls: 0.4409  loss_box_reg: 0.5294  loss_rpn_cls: 0.08599  loss_rpn_loc: 0.1322    time: 0.1619  last_time: 0.1596  data_time: 0.0049  last_data_time: 0.0046   lr: 0.0011988  max_mem: 2125M
[10/28 15:41:32 d2.utils.events]:  eta: 0:00:03  iter: 979  total_loss: 1.227  loss_cls: 0.4864  loss_box_reg: 0.5101  loss_rpn_cls: 0.07952  loss_rpn_loc: 0.1065    time: 0.1619  last_time: 0.1614  data_time: 0.0058  last_data_time: 0.0060   lr: 0.0012238  max_mem: 2125M
[10/28 15:41:36 d2.utils.events]:  eta: 0:00:00  iter: 999  total_loss: 1.234  loss_cls: 0.4387  loss_box_reg: 0.5025  loss_rpn_cls: 0.09562  loss_rpn_loc: 0.1306    time: 0.1619  last_time: 0.1593  data_time: 0.0055  last_data_time: 0.0054   lr: 0.0012488  max_mem: 2125M
[10/28 15:41:36 d2.engine.hooks]: Overall training speed: 998 iterations in 0:02:41 (0.1619 s / it)
[10/28 15:41:36 d2.engine.hooks]: Total training time: 0:02:56 (0:00:14 on hooks)
WARNING [10/28 15:41:36 d2.data.datasets.coco]: 
Category ids in annotations are not in [1, #categories]! We'll apply a mapping for you.

[10/28 15:41:36 d2.data.datasets.coco]: Loaded 420 images in COCO format from /kaggle/input/satellite-small-objects-dataset-coco-json-format/valid/_annotations.coco.json
[10/28 15:41:36 d2.data.build]: Distribution of instances among all 12 categories:
|   category    | #instances   |  category  | #instances   |   category    | #instances   |
|:-------------:|:-------------|:----------:|:-------------|:-------------:|:-------------|
| car-plane-b.. | 0            |  airplane  | 176          |     boat      | 719          |
|      car      | 1731         | cargo ship | 46           |  fighter jet  | 242          |
|  helicopter   | 125          |    jet     | 198          | military ai.. | 82           |
| propeller p.. | 129          |   truck    | 545          |     yacht     | 31           |
|               |              |            |              |               |              |
|     total     | 4024         |            |              |               |              |
[10/28 15:41:36 d2.data.dataset_mapper]: [DatasetMapper] Augmentations used in inference: [ResizeShortestEdge(short_edge_length=(800, 800), max_size=1333, sample_style='choice')]
[10/28 15:41:36 d2.data.common]: Serializing the dataset using: <class 'detectron2.data.common._TorchSerializedList'>
[10/28 15:41:36 d2.data.common]: Serializing 420 elements to byte tensors and concatenating them all ...
[10/28 15:41:36 d2.data.common]: Serialized dataset takes 0.26 MiB
WARNING [10/28 15:41:36 d2.engine.defaults]: No evaluator found. Use `DefaultTrainer.test(evaluators=)`, or implement its `build_evaluator` method.

Na trénování jsem měl vyhrazeno 1000 iterací. Výsledkem je model zapsaný do souboru model_final.pth.

Ověření výsledků modelu

A nyní se zkusím podívat, jak jsem byl při trénování modelu úspěšný.

Pro predikci na testovací sadě budu potřebovat vytvořit objekt DefaultPredictor. Na jeho vytvoření potřebuji konfiguraci. Tu ale již mám, vytvořil jsem si jí předtím, než jsem začal s trénováním. Pouze potřebuji nastavit nové váhy modelu, proto ta změna parametru cfg.MODEL.WEIGHTS.

Dále si z testovací sady vyberu náhodně několik vzorků – satelitních obrázků. Každý takový satelitní obrázek si zobrazím dvakrát vedle sebe. Na levé straně je snímek včetně anotací definovaných v datové sadě (tedy to je dle autorů datové sady pravda). Na pravé straně je pak tento snímek s anotacemi predikovanými modelem (to je tedy to, co vidí můj model).

A zde je výsledek:

cfg.MODEL.WEIGHTS = os.path.join(cfg.OUTPUT_DIR, "model_final.pth")

predictor = DefaultPredictor(cfg)

metadata = detectron2.data.MetadataCatalog.get('satellite_test')
dataset = detectron2.data.DatasetCatalog.get('satellite_test')

for sample in random.sample(dataset, 5):
    image = cv2.imread(sample["file_name"])

    outputs = predictor(image)

    v_gt = Visualizer(image, metadata=metadata, scale=1.0)
    gt_vis = v_gt.draw_dataset_dict(sample)

    v_pred = Visualizer(image, metadata=metadata, scale=1.0)
    pred_vis = v_pred.draw_instance_predictions(outputs["instances"].to("cpu"))

    plt.figure(figsize=(16, 8))
    plt.subplot(1, 2, 1)
    plt.imshow(gt_vis.get_image()[:,:,::-1])
    plt.title("Ground Truth")

    plt.subplot(1, 2, 2)
    plt.imshow(pred_vis.get_image()[:,:,::-1])
    plt.title("Predictions")
    plt.show()
[10/28 15:41:37 d2.checkpoint.detection_checkpoint]: [DetectionCheckpointer] Loading from ./output/model_final.pth ...
Category ids in annotations are not in [1, #categories]! We'll apply a mapping for you.

[10/28 15:41:38 d2.data.datasets.coco]: Loaded 208 images in COCO format from /kaggle/input/satellite-small-objects-dataset-coco-json-format/test/_annotations.coco.json