sizhky

from snippets.loader import *
from PIL import Image
import xml.etree.ElementTree as ET
from torchvision import transforms

device = 'cuda'
voc_labels = ('aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable',
              'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor')
label_map = {k: v + 1 for v, k in enumerate(voc_labels)}
label_map['background'] = 0

sizhky

import sys; sys.path.append('/home/yyr/data/VOCdevkit/')
from load_data import VOCDataset, get_items, aug_trn
from snippets.loader import *

_2007_root = Path("/home/yyr/data/VOCdevkit/VOC2007")
_2012_root = Path("/home/yyr/data/VOCdevkit/VOC2012")
train_items = get_items(_2007_root, 'train') + get_items(_2012_root, 'train')
val_items   = get_items(_2007_root, 'val') + get_items(_2012_root, 'val')
logger.info(f'\n{len(train_items)} training images\n{len(val_items)} validation images')
x = VOCDataset(train_items, tfms=aug_trn)

sizhky

sizhky

project_url = 'https://github.com/sizhky/rcnn'
import os
if not os.path.exists('faster-rcnn'):
  !git clone --quiet {project_url} rcnn
  !pip install -q --upgrade imgaug fire torch_snippets
  from google.colab import files
  files.upload() # upload kaggle.json
  !mkdir -p ~/.kaggle
  !mv kaggle.json ~/.kaggle/
  !ls ~/.kaggle

sizhky

df = pd.read_csv('../validation-annotations-bbox.csv')
df['LabelName'] = df['LabelName'].map(lambda x: code2label[x])
unique_classes = df['LabelName'].value_counts()
unique_classes = unique_classes[unique_classes < 501]
df = df[df['LabelName'].map(lambda x: x in unique_classes)]
print(df.ImageID.nunique())

!mkdir -p open-images-mini
from tqdm import tqdm
for f in tqdm(df.ImageID.unique()):

sizhky

class OpenImages(Dataset):
  def __init__(self, image_folder, df):
    self.root = image_folder
    self.df = df
    self.unique_images = df['ImageID'].unique()
  def __len__(self): return len(self.unique_images)
  def __getitem__(self, ix):
    image_id = self.unique_images[ix]
    image_path = f'{self.root}/{image_id}.jpg'
    image = cv2.imread(image_path, 1)[...,::-1] # conver BGR to RGB

sizhky

jt -t grade3 -fs 95 -tfs 11 -nfs 115 -cellw 95% -T
jt -t onedork -fs 95 -tfs 11 -nfs 115 -cellw 88% -T -N

sizhky

import torch, math
import torch.nn as nn
import torch.nn.functional as F
class RetinaNet(nn.Module):
    num_anchors = 9

    def __init__(self, num_classes):
        super(RetinaNet, self).__init__()
        self.fpn = FPN50()
        self.num_classes = num_classes

sizhky

from IPython.core.magic import register_cell_magic
from IPython.display import HTML, display

def set_background(color):    
    script = (
        "var cell = this.closest('.jp-CodeCell');"
        "var editor = cell.querySelector('.jp-Editor');"
        "editor.style.background='{}';"
        "this.parentNode.removeChild(this)"
    ).format(color)

sizhky

document.getElementsByTagName("video")[0].playbackRate = 9001