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

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

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

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

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

sizhky

8x+4y+z = 96
3x+2y+4z = 53
9x+2y+z = 119

sizhky

Inception Networks. Role of Auxiliary loss in Training and Inference

• Deep networks suffer the problem of vanishing gradient, i.e., gradients near to the loss layer are much larger and as they get distributed to the previous layers backward, the magnitude of gradients diminish.

Training

• Inception tackles this problem by using the concept of auxiliary losses where (see fig) it branches out convolution outputs from 4a and 4d into their respective mini classification heads that employ standard softmax - predicting on the same classes as the main task

• Hence during training, there will be a strong gradient flow from loss:0 to adjust layers 4a (and preceding layers) more aggressively. Similar is the case with layers "4d,4c and 4b" where the gradients are high from "loss:1" and for layers 4e onwards which are closer to "loss:2".

sizhky

import time
info = lambda report: '\t'.join([f'{k}: {v:.3f}' for k,v in report.items()])

def report_metrics(pos, **report):
    elapsed = time.time()-start
    end = report.pop('end','\n')
    elapsed = '\t({:.2f}s - {:.2f}s remaining)'.format(time.time()-start, ((n_epochs-pos)/pos)*elapsed)
    current_iteration = f'EPOCH: {pos:.3f}\t'
    print(current_iteration + info(report) + elapsed, end=end)