如何使用900万张开放图像训练600类图片分类器

import sys
import os
import pandas as pd
import requests

from tqdm import tqdm
import ratelim
from checkpoints import checkpoints
checkpoints.enable()


def download(categories):
    # Download the metadata
    kwargs = {'header': None, 'names': ['LabelID', 'LabelName']}
    orig_url = "https://storage.googleapis.com/openimages/2018_04/class-descriptions-boxable.csv"
    class_names = pd.read_csv(orig_url, **kwargs)
    orig_url = "https://storage.googleapis.com/openimages/2018_04/train/train-annotations-bbox.csv"
    train_boxed = pd.read_csv(orig_url)
    orig_url = "https://storage.googleapis.com/openimages/2018_04/train/train-images-boxable-with-rotation.csv"
    image_ids = pd.read_csv(orig_url)

    # Get category IDs for the given categories and sub-select train_boxed with them.
    label_map = dict(class_names.set_index('LabelName').loc[categories, 'LabelID']
                     .to_frame().reset_index().set_index('LabelID')['LabelName'])
    label_values = set(label_map.keys())
    relevant_training_images = train_boxed[train_boxed.LabelName.isin(label_values)]

    # Start from prior results if they exist and are specified, otherwise start from scratch.
    relevant_flickr_urls = (relevant_training_images.set_index('ImageID')
                            .join(image_ids.set_index('ImageID'))
                            .loc[:, 'OriginalURL'])
    relevant_flickr_img_metadata = (relevant_training_images.set_index('ImageID').loc[relevant_flickr_urls.index]
                                    .pipe(lambda df: df.assign(LabelValue=df.LabelName.map(lambda v: label_map[v]))))
    remaining_todo = len(relevant_flickr_urls) if checkpoints.results is None else\
        len(relevant_flickr_urls) - len(checkpoints.results)

    # Download the images
    with tqdm(total=remaining_todo) as progress_bar:
        relevant_image_requests = relevant_flickr_urls.safe_map(lambda url: _download_image(url, progress_bar))
        progress_bar.close()

    # Write the images to files, adding them to the package as we go along.
    if not os.path.isdir("temp/"):
        os.mkdir("temp/")
    for ((_, r), (_, url), (_, meta)) in zip(relevant_image_requests.iteritems(), relevant_flickr_urls.iteritems(),
                                             relevant_flickr_img_metadata.iterrows()):
        image_name = url.split("/")[-1]
        image_label = meta['LabelValue']

        _write_image_file(r, image_name)


@ratelim.patient(5, 5)
def _download_image(url, pbar):
    """Download a single image from a URL, rate-limited to once per second"""
    r = requests.get(url)
    r.raise_for_status()
    pbar.update(1)
    return r


def _write_image_file(r, image_name):
    """Write an image to a file"""
    filename = f"temp/{image_name}"
    with open(filename, "wb") as f:
        f.write(r.content)
        
        
if __name__ == '__main__':
  categories = sys.argv[1:]
  download(categories)

$ git clone https://github.com/quiltdata/open-images.git
$ cd open-images/
$ conda env create -f environment.yml
$ source activate quilt-open-images-dev
$ cd src/openimager/
$ python openimager.py "Sandwiches" "Hamburgers"

import matplotlib.pyplot as plt
from matplotlib.image import imread
%matplotlib inline
import os

fig, axarr = plt.subplots(1, 5, figsize=(24, 4))
for i, img in enumerate(os.listdir('../data/images/')[:5]):
    axarr[i].imshow(imread('../data/images/' + img))

images_cropped/
    sandwich/
        some_image.jpg
        some_other_image.jpg
        ...
    hamburger/
        yet_another_image.jpg
        ...

from keras.preprocessing.image import ImageDataGenerator

train_datagen = ImageDataGenerator(
    rotation_range=40,
    width_shift_range=0.2,
    height_shift_range=0.2,
    rescale=1/255,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True,
    fill_mode='nearest'
)

test_datagen = ImageDataGenerator(
    rescale=1/255
)

train_generator = train_datagen.flow_from_directory(
    '../data/images_cropped/quilt/open_images/',
    target_size=(128, 128),
    batch_size=16,
    class_mode='binary'
)

validation_generator = test_datagen.flow_from_directory(
    '../data/images_cropped/quilt/open_images/',
    target_size=(128, 128),
    batch_size=16,
    class_mode='binary'
)

from keras.models import Sequential
from keras.layers import Conv2D, MaxPooling2D
from keras.layers import Activation, Dropout, Flatten, Dense
from keras.losses import binary_crossentropy
from keras.callbacks import EarlyStopping
from keras.optimizers import RMSprop


model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3), input_shape=(128, 128, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(32, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Flatten())  # this converts our 3D feature maps to 1D feature vectors
model.add(Dense(64, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(1))
model.add(Activation('sigmoid'))

model.compile(loss=binary_crossentropy,
              optimizer=RMSprop(lr=0.0005),  # half of the default lr
              metrics=['accuracy'])

import pathlib

sample_size = len(list(pathlib.Path('../data/images_cropped/').rglob('./*')))
batch_size = 16

hist = model.fit_generator(
    train_generator,
    steps_per_epoch=sample_size // batch_size,
    epochs=50,
    validation_data=validation_generator,
    validation_steps=round(sample_size * 0.2) // batch_size,
    callbacks=[EarlyStopping(monitor='val_loss', min_delta=0, patience=4)]
)

model.save("clf.h5")

precision    recall  f1-score   support

           0       0.90      0.59      0.71      1399
           1       0.64      0.92      0.75      1109   
   
   micro avg       0.73      0.73      0.73      2508
   macro avg       0.77      0.75      0.73      2508
weighted avg       0.78      0.73      0.73      2508

git clone 
conda env create -f open-images/environment.yml
source activate quilt-open-images-dev
python -c "import t4; t4.Package.install('quilt/open_images', dest='open-images/', registry='s3://quilt-example')"