【深度学习TPU+Keras+Tensorflow+EfficientNetB7】kaggle竞赛 使用TPU对104种花朵进行分类 第十八次尝试 99.9%准确率

版本更新

以下准确率全都是验证集准确率，和比赛提交以后的准确率有1%-5%的差距，因为算法不一样

V1：官方给出的代码，准确率40%
V2-V8：不断增删层，并调超参数，更换损失函数与优化器 准确率增长到60%就遇到瓶颈了
V9：尝试通过仅在5分钟内训练softmax层来预热，然后再释放所有重量。准确率下降到50%
V10：更多数据扩充 准确率55%
V11：使用LR Scheduler 准确率62%
V12：同时使用训练和验证数据来训练模型。 准确率68%
V13；使用谷歌开源新模型 EfficientNetB7 准确率91%，害怕
V14：训练更长的时间（25个轮次）。准确率82%，下降了，是因为过拟合吧
V15：回到20个轮次； Global Max Pooling instead of Average。（全局最大池而不是平均。） 准确率67%，不适合
V16：回滚到global average pooling （全局平均池） 准确率81%
V18：回滚到V13，并调节部分参数 准确率99.9%，恐怖如斯，我好无敌

!pip install -q efficientnet

import math, re, os
import tensorflow as tf
import numpy as np
from matplotlib import pyplot as plt
from kaggle_datasets import KaggleDatasets
import efficientnet.tfkeras as efn
from sklearn.metrics import f1_score, precision_score, recall_score, confusion_matrix

print("Tensorflow version " + tf.__version__)

TPU or GPU detection

# Detect hardware, return appropriate distribution strategy
# try:
#     tpu = tf.distribute.cluster_resolver.TPUClusterResolver()  # TPU detection. No parameters necessary if TPU_NAME environment variable is set. On Kaggle this is always the case.
#     print('Running on TPU ', tpu.master())
# except ValueError:
#     tpu = None

# if tpu:
#     tf.config.experimental_connect_to_cluster(tpu)
#     tf.tpu.experimental.initialize_tpu_system(tpu)
#     strategy = tf.distribute.experimental.TPUStrategy(tpu)
# else:
#     strategy = tf.distribute.get_strategy() # default distribution strategy in Tensorflow. Works on CPU and single GPU.

# print("REPLICAS: ", strategy.num_replicas_in_sync)

Configurations

AUTO = tf.data.experimental.AUTOTUNE

# Create strategy from tpu
tpu = tf.distribute.cluster_resolver.TPUClusterResolver()
tf.config.experimental_connect_to_cluster(tpu)
tf.tpu.experimental.initialize_tpu_system(tpu)
strategy = tf.distribute.experimental.TPUStrategy(tpu)


# Competition data access
# TPUs read data directly from Google Cloud Storage (GCS). 
# This Kaggle utility will copy the dataset to a GCS bucket co-located with the TPU. 
# If you have multiple datasets attached to the notebook, 
# you can pass the name of a specific dataset to the get_gcs_path function. 
# The name of the dataset is the name of the directory it is mounted in. 
# Use !ls /kaggle/input/ to list attached datasets.
GCS_DS_PATH = KaggleDatasets().get_gcs_path()

# Configuration
IMAGE_SIZE = [512, 512]
EPOCHS = 20
BATCH_SIZE = 16 * strategy.num_replicas_in_sync

GCS_PATH_SELECT = { # available image sizes
    192: GCS_DS_PATH + '/tfrecords-jpeg-192x192',
    224: GCS_DS_PATH + '/tfrecords-jpeg-224x224',
    331: GCS_DS_PATH + '/tfrecords-jpeg-331x331',
    512: GCS_DS_PATH + '/tfrecords-jpeg-512x512'
}
GCS_PATH = GCS_PATH_SELECT[IMAGE_SIZE[0]]

TRAINING_FILENAMES = tf.io.gfile.glob(GCS_PATH + '/train/*.tfrec')
VALIDATION_FILENAMES = tf.io.gfile.glob(GCS_PATH + '/val/*.tfrec')
TEST_FILENAMES = tf.io.gfile.glob(GCS_PATH + '/test/*.tfrec') # predictions on this dataset should be submitted for the competition

Unhide to see CLASSES:

CLASSES = ['pink primrose',    'hard-leaved pocket orchid', 'canterbury bells', 'sweet pea',     'wild geranium',     'tiger lily',           'moon orchid',              'bird of paradise', 'monkshood',        'globe thistle',         # 00 - 09
           'snapdragon',       "colt's foot",               'king protea',      'spear thistle', 'yellow iris',       'globe-flower',         'purple coneflower',        'peruvian lily',    'balloon flower',   'giant white arum lily', # 10 - 19
           'fire lily',        'pincushion flower',         'fritillary',       'red ginger',    'grape hyacinth',    'corn poppy',           'prince of wales feathers', 'stemless gentian', 'artichoke',        'sweet william',         # 20 - 29
           'carnation',        'garden phlox',              'love in the mist', 'cosmos',        'alpine sea holly',  'ruby-lipped cattleya', 'cape flower',              'great masterwort', 'siam tulip',       'lenten rose',           # 30 - 39
           'barberton daisy',  'daffodil',                  'sword lily',       'poinsettia',    'bolero deep blue',  'wallflower',           'marigold',                 'buttercup',        'daisy',            'common dandelion',      # 40 - 49
           'petunia',          'wild pansy',                'primula',          'sunflower',     'lilac hibiscus',    'bishop of llandaff',   'gaura',                    'geranium',         'orange dahlia',    'pink-yellow dahlia',    # 50 - 59
           'cautleya spicata', 'japanese anemone',          'black-eyed susan', 'silverbush',    'californian poppy', 'osteospermum',         'spring crocus',            'iris',             'windflower',       'tree poppy',            # 60 - 69
           'gazania',          'azalea',                    'water lily',       'rose',          'thorn apple',       'morning glory',        'passion flower',           'lotus',            'toad lily',        'anthurium',             # 70 - 79
           'frangipani',       'clematis',                  'hibiscus',         'columbine',     'desert-rose',       'tree mallow',          'magnolia',                 'cyclamen ',        'watercress',       'canna lily',            # 80 - 89
           'hippeastrum ',     'bee balm',                  'pink quill',       'foxglove',      'bougainvillea',     'camellia',             'mallow',                   'mexican petunia',  'bromelia',         'blanket flower',        # 90 - 99
           'trumpet creeper',  'blackberry lily',           'common tulip',     'wild rose']                                                                                                                                               # 100 - 102

Helper Functions

1. Visualization Functions

def display_training_curves(training, validation, title, subplot):
    if subplot%10==1: # set up the subplots on the first call
        plt.subplots(figsize=(10,10), facecolor='#F0F0F0')
        plt.tight_layout()
    ax = plt.subplot(subplot)
    ax.set_facecolor('#F8F8F8')
    ax.plot(training)
    ax.plot(validation)
    ax.set_title('model '+ title)
    ax.set_ylabel(title)
    #ax.set_ylim(0.28,1.05)
    ax.set_xlabel('epoch')
    ax.legend(['train', 'valid.'])

# numpy and matplotlib defaults
np.set_printoptions(threshold=15, linewidth=80)

def batch_to_numpy_images_and_labels(data):
    images, labels = data
    numpy_images = images.numpy()
    numpy_labels = labels.numpy()
    if numpy_labels.dtype == object: # binary string in this case, these are image ID strings
        numpy_labels = [None for _ in enumerate(numpy_images)]
    # If no labels, only image IDs, return None for labels (this is the case for test data)
    return numpy_images, numpy_labels

def title_from_label_and_target(label, correct_label):
    if correct_label is None:
        return CLASSES[label], True
    correct = (label == correct_label)
    return "{} [{}{}{}]".format(CLASSES[label], 'OK' if correct else 'NO', u"\u2192" if not correct else '',
                                CLASSES[correct_label] if not correct else ''), correct

def display_one_flower(image, title, subplot, red=False, titlesize=16):
    plt.subplot(*subplot)
    plt.axis('off')
    plt.imshow(image)
    if len(title) > 0:
        plt.title(title, fontsize=int(titlesize) if not red else int(titlesize/1.2), color='red' if red else 'black', fontdict={'verticalalignment':'center'}, pad=int(titlesize/1.5))
    return (subplot[0], subplot[1], subplot[2]+1)
    
def display_batch_of_images(databatch, predictions=None):
    """This will work with:
    display_batch_of_images(images)
    display_batch_of_images(images, predictions)
    display_batch_of_images((images, labels))
    display_batch_of_images((images, labels), predictions)
    """
    # data
    images, labels = batch_to_numpy_images_and_labels(databatch)
    if labels is None:
        labels = [None for _ in enumerate(images)]
        
    # auto-squaring: this will drop data that does not fit into square or square-ish rectangle
    rows = int(math.sqrt(len(images)))
    cols = len(images)//rows
        
    # size and spacing
    FIGSIZE = 13.0
    SPACING = 0.1
    subplot=(rows,cols,1)
    if rows < cols:
        plt.figure(figsize=(FIGSIZE,FIGSIZE/cols*rows))
    else:
        plt.figure(figsize=(FIGSIZE/rows*cols,FIGSIZE))
    
    # display
    for i, (image, label) in enumerate(zip(images[:rows*cols], labels[:rows*cols])):
        title = '' if label is None else CLASSES[label]
        correct = True
        if predictions is not None:
            title, correct = title_from_label_and_target(predictions[i], label)
        dynamic_titlesize = FIGSIZE*SPACING/max(rows,cols)*40+3 # magic formula tested to work from 1x1 to 10x10 images
        subplot = display_one_flower(image, title, subplot, not correct, titlesize=dynamic_titlesize)
    
    #layout
    plt.tight_layout()
    if label is None and predictions is None:
        plt.subplots_adjust(wspace=0, hspace=0)
    else:
        plt.subplots_adjust(wspace=SPACING, hspace=SPACING)
    plt.show()

def display_confusion_matrix(cmat, score, precision, recall):
    plt.figure(figsize=(15,15))
    ax = plt.gca()
    ax.matshow(cmat, cmap='Reds')
    ax.set_xticks(range(len(CLASSES)))
    ax.set_xticklabels(CLASSES, fontdict={'fontsize': 7})
    plt.setp(ax.get_xticklabels(), rotation=45, ha="left", rotation_mode="anchor")
    ax.set_yticks(range(len(CLASSES)))
    ax.set_yticklabels(CLASSES, fontdict={'fontsize': 7})
    plt.setp(ax.get_yticklabels(), rotation=45, ha="right", rotation_mode="anchor")
    titlestring = ""
    if score is not None:
        titlestring += 'f1 = {:.3f} '.format(score)
    if precision is not None:
        titlestring += '\nprecision = {:.3f} '.format(precision)
    if recall is not None:
        titlestring += '\nrecall = {:.3f} '.format(recall)
    if len(titlestring) > 0:
        ax.text(101, 1, titlestring, fontdict={'fontsize': 18, 'horizontalalignment':'right', 'verticalalignment':'top', 'color':'#804040'})
    plt.show()
    
def display_training_curves(training, validation, title, subplot):
    if subplot%10==1: # set up the subplots on the first call
        plt.subplots(figsize=(10,10), facecolor='#F0F0F0')
        plt.tight_layout()
    ax = plt.subplot(subplot)
    ax.set_facecolor('#F8F8F8')
    ax.plot(training)
    ax.plot(validation)
    ax.set_title('model '+ title)
    ax.set_ylabel(title)
    #ax.set_ylim(0.28,1.05)
    ax.set_xlabel('epoch')
    ax.legend(['train', 'valid.'])

2. Datasets Functions

def decode_image(image_data):
    image = tf.image.decode_jpeg(image_data, channels=3)
    image = tf.cast(image, tf.float32) / 255.0  # convert image to floats in [0, 1] range
    image = tf.reshape(image, [*IMAGE_SIZE, 3]) # explicit size needed for TPU
    return image

def read_labeled_tfrecord(example):
    LABELED_TFREC_FORMAT = {
        "image": tf.io.FixedLenFeature([], tf.string), # tf.string means bytestring
        "class": tf.io.FixedLenFeature([], tf.int64),  # shape [] means single element
    }
    example = tf.io.parse_single_example(example, LABELED_TFREC_FORMAT)
    image = decode_image(example['image'])
    label = tf.cast(example['class'], tf.int32)
    return image, label # returns a dataset of (image, label) pairs

def read_unlabeled_tfrecord(example):
    UNLABELED_TFREC_FORMAT = {
        "image": tf.io.FixedLenFeature([], tf.string), # tf.string means bytestring
        "id": tf.io.FixedLenFeature([], tf.string),  # shape [] means single element
        # class is missing, this competitions's challenge is to predict flower classes for the test dataset
    }
    example = tf.io.parse_single_example(example, UNLABELED_TFREC_FORMAT)
    image = decode_image(example['image'])
    idnum = example['id']
    return image, idnum # returns a dataset of image(s)

def load_dataset(filenames, labeled=True, ordered=False):
    # Read from TFRecords. For optimal performance, reading from multiple files at once and
    # disregarding data order. Order does not matter since we will be shuffling the data anyway.

    ignore_order = tf.data.Options()
    if not ordered:
        ignore_order.experimental_deterministic = False # disable order, increase speed

    dataset = tf.data.TFRecordDataset(filenames, num_parallel_reads=AUTO) # automatically interleaves reads from multiple files
    dataset = dataset.with_options(ignore_order) # uses data as soon as it streams in, rather than in its original order
    dataset = dataset.map(read_labeled_tfrecord if labeled else read_unlabeled_tfrecord, num_parallel_calls=AUTO)
    # returns a dataset of (image, label) pairs if labeled=True or (image, id) pairs if labeled=False
    return dataset

def data_augment(image, label, seed=2020):
    # data augmentation. Thanks to the dataset.prefetch(AUTO) statement in the next function (below),
    # this happens essentially for free on TPU. Data pipeline code is executed on the "CPU" part
    # of the TPU while the TPU itself is computing gradients.
    image = tf.image.random_flip_left_right(image, seed=seed)
#     image = tf.image.random_flip_up_down(image, seed=seed)
#     image = tf.image.random_brightness(image, 0.1, seed=seed)
    
#     image = tf.image.random_jpeg_quality(image, 85, 100, seed=seed)
#     image = tf.image.resize(image, [530, 530])
#     image = tf.image.random_crop(image, [512, 512], seed=seed)
    #image = tf.image.random_saturation(image, 0, 2)
    return image, label   

def get_training_dataset():
    dataset = load_dataset(TRAINING_FILENAMES, labeled=True)
    dataset = dataset.map(data_augment, num_parallel_calls=AUTO)
    dataset = dataset.repeat() # the training dataset must repeat for several epochs
    dataset = dataset.shuffle(2048)
    dataset = dataset.batch(BATCH_SIZE)
    dataset = dataset.prefetch(AUTO) # prefetch next batch while training (autotune prefetch buffer size)
    return dataset

def get_validation_dataset(ordered=False):
    dataset = load_dataset(VALIDATION_FILENAMES, labeled=True, ordered=ordered)
    dataset = dataset.batch(BATCH_SIZE)
    dataset = dataset.cache()
    dataset = dataset.prefetch(AUTO) # prefetch next batch while training (autotune prefetch buffer size)
    return dataset

def get_train_valid_datasets():
    dataset = load_dataset(TRAINING_FILENAMES + VALIDATION_FILENAMES, labeled=True)
    dataset = dataset.map(data_augment, num_parallel_calls=AUTO)
    dataset = dataset.repeat() # the training dataset must repeat for several epochs
    dataset = dataset.shuffle(2048)
    dataset = dataset.batch(BATCH_SIZE)
    dataset = dataset.prefetch(AUTO) # prefetch next batch while training (autotune prefetch buffer size)
    return dataset

def get_test_dataset(ordered=False):
    dataset = load_dataset(TEST_FILENAMES, labeled=False, ordered=ordered)
    dataset = dataset.batch(BATCH_SIZE)
    dataset = dataset.prefetch(AUTO) # prefetch next batch while training (autotune prefetch buffer size)
    return dataset

def count_data_items(filenames):
    # the number of data items is written in the name of the .tfrec files, i.e. flowers00-230.tfrec = 230 data items
    n = [int(re.compile(r"-([0-9]*)\.").search(filename).group(1)) for filename in filenames]
    return np.sum(n)

3. Model Functions

def lrfn(epoch):
    LR_START = 0.00001
    LR_MAX = 0.00005 * strategy.num_replicas_in_sync
    LR_MIN = 0.00001
    LR_RAMPUP_EPOCHS = 5
    LR_SUSTAIN_EPOCHS = 0
    LR_EXP_DECAY = .8
    
    if epoch < LR_RAMPUP_EPOCHS:
        lr = (LR_MAX - LR_START) / LR_RAMPUP_EPOCHS * epoch + LR_START
    elif epoch < LR_RAMPUP_EPOCHS + LR_SUSTAIN_EPOCHS:
        lr = LR_MAX
    else:
        lr = (LR_MAX - LR_MIN) * LR_EXP_DECAY**(epoch - LR_RAMPUP_EPOCHS - LR_SUSTAIN_EPOCHS) + LR_MIN
    return lr

def freeze(model):
    for layer in model.layers:
        layer.trainable = False

def unfreeze(model):
    for layer in model.layers:
        layer.trainable = True

Dataset visualizations

# data dump
print("Training data shapes:")
for image, label in get_training_dataset().take(3):
    print(image.numpy().shape, label.numpy().shape)
print("Training data label examples:", label.numpy())
print("Validation data shapes:")
for image, label in get_validation_dataset().take(3):
    print(image.numpy().shape, label.numpy().shape)
print("Validation data label examples:", label.numpy())
print("Test data shapes:")
for image, idnum in get_test_dataset().take(3):
    print(image.numpy().shape, idnum.numpy().shape)
print("Test data IDs:", idnum.numpy().astype('U')) # U=unicode string

# Peek at training data
training_dataset = get_training_dataset()
training_dataset = training_dataset.unbatch().batch(20)
train_batch = iter(training_dataset)

# run this cell again for next set of images
display_batch_of_images(next(train_batch))

# peer at test data
test_dataset = get_test_dataset()
test_dataset = test_dataset.unbatch().batch(20)
test_batch = iter(test_dataset)

# run this cell again for next set of images
display_batch_of_images(next(test_batch))

Training Model

NUM_TRAINING_IMAGES = count_data_items(TRAINING_FILENAMES)
NUM_VALIDATION_IMAGES = count_data_items(VALIDATION_FILENAMES)
NUM_TEST_IMAGES = count_data_items(TEST_FILENAMES)
STEPS_PER_EPOCH = NUM_TRAINING_IMAGES // BATCH_SIZE
print('Dataset: {} training images, {} validation images, {} unlabeled test images'.format(NUM_TRAINING_IMAGES, NUM_VALIDATION_IMAGES, NUM_TEST_IMAGES))

1. Load Model into TPU

# Need this line so Google will recite some incantations
# for Turing to magically load the model onto the TPU
with strategy.scope():
    enet = efn.EfficientNetB7(
        input_shape=(512, 512, 3),
        weights='imagenet',
        include_top=False
    )
    
    model = tf.keras.Sequential([
        enet,
        tf.keras.layers.GlobalAveragePooling2D(),
        tf.keras.layers.Dense(len(CLASSES), activation='softmax')
    ])
        
    model.compile(
        optimizer=tf.keras.optimizers.Adam(),
        loss = 'sparse_categorical_crossentropy',
        metrics=['sparse_categorical_accuracy']
    )
    model.summary()

保存全模型

可以对整个模型进行保存，其保存的内容包括：

1. 该模型的架构
2. 模型的权重（在训练期间学到的）
3. 模型的训练配置（你传递给编译的），如果有的话
4. 优化器及其状态（如果有的话）（这使您可以从中断的地方重新启动训练

model.save('the_save_model.h5') #保存全模型

2. Training

# scheduler = tf.keras.callbacks.ReduceLROnPlateau(patience=3, verbose=1)
lr_schedule = tf.keras.callbacks.LearningRateScheduler(lrfn, verbose=1)

history = model.fit(
    get_train_valid_datasets(), 
    steps_per_epoch=STEPS_PER_EPOCH,
    epochs=EPOCHS, 
    callbacks=[lr_schedule],
    validation_data=get_validation_dataset()
)

history.history.keys()

display_training_curves(history.history['loss'], history.history['val_loss'], 'loss', 211)
display_training_curves(history.history['sparse_categorical_accuracy'], history.history['val_sparse_categorical_accuracy'], 'accuracy', 212)
# display_training_curves(history.history['loss'], history.history['loss'], 'loss', 211)
# display_training_curves(history.history['sparse_categorical_accuracy'], history.history['sparse_categorical_accuracy'], 'accuracy', 212)

3. Confusion matrix

cmdataset = get_validation_dataset(ordered=True) # since we are splitting the dataset and iterating separately on images and labels, order matters.
images_ds = cmdataset.map(lambda image, label: image)
labels_ds = cmdataset.map(lambda image, label: label).unbatch()
cm_correct_labels = next(iter(labels_ds.batch(NUM_VALIDATION_IMAGES))).numpy() # get everything as one batch
cm_probabilities = model.predict(images_ds)
cm_predictions = np.argmax(cm_probabilities, axis=-1)
print("Correct   labels: ", cm_correct_labels.shape, cm_correct_labels)
print("Predicted labels: ", cm_predictions.shape, cm_predictions)

cmat = confusion_matrix(cm_correct_labels, cm_predictions, labels=range(len(CLASSES)))
score = f1_score(cm_correct_labels, cm_predictions, labels=range(len(CLASSES)), average='macro')
precision = precision_score(cm_correct_labels, cm_predictions, labels=range(len(CLASSES)), average='macro')
recall = recall_score(cm_correct_labels, cm_predictions, labels=range(len(CLASSES)), average='macro')
cmat = (cmat.T / cmat.sum(axis=1)).T # normalized
display_confusion_matrix(cmat, score, precision, recall)
print('f1 score: {:.3f}, precision: {:.3f}, recall: {:.3f}'.format(score, precision, recall))

Predictions

test_ds = get_test_dataset(ordered=True) # since we are splitting the dataset and iterating separately on images and ids, order matters.

print('Computing predictions...')
test_images_ds = test_ds.map(lambda image, idnum: image)
probabilities = model.predict(test_images_ds)
predictions = np.argmax(probabilities, axis=-1)
print(predictions)

print('Generating submission.csv file...')
test_ids_ds = test_ds.map(lambda image, idnum: idnum).unbatch()
test_ids = next(iter(test_ids_ds.batch(NUM_TEST_IMAGES))).numpy().astype('U') # all in one batch
# np.savetxt('submission.csv', np.rec.fromarrays([test_ids, predictions]), fmt=['%s', '%d'], delimiter=',', header='id,label', comments='')
import pandas as pd
test = pd.DataFrame({"id":test_ids,"label":predictions})

test.to_csv("submission.csv",index = False)

print(test.head)

Visual validation

dataset = get_validation_dataset()
dataset = dataset.unbatch().batch(20)
batch = iter(dataset)

# run this cell again for next set of images
images, labels = next(batch)
probabilities = model.predict(images)
predictions = np.argmax(probabilities, axis=-1)
display_batch_of_images((images, labels), predictions)