【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
目录
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：
- 该模型的架构
- 模型的权重（在训练期间学到的）
- 模型的训练配置（你传递给编译的），如果有的话
- 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)
大家好，我是[爱做梦的子浩](https://blog.csdn.net/weixin_43124279)，我是东北大学大数据实验班大三的小菜鸡，非常向往优秀，羡慕优秀的人，已拿两个暑假offer，欢迎大家找我进行交流😂😂😂
这是我的博客地址：[子浩的博客https://blog.csdn.net/weixin_43124279]
——
版权声明：本文为CSDN博主「爱做梦的子浩」的原创文章，遵循CC 4.0 BY-SA版权协议，转载请附上原文出处链接及本声明。
