TensorBoardのProjectorに学習結果を表示させたい

実現したいこと

https://qiita.com/notfolder/items/31bb8be368239914a622
こちらの記事を参考に自分の学習の結果を表示させたいと思い、見よう見まねでやってみたもののメモリが足りなくなり動きません
自分のコードで動かせるようにはどうすればいいか知りたいです

該当のソースコード

Projector用のコードは記述してないものを記述します

python
1import tensorflow as tf
2from tensorflow.keras import backend, models, layers, regularizers , optimizers
3from tensorflow.keras.models import Model, Sequential
4from tensorflow.keras.layers import BatchNormalization , Concatenate
5from keras.layers import Activation, Flatten, Dense, Dropout ,Input
6from tensorflow.keras.layers import Conv2D, MaxPooling2D, AveragePooling2D
7from tensorflow.keras.utils import to_categorical
8from tensorflow.keras.metrics import categorical_crossentropy
9from tensorflow.keras.callbacks import EarlyStopping
10from sklearn.model_selection import train_test_split
11from tensorflow.keras.preprocessing import image
12from tensorflow.keras.preprocessing.image import ImageDataGenerator
13from tensorflow.keras.applications import InceptionV3
14from tensorflow.keras.utils import plot_model
15from tensorflow.keras import backend
16import matplotlib.image as mpimg
17import matplotlib.pyplot as plt
18%matplotlib inline
19import random
20import os,datetime
21import numpy as np
22import pandas as pd
23
24train_path = "../train"
25valid_path = "../valid"
26test_path  = "../test"
27
28train_datagen_augmented = ImageDataGenerator(
29    rescale=1/255,
30    horizontal_flip=True,    # 水平反転
31    rotation_range=60,       # 回転
32    zoom_range=0.1,          # 拡大
33    width_shift_range =0.5,  #左右水平移動
34    height_shift_range =0.5, #上下水平移動
35    shear_range = 30,        #せん断
36    fill_mode = "nearest" ,  #補完
37)
38valid_datagen = ImageDataGenerator(
39    rescale=1/255
40    )
41test_datagen = ImageDataGenerator(
42    rescale=1/255
43    )
44
45train_generator_augmented = train_datagen_augmented.flow_from_directory(
46    train_path,
47    target_size=(224, 224),
48    batch_size=32,
49    color_mode="rgb",
50    class_mode="sparse",
51    shuffle=True,
52)
53validation_generator = valid_datagen.flow_from_directory(
54    valid_path,
55    target_size=(224, 224),
56    batch_size=32,
57    color_mode="rgb",
58    class_mode="sparse")
59
60test_generator = test_datagen.flow_from_directory(
61    test_path,
62    target_size=(224, 224),
63    batch_size=32,
64    color_mode="rgb",
65    class_mode="sparse")
66
67# ここからモデル定義とか
68backend.clear_session()
69
70# pre-trainモデルのload
71incbasemodel4 = InceptionV3(weights="imagenet", include_top=False, input_shape= (224, 224, 3))
72finetune_at = 20
73incbasemodel4.trainable = True
74
75# 最上層あたり以外をフリーズ
76for layer in incbasemodel4.layers[:finetune_at - 1]:
77    layer.trainable =  False
78    # Batch Normalization の freeze解除
79    if layer.name.startswith('batch_normalization'):
80        layer.trainable = True
81
82# Sequential に追加していく
83model_inceptionv3_finetune = models.Sequential()
84model_inceptionv3_finetune.add(incbasemodel4)
85model_inceptionv3_finetune.add(layers.GlobalAveragePooling2D())
86model_inceptionv3_finetune.add(layers.Dense(1024, activation="relu"))
87model_inceptionv3_finetune.add(layers.Dense(20, activation="sigmoid"))
88
89model_inceptionv3_finetune.compile(optimizer=optimizers.Adam(),
90               loss="sparse_categorical_crossentropy",
91               metrics=["accuracy"])
92
93checkpoint_path = "../checkpoints/cp-{epoch:04d}.ckpt"
94checkpoint_dir = os.path.dirname(checkpoint_path)
95
96# checkpoint保存用コールバック
97cp_callback =tf.keras.callbacks.ModelCheckpoint(
98    filepath=checkpoint_path, 
99    verbose=1, 
100    save_weights_only=True,
101    period=1
102)
103
104log_dir = "logs/fit/"+ datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
105tb_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir, histogram_freq=1, write_graph=True)
106
107history = model_inceptionv3_finetune.fit_generator(
108          train_generator_augmented,
109          epochs=100,
110          validation_data=validation_generator,
111          verbose=1, 
112          shuffle=True,
113          callbacks=[
114              EarlyStopping(monitor="val_accuracy", patience=10, restore_best_weights=True),
115              cp_callback,
116              tb_callback,
117          ]
118)
119