如何使用 inception resnet v2 训练我的 model？

Question

I am working on model to train images using tensorflow and inception resnet v2 architecture and can't train this model, I have tried to train it but everytime I get

AttributeError: module 'tensorflow.compat.v1' has no attribute 'fit'

import tensorflow.compat.v1 as tf
import inception_resnet_v2 as incep_v2
import os
import cv2
import numpy as np
import matplotlib.pyplot as plt
from tqdm import tqdm
import selectivesearch
import matplotlib.patches as mpatches
import pandas as pd
import random
tf.disable_eager_execution()
# ------------------------------------------------------------------------------
# Configurations
# ------------------------------------------------------------------------------
IMG_SIZE = 150
TRAIN_DIR = "./dataset/train_images"
TEST_DIR = "./dataset/test_images"
data = pd.read_csv("./dataset/train.csv")
data = data.iloc[0:100, :]
# ------------------------------------------------------------------------------
# Read Train Image
# ------------------------------------------------------------------------------
def create_train_data():
    train_data = []
    for ind in data.index:
        path = os.path.join(TRAIN_DIR, data["image_name"][ind])
        img_data = cv2.imread(path)
        img_data = cv2.resize(img_data, (IMG_SIZE, IMG_SIZE))
        train_data.append([np.array(img_data), data["label"][ind]])
        # fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
        # ax.imshow(img_data)
        # plt.show()
    random.shuffle(train_data)
    np.save('train_data.npy', train_data)
    return train_data


def create_test_data():
    test_data = []
    for img in os.listdir(TEST_DIR):
        path = os.path.join(TEST_DIR, img)
        img_data = cv2.imread(path)
        img_data = cv2.resize(img_data, (IMG_SIZE, IMG_SIZE))
        test_data.append(np.array(img_data))
        break
    random.shuffle(test_data)
    return test_data


train_data = create_train_data()
test_data = create_test_data()

# ------------------------------------------------------------------------------
# Declarations
# ------------------------------------------------------------------------------
def define_model(model, is_training):
    model.Image = tf.placeholder(tf.float32, shape=[None, IMG_SIZE, IMG_SIZE, 3])
    with incep_v2.slim.arg_scope(incep_v2.inception_resnet_v2_arg_scope()):
        model.logits, model.end_points = incep_v2.inception_resnet_v2(model.Image, is_training=is_training)


class Model_Class:
    def __init__(self, is_training):
        define_model(self, is_training=is_training)


sess = tf.Session()
# ------------------------------------------------------------------------------
# Create Model
# ------------------------------------------------------------------------------
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
with tf.device('/cpu:0'):
    model = Model_Class(True)

This is the code I use to make a tensorflow model using inception resnet v2 architecture and I don't know how I can train my dataset. Any help?

Answer 1

Actually, with Tensorflow 2, you can use Inception Resnet V2 directly from tensorflow.keras.applications . Below is the demo.

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import tensorflow as tf

import os
import sys
from glob import glob
import cv2 
import time
import datetime

from tensorflow.keras.models import Model, load_model
from tensorflow.keras.layers import Input, Dense, Conv2D, Flatten
from tensorflow.keras.optimizers import SGD, Adam
from tensorflow.python.keras.applications.inception_resnet_v2 import InceptionResNetV2 as  PretrainedModel,preprocess_input

from tensorflow.keras.preprocessing import image
from tensorflow.keras.preprocessing.image import ImageDataGenerator

Now after this, you can prepare your data as (you might have already done this, I guess)

• Folder Name = train -> 1) Class 1 folder 2) Class 2 folder...
• Folder Name = test -> 1) Class 1 folder 2) Class 2 folder...

train_path = r"./train"
test_path = r"./test"

IMAGE_SIZE = [150,150]

image_files = glob(train_path + '/*/*.png')
test_image_files = glob(test_path + '/*/*.png')

folders = glob(train_path + "/*")

Now, let's get the pretrained model. Assuming you want to do transfer learning.

ptm = PretrainedModel(
        input_shape = IMAGE_SIZE + [3],
        weights = 'imagenet',
        include_top = False

)

ptm.trainable = False
K = len(folders)

x = Flatten()(ptm.output)
x = Dense(K, activation = 'softmax')(x)

model = Model(inputs = ptm.input , outputs = x)

Now, let's get the generators which will fetch the data from our folders.

gen = ImageDataGenerator(
        rotation_range = 20,
        width_shift_range = 0.1,
        height_shift_range = 0.1,
        shear_range = 0.1,
        zoom_range = 0.2,
        horizontal_flip = True,
        preprocessing_function = preprocess_input
)

batch_size = 64

train_generator = gen.flow_from_directory(
            
    train_path,
    shuffle = True,
    target_size = IMAGE_SIZE,
    batch_size = batch_size
    )

test_generator = gen.flow_from_directory(
    
    test_path,
    target_size = IMAGE_SIZE,
    batch_size = batch_size
)

Now, let's compile our model.

model.compile(loss = 'categorical_crossentropy' , optimizer = 'adam' , metrics = ['accuracy'])

Finally, let's fit our model to the data

r = model.fit(
        train_generator,
        validation_data = test_generator,
        epochs = 8,
        steps_per_epoch = int(np.ceil(len(image_files)/batch_size)),
        validation_steps = int(np.ceil(len(test_image_files)/batch_size)),
        callbacks=[myCall]
)

Let's get some plots as well

plt.plot(r.history['loss'] , color = 'red' , label = 'loss')
plt.plot(r.history['val_loss'] , color = 'blue' , label = 'val_loss')

Some more..

plt.plot(r.history['accuracy'] , color = 'red' , label = 'loss')
plt.plot(r.history['val_accuracy'] , color = 'blue' , label = 'val_loss')

Let's ave the model for future use..

model.save("model.h5")

We can generate predictions later if we want.

model = load_model("model.h5")
im = cv2.imread(r".\sample.png")
im  = cv2.resize(im , (150,150))
np.argmax(model.predict(im.reshape(1,150,150,3)))