python - 从 tf.keras.utils.Sequence 构建的自定义数据生成器不适用于 tensorflow 模型的 fit api

Question

我根据来自 link 的指南实现了一个序列生成器对象.

import tensorflow as tf
from cv2 import imread, resize
from sklearn.utils import shuffle
from cv2 import imread, resize
import numpy as np
from tensorflow.keras import utils
import math
import keras as ks

class reader(tf.keras.utils.Sequence):

    def __init__(self, x, y, batch_size, n_class):
        self.x, self.y = x, y
        self.batch_size = batch_size
        self.n_class = n_class
        
    def __len__(self):
        return math.ceil(len(self.x) / self.batch_size)

    def __getitem__(self, idx):
        print('getitem', idx)
        batch_x = self.x[idx * self.batch_size:(idx + 1) *
        self.batch_size]
        batch_y = self.y[idx * self.batch_size:(idx + 1) *
        self.batch_size]
        
        
        data_x = list()
        for batch in batch_x:
            tmp = list()
            for img_path in batch:
                try:
                    img = imread(img_path)
                    tmp.append(img)
                except Exception as e:
                    print(e)
                    print('failed to find path {}'.format(img_path))
            data_x.append(tmp)
        # 
        data_x = np.array(data_x, dtype='object')
        data_y = np.array(batch_y)
        data_y = utils.to_categorical(data_y, self.n_class)
        print('return item')
        print(data_x.shape)
        return (data_x, data_y)
    
    def on_epoch_end(self):
        # option method to run some logic at the end of each epoch: e.g. reshuffling
        print('on epoch end')
        seed = np.random.randint()
        self.x = shuffle(self.x, random_state=seed)
        self.y = shuffle(self.y, random_state=seed)

但是，它不适用于 tensorflow 模型的 fit api。下面是我用来复制这个问题的简单模型架构。

model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Conv3D(10, input_shape=(TEMPORAL_LENGTH,HEIGHT,WIDTH,CHANNEL), kernel_size=(2,2,2), strides=2))
model.add(tf.keras.layers.Conv3D(10, kernel_size=(2,3,3), strides=2))
model.add(tf.keras.layers.Conv3D(10, kernel_size=(2,3,3), strides=2))
model.add(tf.keras.layers.Conv3D(10, kernel_size=(2,3,3), strides=2))
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(10))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy', tf.keras.metrics.Precision(), tf.keras.metrics.Recall()])
model.summary()

让我创建一个阅读器 r1 = reader(x_train, y_train, 20, 10)然后我调用model.fit api。

train_history = model.fit(r1, epochs=3, steps_per_epoch=5, verbose=1)
### output ###
getitem 0
return item
(20, 16, 192, 256, 3)
WARNING:tensorflow:sample_weight modes were coerced from
  ...
    to  
  ['...']
Train for 5 steps
Epoch 1/3

如果我不打扰，它会一直这样。出于好奇，我用 Keras api 创建的模型尝试了这种方法，令我惊讶的是它确实有效!

model = ks.models.Sequential()
model.add(ks.layers.Conv3D(10, input_shape=(TEMPORAL_LENGTH,HEIGHT,WIDTH,CHANNEL), kernel_size=(2,2,2), strides=2))
model.add(ks.layers.Conv3D(10, kernel_size=(2,3,3), strides=2))
model.add(ks.layers.Conv3D(10, kernel_size=(2,3,3), strides=2))
model.add(ks.layers.Conv3D(10, kernel_size=(2,3,3), strides=2))
model.add(ks.layers.Flatten())
model.add(ks.layers.Dense(10))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()
train_history = model.fit(r1, epochs=3, steps_per_epoch=5, verbose=1)
### output ###
Epoch 1/3
getitem 586
return item
(20, 16, 192, 256, 3)
getitem 169
1/5 [=====>........................] - ETA: 22s - loss: 11.0373 - accuracy: 0.0000e+00return item
(20, 16, 192, 256, 3)
getitem 601
2/5 [===========>..................] - ETA: 12s - loss: 7.9983 - accuracy: 0.0250     return item
(20, 16, 192, 256, 3)
getitem 426
3/5 [=================>............] - ETA: 8s - loss: 10.7049 - accuracy: 0.2500return item
(20, 16, 192, 256, 3)
getitem 243
4/5 [=======================>......] - ETA: 3s - loss: 8.5093 - accuracy: 0.1875

依赖

tensorflow -GPU:2.1

keras-gpu:2.3.1

Answer 1

老年人。对于迟到的回复，我深表歉意。我已经找到了解决此问题的方法。
我需要更改的只是在函数 self 处将 data_x 转换为 dtype='float32'。 getitem ().要复制该问题，只需将 dtype 更改为“对象”。
除此之外，请允许我分享该类 ActionDataGenerator 是从 Anujshah's 修改而来的。教程。

import tensorflow as tf
from sklearn.utils import shuffle
import cv2
from cv2 import imread, resize
from tensorflow.keras import utils
import math
import keras as ks
import pandas as pd
import numpy as np
import os
from collections import deque
import copy

class reader(tf.keras.utils.Sequence):

    def __init__(self, x, y, batch_size, n_class):
        self.x, self.y = x, y
        self.batch_size = batch_size
        self.n_class = n_class
        
    def __len__(self):
        return math.ceil(len(self.x) / self.batch_size)

    def __getitem__(self, idx):
        batch_x = self.x[idx * self.batch_size:(idx + 1) *
        self.batch_size]
        batch_y = self.y[idx * self.batch_size:(idx + 1) *
        self.batch_size]
        
        
        data_x = list()
        for batch in batch_x:
            tmp = list()
            for img_path in batch:
                try:
                    img = imread(img_path)
                    if img.shape != (192, 256, 3):
                        img = cv2.resize(img,(256, 192))
                    tmp.append(img)
                except Exception as e:
                    print(e)
                    print('failed to find path {}'.format(img_path))
            data_x.append(tmp)
        # 
        data_x = np.array(data_x, dtype='float32')
        data_y = np.array(batch_y)
        data_y = utils.to_categorical(data_y, self.n_class)
        return data_x, data_y
    
    def on_epoch_end(self):
        # option method to run some logic at the end of each epoch: e.g. reshuffling
        seed = np.random.randint()
        self.x = shuffle(self.x, random_state=seed)
        self.y = shuffle(self.y, random_state=seed)

class ActionDataGenerator(object):
    
    def __init__(self,root_data_path,temporal_stride=1,temporal_length=16,resize=224, max_sample=20):
        
        self.root_data_path = root_data_path
        self.temporal_length = temporal_length
        self.temporal_stride = temporal_stride
        self.resize=resize
        self.max_sample=max_sample

    def file_generator(self,data_path,data_files):
        '''
        data_files - list of csv files to be read.
        '''
        for f in data_files:       
            tmp_df = pd.read_csv(os.path.join(data_path,f))
            label_list = list(tmp_df['Label'])
            total_images = len(label_list) 
            if total_images>=self.temporal_length:
                num_samples = int((total_images-self.temporal_length)/self.temporal_stride)+1
                
                img_list = list(tmp_df['FileName'])
            else:
                print ('num of frames is less than temporal length; hence discarding this file-{}'.format(f))
                continue
            
            samples = deque()
            samp_count=0
            for img in img_list:
                if samp_count == self.max_sample:
                    break
                samples.append(img)
                if len(samples)==self.temporal_length:
                    samples_c=copy.deepcopy(samples)
                    samp_count+=1
                    for t in range(self.temporal_stride):
                        samples.popleft()
                    yield samples_c,label_list[0]

    def load_samples(self,data_cat='train', test_ratio=0.1):
        data_path = os.path.join(self.root_data_path,data_cat)
        csv_data_files = os.listdir(data_path)
        file_gen = self.file_generator(data_path,csv_data_files)
        iterator = True
        data_list = []
        while iterator:
            try:
                x,y = next(file_gen)
                x=list(x)
                data_list.append([x,y])
            except Exception as e:
                print ('the exception: ',e)
                iterator = False
                print ('end of data generator')
        # data_list = self.shuffle_data(data_list)
        return data_list
    
    def train_validation_split(self, data_list, target_column, val_size=0.1, ks_sequence=False):
        dataframe = pd.DataFrame(data_list)
        dataframe.columns = ['Feature', target_column]
        data_dict = dict()
        for i in range(len(np.unique(dataframe[target_column]))):
            data_dict[i] = dataframe[dataframe[target_column]==i]
        train, validation = pd.DataFrame(), pd.DataFrame()
        for df in data_dict.values():
            cut = int(df.shape[0] * val_size)
            val = df[:cut]
            rem = df[cut:]
            train = train.append(rem, ignore_index=True)
            validation = validation.append(val, ignore_index=True)
        if ks_sequence:
            return train['Feature'].values.tolist(), train['Label'].values.tolist(), \
                validation['Feature'].values.tolist(), validation['Label'].values.tolist() # without shuffle
        return train.values.tolist(), validation.values.tolist() # without shuffle

root_data_path = 'C:\\Users\\AI-lab\\Documents\\activity_file\\UCF101\\csv_files\\' # machine specific
CLASSES = 101
BATCH_SIZE = 10
EPOCHS = 1
TEMPORAL_STRIDE = 8
TEMPORAL_LENGTH = 16
MAX_SAMPLE = 20
HEIGHT = 192
WIDTH = 256
CHANNEL = 3

data_gen_obj = ActionDataGenerator(root_data_path, temporal_stride=TEMPORAL_STRIDE, \
                                  temporal_length=TEMPORAL_LENGTH, max_sample=MAX_SAMPLE)
train_data = data_gen_obj.load_samples(data_cat='train')
x_train, y_train, x_val, y_val = data_gen_obj.train_validation_split(train_data, 'Label', 0.1, True)
r1 = reader(x_train, y_train, BATCH_SIZE, CLASSES)
r2 = reader(x_val, y_val, BATCH_SIZE, CLASSES)
print(type(r1), type(r2))

model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Conv3D(10, input_shape=(TEMPORAL_LENGTH,HEIGHT,WIDTH,CHANNEL), kernel_size=(2,2,2), strides=2))
model.add(tf.keras.layers.Conv3D(10, kernel_size=(2,3,3), strides=2))
model.add(tf.keras.layers.Conv3D(10, kernel_size=(2,3,3), strides=2))
model.add(tf.keras.layers.Conv3D(10, kernel_size=(2,3,3), strides=2))
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(101, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()

train_history = model.fit(r1, epochs=3, steps_per_epoch=r1.__len__(), verbose=1)
score = model.evaluate(r2, steps=5)
print(score)

输出

the exception:  
end of data generator
<class '__main__.reader'> <class '__main__.reader'>
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv3d (Conv3D)              (None, 8, 96, 128, 10)    250       
_________________________________________________________________
conv3d_1 (Conv3D)            (None, 4, 47, 63, 10)     1810      
_________________________________________________________________
conv3d_2 (Conv3D)            (None, 2, 23, 31, 10)     1810      
_________________________________________________________________
conv3d_3 (Conv3D)            (None, 1, 11, 15, 10)     1810      
_________________________________________________________________
flatten (Flatten)            (None, 1650)              0         
_________________________________________________________________
dense (Dense)                (None, 101)               166751    
=================================================================
Total params: 172,431
Trainable params: 172,431
Non-trainable params: 0
_________________________________________________________________
WARNING:tensorflow:sample_weight modes were coerced from
  ...
    to  
  ['...']
Train for 17562 steps
Epoch 1/3
   77/17562 [..............................] - ETA: 1:35:53 - loss: 67.0937 - accuracy: 0.0156