我在python中创建了一个自定义图层，以便我可以直接提供数据
但我注意到它运行得非常慢，GPU使用率最多为1％（内存分配，即我可以看到，当我运行脚本时，它分配2100MB VRAM并终止训练，释放大约1G。
我不确定这是预期的行为还是我做错了什么
这是我写的脚本（基于this former pr）：
import json
import caffe
import numpy as np
from random import shuffle
from PIL import Image
class MyDataLayer(caffe.Layer):
"""
This is a simple datalayer for training a network on CIFAR10.
"""
def setup(self, bottom, top):
self.top_names = ['data', 'label']
# === Read input parameters ===
params = eval(self.param_str)
# Check the paramameters for validity.
check_params(params)
# store input as class variables
self.batch_size = params['batch_size']
# Create a batch loader to load the images.
self.batch_loader = BatchLoader(params, None)
# === reshape tops ===
# since we use a fixed input image size, we can shape the data layer
# once. Else, we'd have to do it in the reshape call.
top[0].reshape(self.batch_size, 3, params['im_height'], params['im_width'])
# this is for our label, since we only have one label we set this to 1
top[1].reshape(self.batch_size, 1)
print_info("MyDataLayer", params)
def forward(self, bottom, top):
"""
Load data.
"""
for itt in range(self.batch_size):
# Use the batch loader to load the next image.
im, label = self.batch_loader.load_next_image()
# Add directly to the caffe data layer
top[0].data[itt, ...] = im
top[1].data[itt, ...] = label
def reshape(self, bottom, top):
"""
There is no need to reshape the data, since the input is of fixed size
(rows and columns)
"""
pass
def backward(self, top, propagate_down, bottom):
"""
These layers does not back propagate
"""
pass
class BatchLoader(object):
"""
This class abstracts away the loading of images.
Images can either be loaded singly, or in a batch. The latter is used for
the asyncronous data layer to preload batches while other processing is
performed.
labels:
the format is like :
png_data_batch_1/leptodactylus_pentadactylus_s_000004.png 6
png_data_batch_1/camion_s_000148.png 9
png_data_batch_1/tipper_truck_s_001250.png 9
"""
def __init__(self, params, result):
self.result = result
self.batch_size = params['batch_size']
self.image_root = params['image_root']
self.im_shape = [params['im_height'],params['im_width']]
# get list of images and their labels.
self.image_labels = params['label']
#getting the list of all image filenames along with their labels
self.imagelist = [line.rstrip('\n\r') for line in open(self.image_labels)]
self._cur = 0 # current image
# this class does some simple data-manipulations
self.transformer = SimpleTransformer()
print ("BatchLoader initialized with {} images".format(len(self.imagelist)))
def load_next_image(self):
"""
Load the next image in a batch.
"""
# Did we finish an epoch?
if self._cur == len(self.imagelist):
self._cur = 0
shuffle(self.imagelist)
# Load an image
image_and_label = self.imagelist[self._cur] # Get the image index
#read the image filename
image_file_name = image_and_label[0:-1]
#load the image
im = np.asarray(Image.open(self.image_root +'/'+image_file_name))
#im = scipy.misc.imresize(im, self.im_shape) # resize
# do a simple horizontal flip as data augmentation
flip = np.random.choice(2)*2-1
im = im[:, ::flip, :]
# Load and prepare ground truth
#read the label
label = image_and_label[-1]
#convert to onehot encoded vector
#fix: caffe automatically converts the label into one hot encoded vector. so we only need to simply use the decimal number (i.e. the plain label number)
#one_hot_label = np.eye(10)[label]
self._cur += 1
return self.transformer.preprocess(im), label
def check_params(params):
"""
A utility function to check the parameters for the data layers.
"""
required = ['batch_size', 'image_root', 'im_width', 'im_height', 'label']
for r in required:
assert r in params.keys(), 'Params must include {}'.format(r)
def print_info(name, params):
"""
Ouput some info regarding the class
"""
print ("{} initialized for split: {}, with bs: {}, im_shape: {}.".format(
name,
params['image_root'],
params['batch_size'],
params['im_height'],
params['im_width'],
params['label']))
class SimpleTransformer:
"""
SimpleTransformer is a simple class for preprocessing and deprocessing
images for caffe.
"""
def __init__(self, mean=[125.30, 123.05, 114.06]):
self.mean = np.array(mean, dtype=np.float32)
self.scale = 1.0
def set_mean(self, mean):
"""
Set the mean to subtract for centering the data.
"""
self.mean = mean
def set_scale(self, scale):
"""
Set the data scaling.
"""
self.scale = scale
def preprocess(self, im):
"""
preprocess() emulate the pre-processing occuring in the vgg16 caffe
prototxt.
"""
im = np.float32(im)
im = im[:, :, ::-1] # change to BGR
im -= self.mean
im *= self.scale
im = im.transpose((2, 0, 1))
return im
def deprocess(self, im):
"""
inverse of preprocess()
"""
im = im.transpose(1, 2, 0)
im /= self.scale
im += self.mean
im = im[:, :, ::-1] # change to RGB
return np.uint8(im)
在我的train_test.prototxt文件中，我有：
name: "CIFAR10_SimpleTest_PythonLayer"
layer {
name: 'MyPythonLayer'
type: 'Python'
top: 'data'
top: 'label'
include {
phase: TRAIN
}
python_param {
#the python script filename
module: 'mypythonlayer'
#the class name
layer: 'MyDataLayer'
#needed parameters in json
param_str: '{"phase":"TRAIN", "batch_size":10, "im_height":32, "im_width":32, "image_root": "G:/Caffe/examples/cifar10/testbed/Train and Test using Pycaffe", "label": "G:/Caffe/examples/cifar10/testbed/Train and Test using Pycaffe/train_cifar10.txt"}'
}
}
layer {
name: 'MyPythonLayer'
type: 'Python'
top: 'data'
top: 'label'
include {
phase: TEST
}
python_param {
#the python script filename
module: 'mypythonlayer'
#the class name
layer: 'MyDataLayer'
#needed parameters in json
param_str: '{"phase":"TEST", "batch_size":10, "im_height":32, "im_width":32, "image_root": "G:/Caffe/examples/cifar10/testbed/Train and Test using Pycaffe", "label": "G:/Caffe/examples/cifar10/testbed/Train and Test using Pycaffe/test_cifar10.txt"}'
}
}
这里有什么不对？
答案 0 :(得分：1)
您的数据层效率不高，需要大部分培训时间（您应该尝试caffe time ...进行更详细的分析）。在每个forward传递中，您正在等待python层一个接一个地从磁盘读取batch_size个图像。这可能需要永远。
在网络处理前几批时，您应该考虑使用Multiprocessing在后台执行读取：这应该可以为您提供良好的CPU / GPU利用率。
有关多处理python数据层的信息，请参阅this example。
答案 1 :(得分：0)
Python层在CPU上而不是GPU上执行，所以它很慢，因为在训练时必须在CPU和GPU之间保持一致。这也是为什么你看到低gpu使用率的原因，因为它等待cpu来执行python层。