"我的NVIDIA开发者之旅” | 征文活动进行中.......

完成我的NVIDIA开发者之旅——Caffe教程（2）[Jetson TK1]Caffe工具环境（Linux）搭建实例-CSDN社区，搭建好Caffe环境后我们就可以开始我们的Caffe实践啦。

不知道大家写的第一个有关深度学习的代码是什么，博主个人是学习吴恩达老师的DeepLearning入门的，也按照课后作业进行了练习，自己第一次动手的就是实现了一个简单的线性回归的实践如下图，到现在依然记忆犹新，哈哈。

接下来我们开始吧，虽然Caffe用于深层网络，但它同样可以表示“浅层”模型，如用于分类的逻辑回归。我们将对合成数据进行简单的逻辑回归，我们将生成这些数据并保存到HDF5中，以向Caffe提供向量。完成该模型后，我们将添加层以提高精度。这就是Caffe的意义：定义一个模型，进行实验，然后部署。

首先我们导入所需要的一些包等资源：

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
 
import os
os.chdir('..')
 
import sys
sys.path.insert(0, './python')
import caffe
 
 
import os
import h5py
import shutil
import tempfile
 
import sklearn
import sklearn.datasets
import sklearn.linear_model
 
import pandas as pd

我们可以通过合成一个包含10000个4向量的数据集，用于具有2个信息特征和2个噪声特征的二元分类：

X, y = sklearn.datasets.make_classification(
    n_samples=10000, n_features=4, n_redundant=0, n_informative=2, 
    n_clusters_per_class=2, hypercube=False, random_state=0
)
print 'data,',X.shape,y.shape # (10000, 4) (10000,) x0,x1,x2,x3, y
 
# Split into train and test
X, Xt, y, yt = sklearn.model_selection.train_test_split(X, y)
print 'train,',X.shape,y.shape  #train: (7500, 4) (7500,)
print 'test,', Xt.shape,yt.shape#test:  (2500, 4) (2500,)
 
# Visualize sample of the data
ind = np.random.permutation(X.shape[0])[:1000] # (7500,)--->(1000,)   x0,x1,x2,x3, y
df = pd.DataFrame(X[ind])
_ = pd.plotting.scatter_matrix(df, figsize=(9, 9), diagonal='kde', marker='o', s=40, alpha=.4, c=y[ind])

data, (10000, 4) (10000,)

train, (7500, 4) (7500,)

test, (2500, 4) (2500,)

使用随机梯度下降（SGD）训练学习和评估scikit Learn的logistic回归。计时并检查分类器的准确性：

%%timeit
# Train and test the scikit-learn SGD logistic regression.
clf = sklearn.linear_model.SGDClassifier(
    loss='log', n_iter=1000, penalty='l2', alpha=5e-4, class_weight='balanced')
 
clf.fit(X, y)
yt_pred = clf.predict(Xt)
print('Accuracy: {:.3f}'.format(sklearn.metrics.accuracy_score(yt, yt_pred)))

Accuracy: 0.781

Accuracy: 0.781

Accuracy: 0.781

Accuracy: 0.781

1 loop, best of 3: 372 ms per loop

然后再将数据集保存到HDF5以加载到Caffe中：

# Write out the data to HDF5 files in a temp directory.
# This file is assumed to be caffe_root/examples/hdf5_classification.ipynb
dirname = os.path.abspath('./examples/hdf5_classification/data')
if not os.path.exists(dirname):
    os.makedirs(dirname)
 
train_filename = os.path.join(dirname, 'train.h5')
test_filename = os.path.join(dirname, 'test.h5')
 
# HDF5DataLayer source should be a file containing a list of HDF5 filenames.
# To show this off, we'll list the same data file twice.
with h5py.File(train_filename, 'w') as f:
    f['data'] = X
    f['label'] = y.astype(np.float32)
with open(os.path.join(dirname, 'train.txt'), 'w') as f:
    f.write(train_filename + '\n')
    f.write(train_filename + '\n')
    
# HDF5 is pretty efficient, but can be further compressed.
comp_kwargs = {'compression': 'gzip', 'compression_opts': 1}
with h5py.File(test_filename, 'w') as f:
    f.create_dataset('data', data=Xt, **comp_kwargs)
    f.create_dataset('label', data=yt.astype(np.float32), **comp_kwargs)
with open(os.path.join(dirname, 'test.txt'), 'w') as f:
    f.write(test_filename + '\n')

我们可以通过Python net规范在Caffe中定义逻辑回归。这是一种快速而自然的定义网络的方法，避免了手动编辑protobuf模型：

from caffe import layers as L
from caffe import params as P
 
def logreg(hdf5, batch_size):
    # logistic regression: data, matrix multiplication, and 2-class softmax loss
    n = caffe.NetSpec()
    n.data, n.label = L.HDF5Data(batch_size=batch_size, source=hdf5, ntop=2)
    n.ip1 = L.InnerProduct(n.data, num_output=2, weight_filler=dict(type='xavier'))
    n.accuracy = L.Accuracy(n.ip1, n.label)
    n.loss = L.SoftmaxWithLoss(n.ip1, n.label)
    return n.to_proto()
 
train_net_path = 'examples/hdf5_classification/logreg_auto_train.prototxt'
with open(train_net_path, 'w') as f:
    f.write(str(logreg('examples/hdf5_classification/data/train.txt', 10)))
 
test_net_path = 'examples/hdf5_classification/logreg_auto_test.prototxt'
with open(test_net_path, 'w') as f:
    f.write(str(logreg('examples/hdf5_classification/data/test.txt', 10)))

现在，我们将定义“解算器”，该解算器通过指定上面定义的训练和测试网络的位置，以及用于学习、显示和“快照”的各种参数的设置值来训练网络：

from caffe.proto import caffe_pb2
 
def solver(train_net_path, test_net_path):
    s = caffe_pb2.SolverParameter()
 
    # Specify locations of the train and test networks.
    s.train_net = train_net_path
    s.test_net.append(test_net_path)
 
    s.test_interval = 1000  # Test after every 1000 training iterations.
    s.test_iter.append(250) # Test 250 "batches" each time we test.
 
    s.max_iter = 10000      # # of times to update the net (training iterations)
 
    # Set the initial learning rate for stochastic gradient descent (SGD).
    s.base_lr = 0.01        
 
    # Set `lr_policy` to define how the learning rate changes during training.
    # Here, we 'step' the learning rate by multiplying it by a factor `gamma`
    # every `stepsize` iterations.
    s.lr_policy = 'step'
    s.gamma = 0.1
    s.stepsize = 5000
 
    # Set other optimization parameters. Setting a non-zero `momentum` takes a
    # weighted average of the current gradient and previous gradients to make
    # learning more stable. L2 weight decay regularizes learning, to help prevent
    # the model from overfitting.
    s.momentum = 0.9
    s.weight_decay = 5e-4
 
    # Display the current training loss and accuracy every 1000 iterations.
    s.display = 1000
 
    # Snapshots are files used to store networks we've trained.  Here, we'll
    # snapshot every 10K iterations -- just once at the end of training.
    # For larger networks that take longer to train, you may want to set
    # snapshot < max_iter to save the network and training state to disk during
    # optimization, preventing disaster in case of machine crashes, etc.
    s.snapshot = 10000
    s.snapshot_prefix = 'examples/hdf5_classification/data/train'
 
    # We'll train on the CPU for fair benchmarking against scikit-learn.
    # Changing to GPU should result in much faster training!
    s.solver_mode = caffe_pb2.SolverParameter.CPU
    
    return s
 
solver_path = 'examples/hdf5_classification/logreg_solver.prototxt'
with open(solver_path, 'w') as f:
    f.write(str(solver(train_net_path, test_net_path)))

是时候查看学习和评估Python中的逻辑回归的loss和拟合效果了：

%%timeit
caffe.set_mode_cpu()
solver = caffe.get_solver(solver_path)
solver.solve()
 
accuracy = 0
batch_size = solver.test_nets[0].blobs['data'].num
test_iters = int(len(Xt) / batch_size)
for i in range(test_iters):
    solver.test_nets[0].forward()
    accuracy += solver.test_nets[0].blobs['accuracy'].data
accuracy /= test_iters
 
print("Accuracy: {:.3f}".format(accuracy))

Accuracy: 0.770

Accuracy: 0.770

Accuracy: 0.770

Accuracy: 0.770

1 loop, best of 3: 195 ms per loop

通过命令行界面执行同样的操作，以获得关于模型和求解的详细输出：

!./build/tools/caffe train -solver examples/hdf5_classification/logreg_solver.prototxt

I0224 00:32:03.232779   655 caffe.cpp:178] Use CPU.
I0224 00:32:03.391911   655 solver.cpp:48] Initializing solver from parameters: 
train_net: "examples/hdf5_classification/logreg_auto_train.prototxt"
test_net: "examples/hdf5_classification/logreg_auto_test.prototxt"
......
I0224 00:32:04.087514   655 solver.cpp:406]     Test net output #0: accuracy = 0.77
I0224 00:32:04.087532   655 solver.cpp:406]     Test net output #1: loss = 0.593815 (* 1 = 0.593815 loss)
I0224 00:32:04.087541   655 solver.cpp:323] Optimization Done.
I0224 00:32:04.087548   655 caffe.cpp:222] Optimization Done.

如果查看输出或logreg_auto_train.prototxt，您将看到该模型是简单的逻辑回归。

我们可以通过在接受输入的权重和给出输出的权重之间引入非线性，使其更高级一些——现在我们有了一个两层网络。

该网络在nonlinear_auto_train.proto,txt中给出，这是t解算器中所做的唯一更改。我们现在将使用的新网络的最终精度应高于逻辑回归！

from caffe import layers as L
from caffe import params as P
 
def nonlinear_net(hdf5, batch_size):
    # one small nonlinearity, one leap for model kind
    n = caffe.NetSpec()
    n.data, n.label = L.HDF5Data(batch_size=batch_size, source=hdf5, ntop=2)
    # define a hidden layer of dimension 40
    n.ip1 = L.InnerProduct(n.data, num_output=40, weight_filler=dict(type='xavier'))
    # transform the output through the ReLU (rectified linear) non-linearity
    n.relu1 = L.ReLU(n.ip1, in_place=True)
    # score the (now non-linear) features
    n.ip2 = L.InnerProduct(n.ip1, num_output=2, weight_filler=dict(type='xavier'))
    # same accuracy and loss as before
    n.accuracy = L.Accuracy(n.ip2, n.label)
    n.loss = L.SoftmaxWithLoss(n.ip2, n.label)
    return n.to_proto()
 
train_net_path = 'examples/hdf5_classification/nonlinear_auto_train.prototxt'
with open(train_net_path, 'w') as f:
    f.write(str(nonlinear_net('examples/hdf5_classification/data/train.txt', 10)))
 
test_net_path = 'examples/hdf5_classification/nonlinear_auto_test.prototxt'
with open(test_net_path, 'w') as f:
    f.write(str(nonlinear_net('examples/hdf5_classification/data/test.txt', 10)))
 
solver_path = 'examples/hdf5_classification/nonlinear_logreg_solver.prototxt'
with open(solver_path, 'w') as f:
    f.write(str(solver(train_net_path, test_net_path)))

%%timeit
caffe.set_mode_cpu()
solver = caffe.get_solver(solver_path)
solver.solve()
 
accuracy = 0
batch_size = solver.test_nets[0].blobs['data'].num
test_iters = int(len(Xt) / batch_size)
for i in range(test_iters):
    solver.test_nets[0].forward()
    accuracy += solver.test_nets[0].blobs['accuracy'].data
accuracy /= test_iters
 
print("Accuracy: {:.3f}".format(accuracy))

Accuracy: 0.838

Accuracy: 0.837

Accuracy: 0.838

Accuracy: 0.834

1 loop, best of 3: 277 ms per loop

再次通过命令行界面执行同样的操作，以获得关于模型和求解的详细输出：

!./build/tools/caffe train -solver examples/hdf5_classification/nonlinear_logreg_solver.prototxt

I0224 00:32:05.654265   658 caffe.cpp:178] Use CPU.
I0224 00:32:05.810444   658 solver.cpp:48] Initializing solver from parameters: 
train_net: "examples/hdf5_classification/nonlinear_auto_train.prototxt"
test_net: "examples/hdf5_classification/nonlinear_auto_test.prototxt"
......
I0224 00:32:06.078208   658 solver.cpp:406]     Test net output #0: accuracy = 0.8388
I0224 00:32:06.078225   658 solver.cpp:406]     Test net output #1: loss = 0.382042 (* 1 = 0.382042 loss)
I0224 00:32:06.078234   658 solver.cpp:323] Optimization Done.
I0224 00:32:06.078241   658 caffe.cpp:222] Optimization Done.

# Clean up (comment this out if you want to examine the hdf5_classification/data directory).
shutil.rmtree(dirname)

相关参考：BVLC/caffe: Caffe: a fast open framework for deep learning. (github.com)