(1)Mxnet的数据格式为NDArray，当需要读取可观看的数据，就要调用：
numpy_d = d.asnumpy()
converts it to a Numpy array.
(2)list_arguments （给出当前符号d的输入变量）与list_outputs(给出符号d的输出变量)的说明
import mxnet as mxa = mx.sym.Variable("A") # represent a placeholder. These can be inputs, weights, or anything else.b = mx.sym.Variable("B")c = (a + b) / 10d = c + 1
调用list_arguments 得到的一定就是用于d计算的所有symbol
d.list_arguments()调用list_outputs()得到的就是输出的名字：
# ['A', 'B']
d.list_outputs()上面在查看名称，下面教你如何查看各个层的大小
# ['_plusscalar0_output'] This is the default name from adding to scalar.，
# define input shapes
inp_shapes = {'A':(10,), 'B':(10,)}
arg_shapes, out_shapes, aux_shapes = d.infer_shape(**inp_shapes)
arg_shapes # the shapes of all the inputs to the graph. Order matches d.list_arguments()
# [(10, ), (10, )]
out_shapes # the shapes of all outputs. Order matches d.list_outputs()
# [(10, )]
aux_shapes # the shapes of auxiliary variables. These are variables that are not trainable such as batch normalization population statistics. For now, they are save to ignore.
# []
关于Grad_req的使用，符号描述完后，需要bind，得到一个executor
在使用bing进行绑定，且不需要做反向递归时：
input_arguments = {}args :指出输入的符号以及大小，以词典类型传入
input_arguments['A'] = mx.nd.ones((10, ), ctx=mx.cpu())
input_arguments['B'] = mx.nd.ones((10, ), ctx=mx.cpu())
executor = d.bind(ctx=mx.cpu(),
args=input_arguments, # this can be a list or a dictionary mapping names of inputs to NDArray
grad_req='null') # don't request gradients
grad_req : 设置为Null,说明不需要进行gradient计算
bind完之后，还需要调用一个forward（），就可以运算整个过程。当然，还可以通过executor,对输入的
变量再次进行相关的赋值。
import numpy as np
# The executor
executor.arg_dict
# {'A': NDArray, 'B': NDArray}
executor.arg_dict['A'][:] = np.random.rand(10,) # Note the [:]. This sets the contents of the array instead of setting the array to a new value instead of overwriting the variable.
executor.arg_dict['B'][:] = np.random.rand(10,)
executor.forward()
executor.outputs
# [NDArray]
output_value = executor.outputs[0].asnumpy()
executor.arg_dict['A']是NDArray类型，再使用executor.arg_dict['A'][:]=赋值，表示以numpy的值覆盖NDArray类型的值，类型依旧是NDArray;如果不加[:],表示以numpy值的array类型直接覆盖。但运算的结果却仍然是以mx.nd.ones(10,)得到的.
获取输出的结果：excutor.outputs[0].asnumpy()
本章最重要的一个环节出现了：与上面的例子的区别在于，添加了一个后向传播过程。那么就需要对grad_req = 'write' ,同时调用backforwad.
# allocate space for inputs在调用Bind时，需要提前手动为gradient分配一个空间args_grad并且传入，同时grad_req 设置为 write。
input_arguments = {}
input_arguments['A'] = mx.nd.ones((10, ), ctx=mx.cpu())
input_arguments['B'] = mx.nd.ones((10, ), ctx=mx.cpu())
# allocate space for gradients
grad_arguments = {}
grad_arguments['A'] = mx.nd.ones((10, ), ctx=mx.cpu())
grad_arguments['B'] = mx.nd.ones((10, ), ctx=mx.cpu())
executor = d.bind(ctx=mx.cpu(),
args=input_arguments, # this can be a list or a dictionary mapping names of inputs to NDArray
args_grad=grad_arguments, # this can be a list or a dictionary mapping names of inputs to NDArray
grad_req='write') # instead of null, tell the executor to write gradients. This replaces the contents of grad_arguments with the gradients computed.
executor.arg_dict['A'][:] = np.random.rand(10,)
executor.arg_dict['B'][:] = np.random.rand(10,)
executor.forward()
# in this particular example, the output symbol is not a scalar or loss symbol.
# Thus taking its gradient is not possible.
# What is commonly done instead is to feed in the gradient from a future computation.
# this is essentially how backpropagation works.
out_grad = mx.nd.ones((10,), ctx=mx.cpu())
executor.backward([out_grad]) # because the graph only has one output, only one output grad is needed.
executor.grad_arrays
# [NDarray, NDArray]
再调用executor.forward()前向运行。
再调用excutor.backward()后向运行。输出的symbol既不是一个单量，也不是loss symbol。需要手动传入梯度。
与bind 相对的是 simple_bind，他有一个好处：不需要手动分配计算的梯度空间大小。
input_shapes = {'A': (10,), 'B': (10, )}
executor = d.simple_bind(ctx=mx.cpu(),
grad_req='write', # instead of null, tell the executor to write gradients
**input_shapes)
executor.arg_dict['A'][:] = np.random.rand(10,)
executor.arg_dict['B'][:] = np.random.rand(10,)
executor.forward()
out_grad = mx.nd.ones((10,), ctx=mx.cpu())
executor.backward([out_grad])
只需要为simple_bind 设定 输入的大小，它会自动推断梯度所需的空间大小。
一套清晰简单的网络流程就为你摆放在面前了：
import mxnet as mx
import numpy as np
# First, the symbol needs to be defined
data = mx.sym.Variable("data") # input features, mxnet commonly calls this 'data'
label = mx.sym.Variable("softmax_label")
# One can either manually specify all the inputs to ops (data, weight and bias)
w1 = mx.sym.Variable("weight1")
b1 = mx.sym.Variable("bias1")
l1 = mx.sym.FullyConnected(data=data, num_hidden=128, name="layer1", weight=w1, bias=b1)
a1 = mx.sym.Activation(data=l1, act_type="relu", name="act1")
# Or let MXNet automatically create the needed arguments to ops
l2 = mx.sym.FullyConnected(data=a1, num_hidden=10, name="layer2")
# Create some loss symbol
cost_classification = mx.sym.SoftmaxOutput(data=l2, label=label)
# Bind an executor of a given batch size to do forward pass and get gradients
batch_size = 128
input_shapes = {"data": (batch_size, 28*28), "softmax_label": (batch_size, )}
executor = cost_classification.simple_bind(ctx=mx.gpu(0),
grad_req='write',
**input_shapes)
此时executor是训练时用
# The above executor computes gradients. When evaluating test data we don't need this.
# We want this executor to share weights with the above one, so we will use bind
# (instead of simple_bind) and use the other executor's arguments.
executor_test = cost_classification.bind(ctx=mx.gpu(0),
grad_req='null',
args=executor.arg_arrays)executor_test 是测试时用
# executor 里含有arg_dict表示每层的名称
:bias1,data,layer2_bias,layer2_weight...
#executor 里含有 arg_arrays对应每层的具体数(诀窍:带arrays的表示数值)
# initialize the weights
for r in executor.arg_arrays:
r[:] = np.random.randn(*r.shape)*0.02
# Using skdata to get mnist data. This is for portability. Can sub in any data loading you like.
from skdata.mnist.views import OfficialVectorClassification
data = OfficialVectorClassification()
trIdx = data.sel_idxs[:]
teIdx = data.val_idxs[:]
for epoch in range(10):
print "Starting epoch", epoch
np.random.shuffle(trIdx)
#每128个样本，作为一个batchsize
for x in range(0, len(trIdx), batch_size):
# extract a batch from mnist
batchX = data.all_vectors[trIdx[x:x+batch_size]]
batchY = data.all_labels[trIdx[x:x+batch_size]]
# our executor was bound to 128 size. Throw out non matching batches.
if batchX.shape[0] != batch_size:
continue
# Store batch in executor 'data'
#通过executor的 arg_dict 给予“名称”，就能获取该层的数值信息，例如设置'data'，也就是赋予
#输入数据信息。一定要加上[:] ,表示overwritting
executor.arg_dict['data'][:] = batchX / 255.
# Store label's in 'softmax_label'
executor.arg_dict['softmax_label'][:] = batchY
executor.forward()
executor.backward()
#进行一次forward以及一次backward之后，需要对权值进行一次更新。
#pname表示
# do weight updates in imperative
for pname, W, G in zip(cost_classification.list_arguments(), executor.arg_arrays, executor.grad_arrays):
# Don't update inputs
# MXNet makes no distinction between weights and data.
if pname in ['data', 'softmax_label']:
continue
# what ever fancy update to modify the parameters
W[:] = W - G * .001
# Evaluation at each epoch
num_correct = 0
num_total = 0
for x in range(0, len(teIdx), batch_size):
batchX = data.all_vectors[teIdx[x:x+batch_size]]
batchY = data.all_labels[teIdx[x:x+batch_size]]
if batchX.shape[0] != batch_size:
continue
# use the test executor as we don't care about gradients
executor_test.arg_dict['data'][:] = batchX / 255.
executor_test.forward()
num_correct += sum(batchY == np.argmax(executor_test.outputs[0].asnumpy(), axis=1))
num_total += len(batchY)
print "Accuracy thus far", num_correct / float(num_total)