Mxnet速查_CPU和GPU的mnist預測訓練_模型匯出_模型匯入再預測_匯出onnx並預測

需要做點什麼

方便廣大菸酒生研究生、人工智障煉丹師演算法工程師快速使用mxnet，所以特寫此文章，預設使用者已有基本的深度學習概念、資料集概念。

系統環境

python 3.7.4
mxnet 1.9.0
mxnet-cu112 1.9.0
onnx 1.9.0
onnxruntime-gpu 1.9.0

資料準備

MNIST資料集csv檔案是一個42000x785的矩陣
42000表示有42000張圖片
785中第一列是圖片的類別(0,1,2,..,9),第二列到最後一列是圖片資料向量 (28x28的圖片張成784的向量)， 資料集長這個樣子:

1 0 0 0 0 0 0 0 0 0 ..
0 0 0 0 0 0 0 0 0 0 ..
1 0 0 0 0 0 0 0 0 0 ..

1. 匯入需要的包

import time
import copy
import onnx
import logging
import platform
import mxnet as mx
import numpy as np
import pandas as pd
import onnxruntime as ort
from sklearn.metrics import accuracy_score

logger = logging.getLogger()
logger.setLevel(logging.DEBUG)

# Mxnet Chcek
if platform.system().lower() != 'windows':
    print(mx.runtime.feature_list())
print(mx.context.num_gpus())
a = mx.nd.ones((2, 3), mx.cpu())
b = a * 2 + 1
print(b)

執行輸出

[✔ CUDA, ✔ CUDNN, ✔ NCCL, ✔ CUDA_RTC, ✖ TENSORRT, ✔ CPU_SSE, ✔ CPU_SSE2, ✔ CPU_SSE3, ✖ CPU_SSE4_1, ✖ CPU_SSE4_2, ✖ CPU_SSE4A, ✖ CPU_AVX, ✖ CPU_AVX2, ✔ OPENMP, ✖ SSE, ✖ F16C, ✖ JEMALLOC, ✔ BLAS_OPEN, ✖ BLAS_ATLAS, ✖ BLAS_MKL, ✖ BLAS_APPLE, ✔ LAPACK, ✔ MKLDNN, ✔ OPENCV, ✖ CAFFE, ✖ PROFILER, ✔ DIST_KVSTORE, ✖ CXX14, ✖ INT64_TENSOR_SIZE, ✔ SIGNAL_HANDLER, ✖ DEBUG, ✖ TVM_OP]
1

[[3. 3. 3.]
 [3. 3. 3.]]
<NDArray 2x3 @cpu(0)>

2. 引數準備

N_EPOCH = 1
N_BATCH = 32
N_BATCH_NUM = 900
S_DATA_PATH = r"mnist_train.csv"
S_MODEL_PATH = r"mxnet_cnn"
S_SYM_PATH = './mxnet_cnn-symbol.json'
S_PARAMS_PATH = './mxnet_cnn-0001.params'
S_ONNX_MODEL_PATH = './mxnet_cnn.onnx'
S_DEVICE, N_DEVICE_ID, S_DEVICE_FULL = "cuda", 0, "cuda:0"
# S_DEVICE, N_DEVICE_ID, S_DEVICE_FULL = "cpu", 0, "cpu"
CTX = mx.cpu() if S_DEVICE == "cpu" else mx.gpu(N_DEVICE_ID)
B_IS_UNIX = True

3. 讀取資料

df = pd.read_csv(S_DATA_PATH, header=None)
print(df.shape)
np_mat = np.array(df)
print(np_mat.shape)
X = np_mat[:, 1:]
Y = np_mat[:, 0]
X = X.astype(np.float32) / 255
X_train = X[:N_BATCH * N_BATCH_NUM]
X_test = X[N_BATCH * N_BATCH_NUM:]
Y_train = Y[:N_BATCH * N_BATCH_NUM]
Y_test = Y[N_BATCH * N_BATCH_NUM:]
X_train = X_train.reshape(X_train.shape[0], 1, 28, 28)
X_test = X_test.reshape(X_test.shape[0], 1, 28, 28) 
print(X_train.shape)
print(Y_train.shape)
print(X_test.shape)
print(Y_test.shape)
train_iter = mx.io.NDArrayIter(X_train, Y_train, batch_size=N_BATCH)
test_iter = mx.io.NDArrayIter(X_test, Y_test, batch_size=N_BATCH)
test_iter_2 = copy.copy(test_iter)

執行輸出

(37800, 785)
(37800, 785)
(28800, 1, 28, 28)
(28800,)
(9000, 1, 28, 28)
(9000,)

4. 模型構建

net = mx.gluon.nn.HybridSequential()
with net.name_scope():
    net.add(mx.gluon.nn.Conv2D(channels=32, kernel_size=3, activation='relu'))  # bx28x28 ==>
    net.add(mx.gluon.nn.MaxPool2D(pool_size=2, strides=2))
    net.add(mx.gluon.nn.Flatten())
    net.add(mx.gluon.nn.Dense(128, activation="relu"))
    net.add(mx.gluon.nn.Dense(10))
net.hybridize()
print(net)
net.collect_params().initialize(mx.init.Xavier(), ctx=CTX)
softmax_cross_entropy = mx.gluon.loss.SoftmaxCrossEntropyLoss()
trainer = mx.gluon.Trainer(net.collect_params(), 'adam', {'learning_rate': .001})

執行輸出

HybridSequential(
  (0): Conv2D(None -> 32, kernel_size=(3, 3), stride=(1, 1), Activation(relu))
  (1): MaxPool2D(size=(2, 2), stride=(2, 2), padding=(0, 0), ceil_mode=False, global_pool=False, pool_type=max, layout=NCHW)
  (2): Flatten
  (3): Dense(None -> 128, Activation(relu))
  (4): Dense(None -> 10, linear)
)

5. 模型訓練

for epoch in range(N_EPOCH):
    for batch_num, itr in enumerate(train_iter):
        data = itr.data[0].as_in_context(CTX)
        label = itr.label[0].as_in_context(CTX)
        with mx.autograd.record():
            output = net(data)  # Run the forward pass
            loss = softmax_cross_entropy(output, label)  # Compute the loss
        loss.backward()
        trainer.step(data.shape[0])
        if batch_num % 50 == 0:  # Print loss once in a while
            curr_loss = mx.nd.mean(loss)  # .asscalar()
            pred = mx.nd.argmax(output, axis=1)
            np_pred, np_lable = pred.asnumpy(), label.asnumpy()
            f_acc = accuracy_score(np_lable, np_pred)
            print(f"Epoch: {epoch}; Batch {batch_num}; ACC {f_acc}")
            print(f"loss: {curr_loss}")
            print()
            # print("Epoch: %d; Batch %d; Loss %s; ACC %f" %
            #       (epoch, batch_num, str(curr_loss), f_acc))
    print()

執行輸出

Epoch: 0; Batch 0; ACC 0.09375
loss: 
[2.2868602]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 50; ACC 0.875
loss: 
[0.512461]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 100; ACC 0.90625
loss: 
[0.43415746]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 150; ACC 0.84375
loss: 
[0.3854709]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 200; ACC 1.0
loss: 
[0.04192135]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 250; ACC 0.90625
loss: 
[0.21156572]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 300; ACC 0.9375
loss: 
[0.15938525]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 350; ACC 1.0
loss: 
[0.0379494]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 400; ACC 0.96875
loss: 
[0.17104594]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 450; ACC 0.96875
loss: 
[0.12192786]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 500; ACC 0.96875
loss: 
[0.09210478]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 550; ACC 0.9375
loss: 
[0.13728428]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 600; ACC 0.96875
loss: 
[0.0762211]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 650; ACC 0.96875
loss: 
[0.12162409]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 700; ACC 1.0
loss: 
[0.04334489]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 750; ACC 1.0
loss: 
[0.06458903]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 800; ACC 0.96875
loss: 
[0.07410634]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 850; ACC 0.96875
loss: 
[0.14233188]
<NDArray 1 @gpu(0)>

6.模型預測

for batch_num, itr in enumerate(test_iter_2):
    data = itr.data[0].as_in_context(CTX)
    label = itr.label[0].as_in_context(CTX)

    output = net(data)  # Run the forward pass
    loss = softmax_cross_entropy(output, label)  # Compute the loss

    if batch_num % 50 == 0:  # Print loss once in a while
        curr_loss = mx.nd.mean(loss)  # .asscalar()
        pred = mx.nd.argmax(output, axis=1)
        np_pred, np_lable = pred.asnumpy(), label.asnumpy()
        f_acc = accuracy_score(np_lable, np_pred)
        print(f"Epoch: {epoch}; Batch {batch_num}; ACC {f_acc}")
        print(f"loss: {curr_loss}")
        print()

執行輸出

Epoch: 0; Batch 0; ACC 0.96875
loss: 
[0.22968824]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 50; ACC 0.96875
loss: 
[0.05668993]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 100; ACC 0.96875
loss: 
[0.08171713]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 150; ACC 1.0
loss: 
[0.02264522]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 200; ACC 0.96875
loss: 
[0.080383]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 250; ACC 1.0
loss: 
[0.03774196]
<NDArray 1 @gpu(0)>

7.模型儲存

net.export(S_MODEL_PATH, epoch=N_EPOCH)  # 儲存模型結構和全部引數

8.模型載入和載入模型使用

print("load net and do test")
load_net = mx.gluon.nn.SymbolBlock.imports(S_SYM_PATH, ['data'], S_PARAMS_PATH, ctx=CTX)  # 載入模型
print("load ok")
for batch_num, itr in enumerate(test_iter):  # Test
    data = itr.data[0].as_in_context(CTX)
    label = itr.label[0].as_in_context(CTX)

    output = load_net(data)  # Run the forward pass
    loss = softmax_cross_entropy(output, label)  # Compute the loss

    if batch_num % 50 == 0:  # Print loss once in a while
        curr_loss = mx.nd.mean(loss)  # .asscalar()
        pred = mx.nd.argmax(output, axis=1)
        np_pred, np_lable = pred.asnumpy(), label.asnumpy()
        f_acc = accuracy_score(np_lable, np_pred)
        print(f"Epoch: {epoch}; Batch {batch_num}; ACC {f_acc}")
        print(f"loss: {curr_loss}")
        print()
print("finish")

執行輸出

load net and do test
load ok
Epoch: 0; Batch 0; ACC 0.96875
loss: 
[0.22968824]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 50; ACC 0.96875
loss: 
[0.05668993]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 100; ACC 0.96875
loss: 
[0.08171713]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 150; ACC 1.0
loss: 
[0.02264522]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 200; ACC 0.96875
loss: 
[0.080383]
<NDArray 1 @gpu(0)>

Epoch: 0; Batch 250; ACC 1.0
loss: 
[0.03774196]
<NDArray 1 @gpu(0)>

finish

9.匯出ONNX

if platform.system().lower() != 'windows':
    mx.onnx.export_model(S_SYM_PATH, S_PARAMS_PATH, [(32, 1, 28, 28)], [np.float32], S_ONNX_MODEL_PATH, verbose=True, dynamic=True)

執行輸出

INFO:root:Converting json and weight file to sym and params
INFO:root:Converting idx: 0, op: null, name: data
INFO:root:Converting idx: 1, op: null, name: hybridsequential0_conv0_weight
INFO:root:Converting idx: 2, op: null, name: hybridsequential0_conv0_bias
INFO:root:Converting idx: 3, op: Convolution, name: hybridsequential0_conv0_fwd
INFO:root:Converting idx: 4, op: Activation, name: hybridsequential0_conv0_relu_fwd
INFO:root:Converting idx: 5, op: Pooling, name: hybridsequential0_pool0_fwd
INFO:root:Converting idx: 6, op: Flatten, name: hybridsequential0_flatten0_flatten0
INFO:root:Converting idx: 7, op: null, name: hybridsequential0_dense0_weight
INFO:root:Converting idx: 8, op: null, name: hybridsequential0_dense0_bias
INFO:root:Converting idx: 9, op: FullyConnected, name: hybridsequential0_dense0_fwd
INFO:root:Converting idx: 10, op: Activation, name: hybridsequential0_dense0_relu_fwd
INFO:root:Converting idx: 11, op: null, name: hybridsequential0_dense1_weight
INFO:root:Converting idx: 12, op: null, name: hybridsequential0_dense1_bias
INFO:root:Converting idx: 13, op: FullyConnected, name: hybridsequential0_dense1_fwd
INFO:root:Output node is: hybridsequential0_dense1_fwd
INFO:root:Input shape of the model [(32, 1, 28, 28)] 
INFO:root:Exported ONNX file ./mxnet_cnn.onnx saved to disk

10. 載入ONNX並執行

if platform.system().lower() != 'windows':
    model = onnx.load(S_ONNX_MODEL_PATH)
    print(onnx.checker.check_model(model))  # Check that the model is well formed
    # print(onnx.helper.printable_graph(model.graph))  # Print a human readable representation of the graph
    ls_input_name, ls_output_name = [input.name for input in model.graph.input], [output.name for output in model.graph.output]
    print("input name ", ls_input_name)
    print("output name ", ls_output_name)
    s_input_name = ls_input_name[0]

    x_input = X_train[:N_BATCH*2, :, :, :].astype(np.float32)
    ort_val = ort.OrtValue.ortvalue_from_numpy(x_input, S_DEVICE, N_DEVICE_ID)
    print("val device ", ort_val.device_name())
    print("val shape ", ort_val.shape())
    print("val data type ", ort_val.data_type())
    print("is_tensor ", ort_val.is_tensor())
    print("array_equal ", np.array_equal(ort_val.numpy(), x_input))
    providers = 'CUDAExecutionProvider' if S_DEVICE == "cuda" else 'CPUExecutionProvider'
    print("providers ", providers)
    ort_session = ort.InferenceSession(S_ONNX_MODEL_PATH, providers=[providers])  # gpu執行
    ort_session.set_providers([providers])
    outputs = ort_session.run(None, {s_input_name: ort_val})
    print("sess env ", ort_session.get_providers())
    print(type(outputs))
    print(outputs[0])

    '''
    For example ['CUDAExecutionProvider', 'CPUExecutionProvider']
        means execute a node using CUDAExecutionProvider if capable, otherwise execute using CPUExecutionProvider.
    '''

執行輸出

None
input name  ['data', 'hybridsequential0_conv0_weight', 'hybridsequential0_conv0_bias', 'hybridsequential0_dense0_weight', 'hybridsequential0_dense0_bias', 'hybridsequential0_dense1_weight', 'hybridsequential0_dense1_bias']
output name  ['hybridsequential0_dense1_fwd']
val device  cuda
val shape  [64, 1, 28, 28]
val data type  tensor(float)
is_tensor  True
array_equal  True
providers  CUDAExecutionProvider
sess env  ['CUDAExecutionProvider', 'CPUExecutionProvider']
<class 'list'>
[[-2.69336128e+00  8.42524242e+00 -3.34120363e-01 -1.17912292e+00
   3.82278800e-01 -3.60794234e+00  3.58125120e-01 -2.58064723e+00
   1.55215383e+00 -2.03553891e+00]
 [ 1.02665892e+01 -6.65782404e+00 -2.04501271e-01 -2.25653172e+00
  -6.31941366e+00  1.13084137e+00 -3.83885235e-01  8.22283030e-01
  -1.21192622e+00  3.33601260e+00]
 [-3.27186418e+00  1.00050325e+01  5.39114550e-02 -1.44938648e+00
  -9.89762247e-01 -2.09957671e+00 -1.49389958e+00  6.52510405e-01
   1.73153889e+00 -1.25597775e+00]
 [ 5.72116375e-01 -3.36192799e+00 -6.68362260e-01 -2.81247520e+00
   8.36382389e+00 -3.67477946e-02  2.23792076e+00 -2.91093756e-02
  -4.56922323e-01 -6.77382052e-01]
 [ 1.18602552e+01 -5.09683752e+00  4.54203248e-01 -2.55723000e+00
  -8.68753910e+00  6.96948707e-01 -1.50591761e-01 -3.62227589e-01
   9.83437955e-01  7.46711075e-01]
 [ 7.33289337e+00 -6.65414715e+00  1.57180679e+00 -2.62657452e+00
   4.11511570e-01 -1.35336161e+00 -1.40558392e-01  3.81030589e-01
   1.73799121e+00  8.02671254e-01]
 [-3.02898431e+00  1.26861107e+00 -2.04946566e+00 -2.52499342e-01
  -2.73597687e-01 -3.01714039e+00 -7.10914516e+00  1.10452967e+01
  -5.82177579e-01  1.86712158e+00]
 [-7.78098392e+00 -6.01984358e+00  1.23355007e+00  1.18682652e+01
  -9.83472538e+00  8.27242088e+00 -1.02135544e+01  3.95661980e-01
   6.63226461e+00  3.33681512e+00]
 [-2.72245955e+00 -6.74849796e+00 -6.24665642e+00  3.11165476e+00
  -4.71174330e-01  1.22390661e+01 -1.23519528e+00 -1.24356663e+00
   1.26693976e+00  5.81862879e+00]
 [-5.65229607e+00 -1.25138938e+00  3.68380380e+00  1.24947300e+01
  -8.21508980e+00  1.61641145e+00 -8.01925087e+00  8.37018967e-01
  -2.64613247e+00  7.92313635e-01]
 [-3.73405719e+00 -3.41621947e+00 -7.94842839e-01  4.55352879e+00
  -2.28238964e+00  1.88887548e+00 -5.84129477e+00  6.03430390e-01
   1.05920439e+01  2.25430655e+00]
 [-5.44103146e+00 -5.48421431e+00 -3.62234282e+00  1.20194650e+00
   3.48899674e+00  1.50794566e+00 -6.30612850e+00  4.01568127e+00
   1.61318648e+00  9.87832165e+00]
 [-3.34073186e+00  8.10987663e+00 -6.43497527e-01 -1.64372277e+00
  -4.42907363e-01 -1.46176386e+00 -8.56327295e-01  5.20323329e-02
   1.73289025e+00 -8.17061365e-01]
 [-6.88457203e+00  1.38391244e+00  1.33096969e+00  1.28132534e+01
  -6.20939922e+00  1.48244214e+00 -6.59804583e+00 -1.38118923e+00
   4.26289368e+00 -1.22962976e+00]
 [-6.09051991e+00 -3.15275192e+00  1.79273260e+00  9.92699528e+00
  -5.97349882e+00  3.68225765e+00 -6.47421646e+00 -1.99264419e+00
   2.15714622e+00  2.32836318e+00]
 [-3.25946307e+00  8.14360428e+00 -1.00535810e+00 -2.37552500e+00
   2.38139248e+00 -2.92597318e+00 -1.54173911e+00  2.25682306e+00
  -2.83430189e-01 -1.33554244e+00]
 [-2.99147058e+00  3.86941671e+00  8.82810593e+00  2.20121431e+00
  -8.40485859e+00 -8.66728902e-01 -5.97998762e+00 -5.21699572e+00
   5.80638123e+00 -2.57314467e+00]
 [ 8.64277363e+00 -4.99241495e+00  2.84688592e+00 -4.15350378e-01
  -1.87728360e-01 -2.40291572e+00  4.42544132e-01 -4.54446167e-01
  -1.88113344e+00 -1.23334014e+00]
 [-2.00169897e+00 -2.65497804e+00  1.18750989e+00  9.70900059e-01
  -4.53840446e+00 -2.65584946e+00 -8.23472023e+00  9.93836498e+00
  -5.57100773e-01  3.42955470e+00]
 [-3.57249069e+00 -5.03176594e+00 -1.79369414e+00 -5.03321826e-01
  -1.97100627e+00  9.01608944e+00  6.62497377e+00 -5.48222637e+00
   6.09256268e+00 -4.71334040e-01]
 [-5.27715540e+00 -7.84428477e-01 -6.26944721e-01  3.87298250e+00
  -1.88836837e+00  1.15252662e+00 -2.98473048e+00 -3.10233998e+00
   9.71112537e+00  3.10839200e+00]
 [-9.50223565e-01 -6.47654009e+00  2.26750326e+00  1.95419586e+00
   1.68217969e+00  1.66003108e+00  9.82697105e+00 -9.94868219e-01
  -2.03924966e+00 -1.88321277e-01]
 [-3.11575246e+00  3.43664408e+00  1.19877796e+01  4.36916590e+00
  -1.17812777e+01 -1.69431508e+00 -5.82668829e+00 -5.09948444e+00
   4.15738583e+00 -4.30461359e+00]
 [ 9.72177792e+00 -5.31352401e-01 -1.21784186e+00 -1.07392669e+00
  -7.11223555e+00  1.67838800e+00  1.01826215e+00 -8.88240516e-01
   6.95959151e-01  2.38748863e-01]
 [-2.06619406e+00  1.86608231e+00  1.12100420e+01  4.22539425e+00
  -1.21493711e+01 -4.57662535e+00 -6.88935089e+00 -9.81215835e-01
   3.87611055e+00 -3.28470826e+00]
 [-6.73031902e+00 -2.54390073e+00 -1.10151446e+00  1.51524162e+01
  -1.10052080e+01  6.60323954e+00 -7.94388771e+00  3.31939721e+00
  -1.40840662e+00  2.65730071e+00]
 [-1.96954179e+00 -1.13817227e+00  9.40351069e-01 -1.75684047e+00
   3.60373807e+00  2.01377797e+00  1.00558109e+01 -1.10547984e+00
   5.17374456e-01 -3.94047165e+00]
 [-5.81787634e+00 -1.20211565e+00 -3.53216052e+00  1.17569458e+00
   4.21314144e+00 -2.53644252e+00 -7.64466667e+00  4.19782829e+00
   4.28840429e-01  1.04579344e+01]
 [-4.20310974e+00 -3.19272375e+00 -4.62792778e+00  2.71683741e+00
   4.43899345e+00  3.31357956e-01 -6.24839544e+00  3.80388188e+00
  -1.22620119e-02  9.65024757e+00]
 [-8.26945066e-01 -5.25947523e+00  3.72772887e-02  2.30585241e+00
  -4.95726252e+00 -1.19987357e+00 -1.20395079e+01  1.53253164e+01
  -2.10372299e-01  1.89387524e+00]
 [-5.09596729e+00 -7.76027665e-02 -9.53466833e-01  2.89041376e+00
  -1.50858855e+00  2.27854323e+00 -1.95591903e+00 -3.15785193e+00
   1.00103540e+01  1.08987451e+00]
 [-4.01680946e-01 -4.62062168e+00  3.90530303e-02 -1.66790807e+00
   5.43311167e+00 -1.78802896e+00 -2.88405848e+00  2.93439984e+00
  -2.16558409e+00  8.71198368e+00]
 [-1.29969406e+00 -3.92871022e+00 -3.82151055e+00 -2.93831253e+00
   1.03674269e+01  8.88044477e-01  7.88922787e-01  3.86107159e+00
   1.60807288e+00 -3.76913965e-01]
 [-2.25020099e+00 -8.17249107e+00 -1.82360613e+00 -5.90175152e-01
   4.72407389e+00  9.39436078e-01 -3.85674310e+00  3.95303702e+00
   1.83473241e+00  9.13874054e+00]
 [-3.26617742e+00 -2.91517663e+00  8.37770653e+00  1.61820054e-01
  -2.98638320e+00 -2.47211266e+00  5.08574843e-01  4.65608168e+00
   2.66001201e+00 -4.67262363e+00]
 [-2.29874635e+00  7.77097034e+00  1.11359918e+00 -2.06103897e+00
  -7.61267126e-01  1.00877440e+00  1.47708499e+00 -1.20483887e+00
   1.99922264e+00 -3.81118345e+00]
 [-6.87821198e+00 -9.18823600e-01  2.16773844e+00  1.07671242e+01
  -7.48823595e+00  2.90310860e+00 -1.02075748e+01 -3.83400464e+00
   4.76818371e+00  4.06564474e+00]
 [-2.06487226e+00  8.76828384e+00  1.10449910e+00 -2.29669046e+00
  -1.15668392e+00 -2.50351834e+00 -1.69508122e-02 -1.05916834e+00
   1.91057479e+00 -2.64592767e+00]
 [-2.24318981e+00  9.02024174e+00  1.38990092e+00 -2.72154903e+00
   1.46101296e-01 -4.43454313e+00 -8.21092844e-01 -2.40808502e-01
   3.36577922e-01 -2.63193059e+00]
 [-4.35961342e+00 -7.74704576e-01 -2.74345660e+00 -3.27951574e+00
   1.50971518e+01 -2.80669570e+00 -1.28740633e+00  3.94157290e+00
   4.20372874e-01  8.37333024e-01]
 [-2.80749345e+00 -3.33036280e+00 -1.00865018e+00  4.57633829e+00
  -5.03952360e+00  2.93345642e+00 -8.54609489e+00  2.26549125e+00
   4.73208952e+00  5.93849993e+00]
 [-3.31042576e+00  9.97719002e+00  4.38573778e-01 -1.35296178e+00
  -1.21057940e+00 -2.46178842e+00 -8.34564090e-01  2.19030753e-01
   2.03147411e+00 -1.80211437e+00]
 [-2.14534068e+00 -2.93023801e+00  4.41405416e-01 -2.29865336e+00
   1.47422533e+01 -1.86358702e+00  4.61042017e-01 -6.20108247e-01
  -1.36792552e+00  2.14018896e-01]
 [-2.64241481e+00 -2.28332114e+00  2.01109338e+00  9.67352509e-01
   6.09287119e+00 -2.35626236e-01 -3.02941656e+00  4.32772923e+00
  -4.63955021e+00  3.73136783e+00]
 [-4.55847168e+00  1.04014337e+00  9.12987328e+00  2.06433630e+00
  -1.67355919e+00 -1.49593079e+00  4.09124941e-01  2.41894865e+00
  -6.86871633e-02 -4.42179346e+00]
 [-6.10608578e-01 -2.73860097e+00 -1.09864855e+00 -5.68899512e-01
   2.45831108e+00  5.50326490e+00  1.22601585e+01 -5.23877192e+00
   2.11066246e+00 -2.98584485e+00]
 [-5.96745872e+00 -1.91458237e+00 -3.10774088e-01  1.00216856e+01
  -3.81997776e+00  7.14399862e+00 -4.61386251e+00 -5.18248987e+00
   4.25162363e+00  1.18878789e-01]
 [-4.24126434e+00 -9.63249326e-01  1.06391013e+00  4.45316315e+00
  -4.47125340e+00 -1.21906054e+00 -9.75789547e+00  1.10335569e+01
  -1.17632782e+00  8.78942788e-01]
 [ 3.76867175e-01 -4.75102758e+00 -2.59345794e+00 -3.96257102e-01
  -7.50329159e-03  1.81642962e+00 -6.01041269e+00  9.97849655e+00
  -2.57468176e+00  5.00644207e+00]
 [-2.21995306e+00 -4.23465443e+00 -2.19536662e+00 -3.71420813e+00
   1.49460211e+01 -2.73240638e+00  3.03538777e-02  4.29108334e+00
   1.21963143e+00  6.85284913e-01]
 [-1.93794692e+00 -3.39166284e+00 -3.41372967e-01 -2.16144085e-01
   1.32588074e-01 -3.83050554e-02 -7.32452822e+00  9.68561840e+00
  -1.16044319e+00  3.63913298e+00]
 [-3.31972694e+00 -4.84879112e+00 -3.41381001e+00  1.93332338e+00
  -5.16150045e+00  1.05730495e+01  1.52961123e+00 -4.64916992e+00
   4.11477995e+00  4.80105543e+00]
 [-3.22445488e+00  1.01509037e+01  6.03635088e-02 -1.48001885e+00
  -3.28380197e-01 -2.36782789e+00 -8.66441727e-01  6.68077886e-01
   1.45576179e+00 -1.95271623e+00]
 [-5.64618587e+00 -3.71156931e+00 -2.31174397e+00  1.76912701e+00
   2.95111752e+00  2.09562635e+00 -5.34609461e+00 -2.59399921e-01
   1.10373080e+00  1.10444403e+01]
 [ 1.32743053e+01 -5.11389780e+00  1.10446036e+00 -5.01595545e+00
  -4.87907743e+00  4.46035750e-02  4.57046556e+00 -1.76434004e+00
  -5.40824793e-02 -1.01547205e+00]
 [-4.27589798e+00  1.26832044e+00  6.49948978e+00 -8.06193352e-02
   1.34645328e-01  6.92090929e-01 -8.92272711e-01  2.15252757e+00
  -9.95365143e-01 -2.46636438e+00]
 [-5.08349514e+00 -1.79646879e-01  9.57399654e+00  3.35643005e+00
  -3.42183971e+00 -2.33653522e+00 -1.98645079e+00  1.50538552e+00
  -1.43313253e+00 -9.75638926e-01]
 [-8.48450565e+00 -4.32870531e+00 -1.54253757e+00  1.61029205e+01
  -8.41084957e+00  5.45092726e+00 -1.00705996e+01  2.31331086e+00
   1.05400957e-01  5.19563723e+00]
 [-4.89484310e+00 -4.33120441e+00 -3.71932483e+00 -1.18670315e-01
   4.70187807e+00  2.67010808e-01 -5.52650118e+00  5.62736416e+00
   5.05499423e-01  8.69004250e+00]
 [-1.94048929e+00  7.98250961e+00 -3.75457823e-01 -2.09433126e+00
  -4.63896915e-02 -2.64650702e+00 -5.73348761e-01 -1.34597674e-01
   1.26143038e+00 -1.32178509e+00]
 [-6.36190224e+00  5.44552183e+00  1.89667892e+00  3.91665459e+00
  -1.99860716e+00  2.76620483e+00  3.57649279e+00 -3.53898597e+00
  -1.21411644e-01 -2.60307193e+00]
 [-3.60333633e+00  4.92228556e+00  2.87770915e+00  1.31902504e+00
  -4.28756446e-01 -3.29862523e+00 -2.29294825e+00 -1.10349190e+00
   3.81862259e+00 -1.23572731e+00]
 [ 1.06328213e+00 -6.86575174e+00 -3.61938500e+00  1.15000987e+00
  -2.32747698e+00  1.32029047e+01 -3.19671059e+00  8.91836137e-02
   4.70500660e+00  2.51928687e+00]
 [ 1.00363417e+01 -8.02793884e+00  9.38569680e-02 -2.78522468e+00
  -2.84671545e+00  2.95867848e+00  4.23545599e+00  2.52083421e+00
  -3.35666537e+00  1.45630157e+00]]

你甚至不願意Start的Github

ai_fast_handbook