2023.6.19 CNN + MNISTのサンプル(C×2+F×2)【torch】

(conv + ReLU + MaxPooling)×２ + FC×2

convとpooling層の各種パラメータを調整可能にした、画像サイズが小さくなりすぎなければ対応可能の筈

結果は./result/以下の日付＋時刻に格納する。

code:python

import torch

import torch.nn as nn

import torch.nn.functional as f

import matplotlib.pyplot as plt

import sys, numpy, time

from sklearn.model_selection import train_test_split

import datetime

import os

from sklearn.datasets import fetch_openml

##########

class Net(nn.Module):

def __init__(self, args):

super().__init__()

# Channel number

# Conv-layer

# Pooling, Window & Stride size

# FC-layer

#

OH1 = int(( H1 + 2*P1 - FH1)/S1 + 1) # conv1への入力、タテ

OW1 = int(( W1 + 2*P1 - FW1)/S1 + 1) # conv1への入力、ヨコ

OHP1 = int(OH1 // POOL1) # Pooling1からの出力、タテ

OWP1 = int(OW1 // POOL1) # Pooling1からの出力、ヨコ

OH2 = int((OHP1 + 2*P2 - FH2)/S2 + 1) # conv2への入力、タテ

OW2 = int((OWP1 + 2*P2 - FW2)/S2 + 1) # conv2への入力、ヨコ

OHP2 = int(OH2 // POOL2) # Pooling2からの出力、タテ

OWP2 = int(OW2 // POOL2) # Pooling2からの出力、ヨコ

print(

'output from conv1 :', OH1, OW1,

', outout from pool1 :', OHP1, OWP1,

', output from conv2 :', OH2, OW2,

', outout from pool2 :', OHP2, OWP2

, file=fp)

print(

'output from conv1 :', OH1, OW1,

', outout from pool1 :', OHP1, OWP1,

', output from conv2 :', OH2, OW2,

', outout from pool2 :', OHP2, OWP2

)

self.relu = nn.ReLU()

self.pool1 = nn.MaxPool2d(POOL1, stride=POOL1)

self.pool2 = nn.MaxPool2d(POOL2, stride=POOL2)

self.softmax = nn.Softmax(dim=1)

self.conv1 = nn.Conv2d(CH1_i, CH1_o, kernel_size=(FH1, FW1), padding=P1, stride=S1)

self.conv2 = nn.Conv2d(CH2_i, CH2_o, kernel_size=(FH2, FW2), padding=P2, stride=S2)

self.fc1 = nn.Linear(CH2_o*OHP2*OWP2, HIDDEN_FC)

self.fc2 = nn.Linear(HIDDEN_FC, OUTPUT)

def forward(self, x):

x = self.conv1(x)

x = self.relu(x)

x = self.pool1(x)

x = self.conv2(x)

x = self.relu(x)

x = self.pool2(x)

x = self.fc1(x)

x = self.relu(x)

x = self.fc2(x)

x = self.softmax(x)

return x

##########

def get_datename():

dt = datetime.datetime.now()

result = str(dt.year)

result = result + str(dt.month).zfill(2)

result = result + str(dt.day).zfill(2) + '_'

result = result + str(dt.hour).zfill(2)

result = result + str(dt.minute).zfill(2)

result = result + str(dt.second).zfill(2)

return result

##########

# 結果の保存用

dirname = './result/' + get_datename()

filename = 'result.md'

os.makedirs(dirname)

fp = open(dirname + '/' + filename, mode='w')

# データセットの取得

dataset = fetch_openml('mnist_784', data_home='~/scikit_learn_data', cache=True, parser='auto')

X = torch.tensor(X, dtype=torch.float32).view(-1, 28, 28) # (70000, 28, 28)

X = torch.unsqueeze(X, dim=1) # ここでチャネル数１にしておく (70000, 1, 28, 28)

y = torch.tensor(y) # (70000, )

test_size = 0.25

random_state = 42

output = 10

X_train, X_test, y_train, y_test = train_test_split(

X, y, test_size=test_size, random_state=random_state

)

y_train_onehot = f.one_hot(y_train, num_classes=output).float()

y_test_onehot = f.one_hot(y_test, num_classes=output).float()

train_dataset = torch.utils.data.TensorDataset(X_train, y_train_onehot)

train_loader = torch.utils.data.DataLoader(

dataset=train_dataset,

batch_size=100,

shuffle=True,

num_workers=0

)

# ネットワークの構成、(Conv2d + ReLU + MaxPoolin)×2 + FC×2

args = {}

print(args, file=fp)

print(args)

# 学習

EPOCHS = 10

lr = 0.01

net = Net(args)

optimizer = torch.optim.SGD(net.parameters(), lr=lr)

loss = nn.MSELoss()

obj_hist = []

print(net, file=fp)

print(net)

print(optimizer, file=fp)

print(optimizer)

print('EPOCHS =', EPOCHS, file=fp)

print('EPOCHS =', EPOCHS)

time_t0 = time.time()

for epoch in range(1, EPOCHS + 1):

time_t1 = time.time()

for X, y in train_loader:

optimizer.zero_grad()

y_predict = net(X)

obj = loss(y_predict, y)

obj_hist.append(obj.item())

obj.backward()

optimizer.step()

if epoch%(EPOCHS//10) == 0:

print('progress =', int(epoch/EPOCHS*100),'%, EPOCH =', epoch, ', loss ={:.3e}'.format(obj.item()), ', time =', time.time() - time_t1, file=fp)

print('progress =', int(epoch/EPOCHS*100),'%, EPOCH =', epoch, ', loss ={:.3e}'.format(obj.item()), ', time =', time.time() - time_t1)

print('total time =', time.time() - time_t0, file=fp)

print('total time =', time.time() - time_t0)

#

def calc_accuracy(net, X_test, y_test):

y_test_predict = net.forward(X_test)

y_test_predict_compare = y_test_predict_max_index == y_test

y_test_predict_accuracy = (y_test_predict_compare.sum() / len(y_test_predict_compare))

return y_test_predict_accuracy

#

result_accuracy = calc_accuracy(net, X_test, y_test)

print('accuracy = {:.8f}'.format(result_accuracy.item()), file=fp)

print('accuracy = {:.8f}'.format(result_accuracy.item()))

plt.plot(obj_hist, label='loss function')

plt.grid()

plt.legend()

plt.savefig(dirname + '/loss_function.svg')

torch.save(net, dirname + '/net_cnn_sp.pth') # ネットのパラメータを保存

fp.close()