別玩手機 影象分類比賽
2023-05-10 21:00:35
浙大寧波理工學院資料科學與巨量資料專業 別玩手機影象分類比賽
1 選手禁止互相抄襲，發現結果雷同者將取消成績；
2 請在基線模型基礎上修改程式碼，不允許使用第三方封裝庫、套件或者其他工具，否則做 0 分處理；
3 每位同學請獨立完成比賽，不允許就比賽技術問題進行相互交流，更不允許索要程式碼，請自覺遵守規則，保持良好的品格；
4 晚上 12：00 以後不允許遞交，否則做 0 分處理；
5 結果檔案必須是程式生成，不允許手動修改或者後期處理。
賽題背景
如今，手機已成為大眾離不開的生活工具，而且它的迅速發展使得它的功能不再以通訊為主，手機逐漸發展為可移動的大眾傳播媒體終端裝置，甚至可以比作為第五媒體。當今的大學生群體是智慧手機使用者中的一支巨大的的隊伍，零零後大學生在進入大學以來，學習生活中過度的依賴手機，甚至上課時忘記攜帶手機便會手足無措，神情恍惚。本比賽要求通過監控攝像頭等拍攝到的畫面判斷畫面中的人物是否正在使用手機
資料集介紹
本比賽採用的資料集中,訓練集共 2180 張使用手機的圖片（位於目錄 data/data146247/train/0_phone/）、1971 張沒有使用手機的圖片（位於目錄 data/data146247/train/1_no_phone/）
測試集共 1849 張圖片，無標註資訊
總體思路
本 Baseline 採用 LeNet 模型架構，參賽者可以在此基礎上進行修改，也可以使用全新的網路架構衝榜~~
LeNet 網路結構圖如下
LeNet-5 共有 7 層，不包含輸入，每層都包含可訓練引數；每個層有多個 Feature Map，每個 FeatureMap 通過一種折積濾波器提取輸入的一種特徵，然後每個 FeatureMap 有多個神經元。
- 預處理
- 生成標籤
- 讀取資料集
- 訓練
- 模型組網
- 反向傳播
- 預測
- 模型預測並儲存結果
in[1]
!rm aug.csv
!rm test.csv
!rm train.csv
!rm -r augment
!rm -r train
!rm -r test
in[2]
!unzip -oq /home/aistudio/data/data146247/train.zip
!unzip -oq /home/aistudio/data/data146247/test.zip #解壓資料集
!mkdir augment
!mkdir augment/0_phone
!mkdir augment/1_no_phone
in[3]
import numpy as np
import pandas as pd
import paddle.nn.functional as F
import paddle.vision.transforms as T
from paddle.nn import Conv2D, BatchNorm2D, Linear, Dropout
from paddle.nn import AdaptiveAvgPool2D, MaxPool2D, AvgPool2D
import paddle
import paddle.nn as nn
import os
from paddle.io import Dataset, DataLoader, IterableDataset
from sklearn.utils import shuffle
from PIL import Image
from paddle.vision.transforms import Resize
PLACE = paddle.CUDAPlace(0) # 在gpu上訓練
預處理
讀取解壓好的資料集,將圖片名全部儲存 train.csv(訓練集)和 test.csv(測試集)中
處理完後,就可以定義資料集類並讀取資料集了
in[4]
# 寫入訓練集csv
list1=[]
for path in os.listdir("train/1_no_phone") :
if path[-3:]=='jpg':
k=["train/1_no_phone/"+path,1]
list1.append(k)
for path in os.listdir("train/0_phone") :
if path[-3:]=='jpg':
k=["train/0_phone/"+path,0]
list1.append(k)
result_df = pd.DataFrame(list1)
result_df.columns=['image_id', 'label']
data = shuffle(result_df)
data.to_csv('train.csv', index=False, header=True)
in[5]
# 寫入測試集csv
list1=[]
for path in os.listdir("test") :
if path[-3:]=='jpg':
k=["test/"+path]
list1.append(k)
result_df = pd.DataFrame(list1)
result_df.columns=['image_id']
result_df.to_csv('test.csv', index=False, header=True)
in[6]
pic_list=pd.read_csv("train.csv")
pic_list=np.array(pic_list)
train_list=pic_list[:int(len(pic_list)*0.8)]
test_list=pic_list[int(len(pic_list)*0.8):]
print(len(train_list))
train_list
out[6]
3320
array([['train/1_no_phone/tWuCoxIkbAedBjParJZQGn9XVYLK16hS.jpg', 1],
['train/1_no_phone/rkX9Rj7YqENc85bFxdDaw14Is0yCeAnp.jpg', 1],
['train/0_phone/wfhSF1W5B46doJKEc7VTU2OxbCtRDAI9.jpg', 0],
...,
['train/1_no_phone/uvHPyhakBK89o43fjJRXgZ2CIEpGzSAQ.jpg', 1],
['train/0_phone/rlw854paLXkqc3zhgo9N2idUSeyIPAWf.jpg', 0],
['train/0_phone/Wy59OfsandAo1lp2S7e3cEutPbi4zqYF.jpg', 0]],
dtype=object)
離線影象增廣
對影象進行 70%概率的隨機水平翻轉，並對翻轉後的影象亮度、對比度和飽和度進行變化，將增廣後的影象儲存到 augment 資料夾，將影象資料追加到 train_list 後面，
in[7]
import paddle.vision.transforms as T
directory_name = "augment/"
for i in range(len(train_list)):
img_path = '/'.join(train_list[i][0].split('/')[:4])
img_path[6:-4]
img = Image.open(img_path)
img = T.RandomHorizontalFlip(0.7)(img)
# img.save(directory_name + "/" + img_path[6:-4] + "-tramsforms1.jpg")
img = T.ColorJitter(0.5, 0.5, 0.5, 0.0)(img)
img.save(directory_name + "/" + img_path[6:-4] + "-tramsforms.jpg")
in[8]
# 寫入訓練集csv
list4=[]
for path in os.listdir("augment/1_no_phone") :
if path[-3:]=='jpg':
k=["augment/1_no_phone/"+path,1]
list4.append(k)
for path in os.listdir("augment/0_phone") :
if path[-3:]=='jpg':
k=["augment/0_phone/"+path,0]
list4.append(k)
result_df = pd.DataFrame(list4)
result_df.columns=['image_id', 'label']
data = shuffle(result_df)
data.to_csv('aug.csv', index=False, header=True)
in[9]
aug_list=pd.read_csv("aug.csv")
aug_list=np.array(aug_list)
train_list = np.append(train_list, aug_list, axis=0)
print(len(train_list))
train_list
out[9]
6640
array([['train/1_no_phone/tWuCoxIkbAedBjParJZQGn9XVYLK16hS.jpg', 1],
['train/1_no_phone/rkX9Rj7YqENc85bFxdDaw14Is0yCeAnp.jpg', 1],
['train/0_phone/wfhSF1W5B46doJKEc7VTU2OxbCtRDAI9.jpg', 0],
...,
['augment/1_no_phone/WPNLg72lV8Jhn5TpokKMf34QR1Fd6IcH-tramsforms.jpg',
1],
['augment/0_phone/qJcVtZH2yiAL6w34rOQopT9IemzEDjWK-tramsforms.jpg',
0],
['augment/0_phone/qwoDX3ENaVuCk9IYjAnBHbv7TLUPR5Kr-tramsforms.jpg',
0]], dtype=object)
in[10]
class H2ZDateset(Dataset):
def __init__(self, data_dir):
super(H2ZDateset, self).__init__()
self.pic_list=data_dir
def __getitem__(self, idx):
image_file,label=self.pic_list[idx]
img = Image.open(image_file) # 讀取圖片
img = img.resize((256, 256), Image.ANTIALIAS) # 圖片大小樣式歸一化
img = np.array(img).astype('float32') # 轉換成陣列型別浮點型32位元
img = img.transpose((2, 0, 1))
img = img/255.0 # 資料縮放到0-1的範圍
return img, np.array(label, dtype='int64').reshape(-1)
def __len__(self):
return len(self.pic_list)
in[11]
h2zdateset = H2ZDateset(train_list)
BATCH_SIZE = 32
loader = DataLoader(h2zdateset, places=PLACE, shuffle=True, batch_size=BATCH_SIZE, drop_last=False, num_workers=0, use_shared_memory=False)
data,label = next(loader())
print("讀取的資料形狀：", data.shape,label.shape)
out[11]
讀取的資料形狀： [32, 3, 256, 256] [32, 1]
in[12]
train_data = H2ZDateset(train_list)
test_data = H2ZDateset(test_list)
train_data_reader = DataLoader(train_data, places=PLACE, shuffle=True, batch_size=BATCH_SIZE, drop_last=False, num_workers=2, use_shared_memory=True)
test_data_reader = DataLoader(test_data, places=PLACE, shuffle=True, batch_size=BATCH_SIZE, drop_last=False, num_workers=2, use_shared_memory=True)
訓練
首先進行 LeNet 模型組網
這裡用到了 paddle.Model 的 summary()方法來將模型視覺化,通過 summary()可以快速列印模型的網路結構，並且，執行該語句的時候會執行一次網路。在動態圖中，我們需要手算網路的輸入和輸出層，如果出現一點問題就會報錯非常麻煩，而 summary()能大大縮短 debug 時間
自定義 ResNeXt 類，匯入 resnext101_64x4d 網路，並開啟預訓練模型的選項，增一個線性層，將原本模型的 1000 分類問題變成 2 分類問題
in[13]
class ResNeXt(nn.Layer):
def __init__(self):
super(ResNeXt, self).__init__()
self.layer = paddle.vision.models.resnext101_64x4d(pretrained=True)
self.fc = nn.Sequential(
nn.Dropout(0.5),
nn.Linear(1000, 2)
)
def forward(self, inputs):
outputs = self.layer(inputs)
outputs = self.fc(outputs)
return outputs
model = ResNeXt()
paddle.Model(model).summary((-1, 3, 256, 256))
out[13]
W0510 00:32:22.100802 8743 gpu_context.cc:244] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 10.1
W0510 00:32:22.104816 8743 gpu_context.cc:272] device: 0, cuDNN Version: 7.6.
-------------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param #
===============================================================================
Conv2D-1 [[1, 3, 256, 256]] [1, 64, 128, 128] 9,408
BatchNorm-1 [[1, 64, 128, 128]] [1, 64, 128, 128] 256
ConvBNLayer-1 [[1, 3, 256, 256]] [1, 64, 128, 128] 0
MaxPool2D-1 [[1, 64, 128, 128]] [1, 64, 64, 64] 0
Conv2D-2 [[1, 64, 64, 64]] [1, 256, 64, 64] 16,384
BatchNorm-2 [[1, 256, 64, 64]] [1, 256, 64, 64] 1,024
ConvBNLayer-2 [[1, 64, 64, 64]] [1, 256, 64, 64] 0
Conv2D-3 [[1, 256, 64, 64]] [1, 256, 64, 64] 9,216
BatchNorm-3 [[1, 256, 64, 64]] [1, 256, 64, 64] 1,024
ConvBNLayer-3 [[1, 256, 64, 64]] [1, 256, 64, 64] 0
Conv2D-4 [[1, 256, 64, 64]] [1, 256, 64, 64] 65,536
BatchNorm-4 [[1, 256, 64, 64]] [1, 256, 64, 64] 1,024
ConvBNLayer-4 [[1, 256, 64, 64]] [1, 256, 64, 64] 0
Conv2D-5 [[1, 64, 64, 64]] [1, 256, 64, 64] 16,384
BatchNorm-5 [[1, 256, 64, 64]] [1, 256, 64, 64] 1,024
ConvBNLayer-5 [[1, 64, 64, 64]] [1, 256, 64, 64] 0
BottleneckBlock-1 [[1, 64, 64, 64]] [1, 256, 64, 64] 0
Conv2D-6 [[1, 256, 64, 64]] [1, 256, 64, 64] 65,536
BatchNorm-6 [[1, 256, 64, 64]] [1, 256, 64, 64] 1,024
ConvBNLayer-6 [[1, 256, 64, 64]] [1, 256, 64, 64] 0
Conv2D-7 [[1, 256, 64, 64]] [1, 256, 64, 64] 9,216
BatchNorm-7 [[1, 256, 64, 64]] [1, 256, 64, 64] 1,024
ConvBNLayer-7 [[1, 256, 64, 64]] [1, 256, 64, 64] 0
Conv2D-8 [[1, 256, 64, 64]] [1, 256, 64, 64] 65,536
BatchNorm-8 [[1, 256, 64, 64]] [1, 256, 64, 64] 1,024
ConvBNLayer-8 [[1, 256, 64, 64]] [1, 256, 64, 64] 0
BottleneckBlock-2 [[1, 256, 64, 64]] [1, 256, 64, 64] 0
Conv2D-9 [[1, 256, 64, 64]] [1, 256, 64, 64] 65,536
BatchNorm-9 [[1, 256, 64, 64]] [1, 256, 64, 64] 1,024
ConvBNLayer-9 [[1, 256, 64, 64]] [1, 256, 64, 64] 0
Conv2D-10 [[1, 256, 64, 64]] [1, 256, 64, 64] 9,216
BatchNorm-10 [[1, 256, 64, 64]] [1, 256, 64, 64] 1,024
ConvBNLayer-10 [[1, 256, 64, 64]] [1, 256, 64, 64] 0
Conv2D-11 [[1, 256, 64, 64]] [1, 256, 64, 64] 65,536
BatchNorm-11 [[1, 256, 64, 64]] [1, 256, 64, 64] 1,024
ConvBNLayer-11 [[1, 256, 64, 64]] [1, 256, 64, 64] 0
BottleneckBlock-3 [[1, 256, 64, 64]] [1, 256, 64, 64] 0
Conv2D-12 [[1, 256, 64, 64]] [1, 512, 64, 64] 131,072
BatchNorm-12 [[1, 512, 64, 64]] [1, 512, 64, 64] 2,048
ConvBNLayer-12 [[1, 256, 64, 64]] [1, 512, 64, 64] 0
Conv2D-13 [[1, 512, 64, 64]] [1, 512, 32, 32] 36,864
BatchNorm-13 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-13 [[1, 512, 64, 64]] [1, 512, 32, 32] 0
Conv2D-14 [[1, 512, 32, 32]] [1, 512, 32, 32] 262,144
BatchNorm-14 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-14 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
Conv2D-15 [[1, 256, 64, 64]] [1, 512, 32, 32] 131,072
BatchNorm-15 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-15 [[1, 256, 64, 64]] [1, 512, 32, 32] 0
BottleneckBlock-4 [[1, 256, 64, 64]] [1, 512, 32, 32] 0
Conv2D-16 [[1, 512, 32, 32]] [1, 512, 32, 32] 262,144
BatchNorm-16 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-16 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
Conv2D-17 [[1, 512, 32, 32]] [1, 512, 32, 32] 36,864
BatchNorm-17 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-17 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
Conv2D-18 [[1, 512, 32, 32]] [1, 512, 32, 32] 262,144
BatchNorm-18 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-18 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
BottleneckBlock-5 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
Conv2D-19 [[1, 512, 32, 32]] [1, 512, 32, 32] 262,144
BatchNorm-19 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-19 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
Conv2D-20 [[1, 512, 32, 32]] [1, 512, 32, 32] 36,864
BatchNorm-20 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-20 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
Conv2D-21 [[1, 512, 32, 32]] [1, 512, 32, 32] 262,144
BatchNorm-21 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-21 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
BottleneckBlock-6 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
Conv2D-22 [[1, 512, 32, 32]] [1, 512, 32, 32] 262,144
BatchNorm-22 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-22 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
Conv2D-23 [[1, 512, 32, 32]] [1, 512, 32, 32] 36,864
BatchNorm-23 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-23 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
Conv2D-24 [[1, 512, 32, 32]] [1, 512, 32, 32] 262,144
BatchNorm-24 [[1, 512, 32, 32]] [1, 512, 32, 32] 2,048
ConvBNLayer-24 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
BottleneckBlock-7 [[1, 512, 32, 32]] [1, 512, 32, 32] 0
Conv2D-25 [[1, 512, 32, 32]] [1, 1024, 32, 32] 524,288
BatchNorm-25 [[1, 1024, 32, 32]] [1, 1024, 32, 32] 4,096
ConvBNLayer-25 [[1, 512, 32, 32]] [1, 1024, 32, 32] 0
Conv2D-26 [[1, 1024, 32, 32]] [1, 1024, 16, 16] 147,456
BatchNorm-26 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-26 [[1, 1024, 32, 32]] [1, 1024, 16, 16] 0
Conv2D-27 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-27 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-27 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-28 [[1, 512, 32, 32]] [1, 1024, 16, 16] 524,288
BatchNorm-28 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-28 [[1, 512, 32, 32]] [1, 1024, 16, 16] 0
BottleneckBlock-8 [[1, 512, 32, 32]] [1, 1024, 16, 16] 0
Conv2D-29 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-29 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-29 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-30 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-30 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-30 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-31 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-31 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-31 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-9 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-32 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-32 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-32 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-33 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-33 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-33 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-34 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-34 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-34 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-10 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-35 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-35 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-35 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-36 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-36 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-36 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-37 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-37 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-37 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-11 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-38 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-38 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-38 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-39 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-39 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-39 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-40 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-40 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-40 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-12 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-41 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-41 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-41 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-42 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-42 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-42 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-43 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-43 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-43 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-13 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-44 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-44 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-44 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-45 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-45 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-45 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-46 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-46 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-46 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-14 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-47 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-47 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-47 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-48 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-48 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-48 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-49 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-49 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-49 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-15 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-50 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-50 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-50 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-51 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-51 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-51 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-52 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-52 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-52 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-16 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-53 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-53 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-53 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-54 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-54 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-54 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-55 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-55 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-55 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-17 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-56 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-56 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-56 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-57 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-57 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-57 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-58 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-58 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-58 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-18 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-59 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-59 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-59 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-60 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-60 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-60 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-61 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-61 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-61 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-19 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-62 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-62 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-62 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-63 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-63 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-63 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-64 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-64 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-64 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-20 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-65 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-65 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-65 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-66 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-66 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-66 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-67 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-67 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-67 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-21 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-68 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-68 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-68 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-69 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-69 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-69 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-70 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-70 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-70 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-22 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-71 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-71 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-71 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-72 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-72 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-72 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-73 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-73 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-73 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-23 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-74 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-74 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-74 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-75 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-75 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-75 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-76 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-76 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-76 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-24 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-77 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-77 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-77 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-78 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-78 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-78 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-79 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-79 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-79 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-25 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-80 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-80 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-80 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-81 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-81 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-81 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-82 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-82 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-82 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-26 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-83 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-83 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-83 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-84 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-84 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-84 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-85 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-85 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-85 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-27 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-86 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-86 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-86 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-87 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-87 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-87 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-88 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-88 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-88 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-28 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-89 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-89 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-89 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-90 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-90 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-90 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-91 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-91 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-91 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-29 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-92 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-92 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-92 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-93 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 147,456
BatchNorm-93 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-93 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-94 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 1,048,576
BatchNorm-94 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 4,096
ConvBNLayer-94 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
BottleneckBlock-30 [[1, 1024, 16, 16]] [1, 1024, 16, 16] 0
Conv2D-95 [[1, 1024, 16, 16]] [1, 2048, 16, 16] 2,097,152
BatchNorm-95 [[1, 2048, 16, 16]] [1, 2048, 16, 16] 8,192
ConvBNLayer-95 [[1, 1024, 16, 16]] [1, 2048, 16, 16] 0
Conv2D-96 [[1, 2048, 16, 16]] [1, 2048, 8, 8] 589,824
BatchNorm-96 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 8,192
ConvBNLayer-96 [[1, 2048, 16, 16]] [1, 2048, 8, 8] 0
Conv2D-97 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 4,194,304
BatchNorm-97 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 8,192
ConvBNLayer-97 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 0
Conv2D-98 [[1, 1024, 16, 16]] [1, 2048, 8, 8] 2,097,152
BatchNorm-98 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 8,192
ConvBNLayer-98 [[1, 1024, 16, 16]] [1, 2048, 8, 8] 0
BottleneckBlock-31 [[1, 1024, 16, 16]] [1, 2048, 8, 8] 0
Conv2D-99 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 4,194,304
BatchNorm-99 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 8,192
ConvBNLayer-99 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 0
Conv2D-100 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 589,824
BatchNorm-100 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 8,192
ConvBNLayer-100 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 0
Conv2D-101 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 4,194,304
BatchNorm-101 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 8,192
ConvBNLayer-101 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 0
BottleneckBlock-32 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 0
Conv2D-102 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 4,194,304
BatchNorm-102 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 8,192
ConvBNLayer-102 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 0
Conv2D-103 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 589,824
BatchNorm-103 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 8,192
ConvBNLayer-103 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 0
Conv2D-104 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 4,194,304
BatchNorm-104 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 8,192
ConvBNLayer-104 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 0
BottleneckBlock-33 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 0
AdaptiveAvgPool2D-1 [[1, 2048, 8, 8]] [1, 2048, 1, 1] 0
Linear-1 [[1, 2048]] [1, 1000] 2,049,000
ResNeXt-2 [[1, 3, 256, 256]] [1, 1000] 0
Dropout-1 [[1, 1000]] [1, 1000] 0
Linear-2 [[1, 1000]] [1, 2] 2,002
===============================================================================
Total params: 83,660,154
Trainable params: 83,254,394
Non-trainable params: 405,760
-------------------------------------------------------------------------------
Input size (MB): 0.75
Forward/backward pass size (MB): 1024.04
Params size (MB): 319.14
Estimated Total Size (MB): 1343.93
-------------------------------------------------------------------------------
{'total_params': 83660154, 'trainable_params': 83254394}
預測
模型訓練好之後就可以開始預測
在使用帶有動量的 SGD 學習率優化器後訓練 5 個 epoch，在 valid_list 中已經可以達到 95.3%的準確率
in[14]
epochs_num = 5 #迭代次數
# opt = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters())
opt = paddle.optimizer.Momentum(learning_rate=0.002, parameters=model.parameters(), weight_decay=0.005, momentum=0.9)
train_acc, train_loss, valid_acc = [], [], []
for pass_num in range(epochs_num):
model.train() #訓練模式
accs=[]
for batch_id,data in enumerate(train_data_reader):
images, labels = data
predict = model(images)#預測
loss=F.cross_entropy(predict,labels)
avg_loss=paddle.mean(loss)
acc=paddle.metric.accuracy(predict,labels)#計算精度
accs.append(acc.numpy()[0])
if batch_id % 20 == 0:
print("epoch:{}, iter:{}, loss:{}, acc:{}".format(pass_num, batch_id, avg_loss.numpy(), acc.numpy()[0]))
opt.clear_grad()
avg_loss.backward()
opt.step()
print("train_pass:{}, train_loss:{}, train_acc:{}".format(pass_num, avg_loss.numpy(), np.mean(accs)))
train_acc.append(np.mean(accs)), train_loss.append(avg_loss.numpy())
##得到驗證集的效能
model.eval()
val_accs=[]
for batch_id,data in enumerate(test_data_reader):
images, labels = data
predict=model(images)#預測
acc=paddle.metric.accuracy(predict,labels)#計算精度
val_accs.append(acc.numpy()[0])
print('val_acc={}'.format(np.mean(val_accs)))
valid_acc.append(np.mean(val_accs))
paddle.save(model.state_dict(),'resnext101_64x4d')#儲存模型
out[14]
epoch:0, iter:0, loss:[1.2695745], acc:0.5625
epoch:0, iter:20, loss:[0.72218126], acc:0.78125
epoch:0, iter:40, loss:[0.8521573], acc:0.625
epoch:0, iter:60, loss:[0.9325488], acc:0.78125
epoch:0, iter:80, loss:[0.43677044], acc:0.90625
epoch:0, iter:100, loss:[0.37998652], acc:0.84375
epoch:0, iter:120, loss:[0.17167747], acc:0.9375
epoch:0, iter:140, loss:[0.13679577], acc:0.9375
epoch:0, iter:160, loss:[0.27065688], acc:0.9375
epoch:0, iter:180, loss:[0.27734363], acc:0.875
epoch:0, iter:200, loss:[0.27193183], acc:0.90625
train_pass:0, train_loss:[0.37479421], train_acc:0.8610276579856873
val_acc=0.9120657444000244
epoch:1, iter:0, loss:[0.29650056], acc:0.875
epoch:1, iter:20, loss:[0.03968127], acc:0.96875
epoch:1, iter:40, loss:[0.0579604], acc:0.96875
epoch:1, iter:60, loss:[0.08308301], acc:0.96875
epoch:1, iter:80, loss:[0.02734788], acc:1.0
epoch:1, iter:100, loss:[0.04930278], acc:0.96875
epoch:1, iter:120, loss:[0.08352702], acc:0.96875
epoch:1, iter:140, loss:[0.180908], acc:0.96875
epoch:1, iter:160, loss:[0.00271969], acc:1.0
epoch:1, iter:180, loss:[0.23547903], acc:0.90625
epoch:1, iter:200, loss:[0.02033656], acc:1.0
train_pass:1, train_loss:[0.00160835], train_acc:0.9695011973381042
val_acc=0.9362205266952515
epoch:2, iter:0, loss:[0.06544636], acc:0.96875
epoch:2, iter:20, loss:[0.06581119], acc:0.96875
epoch:2, iter:40, loss:[0.00562927], acc:1.0
epoch:2, iter:60, loss:[0.01040748], acc:1.0
epoch:2, iter:80, loss:[0.03810382], acc:0.96875
epoch:2, iter:100, loss:[0.01447718], acc:1.0
epoch:2, iter:120, loss:[0.12421186], acc:0.96875
epoch:2, iter:140, loss:[0.00112416], acc:1.0
epoch:2, iter:160, loss:[0.003324], acc:1.0
epoch:2, iter:180, loss:[0.01755645], acc:1.0
epoch:2, iter:200, loss:[0.06159591], acc:0.96875
train_pass:2, train_loss:[0.00045859], train_acc:0.990234375
val_acc=0.9543269276618958
epoch:3, iter:0, loss:[0.00377628], acc:1.0
epoch:3, iter:20, loss:[0.0032537], acc:1.0
epoch:3, iter:40, loss:[0.00095566], acc:1.0
epoch:3, iter:60, loss:[0.1955388], acc:0.9375
epoch:3, iter:80, loss:[0.00345089], acc:1.0
epoch:3, iter:100, loss:[0.00279539], acc:1.0
epoch:3, iter:120, loss:[0.01505984], acc:1.0
epoch:3, iter:140, loss:[0.06922489], acc:0.96875
epoch:3, iter:160, loss:[0.05011526], acc:0.96875
epoch:3, iter:180, loss:[0.13640904], acc:0.96875
epoch:3, iter:200, loss:[0.01501912], acc:1.0
train_pass:3, train_loss:[0.00018014], train_acc:0.991135835647583
val_acc=0.944672703742981
epoch:4, iter:0, loss:[0.00011357], acc:1.0
epoch:4, iter:20, loss:[0.00145513], acc:1.0
epoch:4, iter:40, loss:[0.00594005], acc:1.0
epoch:4, iter:60, loss:[0.00403259], acc:1.0
epoch:4, iter:80, loss:[0.00525207], acc:1.0
epoch:4, iter:100, loss:[0.01096269], acc:1.0
epoch:4, iter:120, loss:[0.0166808], acc:1.0
epoch:4, iter:140, loss:[0.00298717], acc:1.0
epoch:4, iter:160, loss:[0.00306062], acc:1.0
epoch:4, iter:180, loss:[0.0080948], acc:1.0
epoch:4, iter:200, loss:[0.08889838], acc:0.9375
train_pass:4, train_loss:[0.07838897], train_acc:0.9887319803237915
val_acc=0.9530086517333984
in[15]
from tqdm import tqdm
test=pd.read_csv("test.csv")
test["label"]=0
with tqdm(total=len(test)) as pbar:
for i in range(len(test)):
image_file=test["image_id"][i]
pbar.set_description(f'Processing: {test["image_id"][i]}')
img = Image.open(image_file) # 讀取圖片
img = img.resize((256, 256), Image.ANTIALIAS) # 圖片大小樣式歸一化
img = np.array(img).astype('float32') # 轉換成陣列型別浮點型32位元
img = img.transpose((2, 0, 1))
img = img/255.0 # 資料縮放到0-1的範圍
img=paddle.to_tensor(img.reshape(-1,3,256,256))
predict = model(img)
result = int(np.argmax(predict.numpy()))
test.loc[i, "image_id"]=test.loc[i, "image_id"][5:]
test.loc[i, "label"]=result
pbar.update(1)
test.to_csv('predict_result.csv', index=False)
print("end")
out[15]
Processing: test/yheWlrsgStOjfLdqHbYE7p1P28ViDBQ9.jpg: 100%|██████████| 1849/1849 [01:58<00:00, 15.65it/s]
end
改進方向
- 可以在基線模型的基礎上通過調參及模型優化進一步提升效果
- 可以對訓練集進行資料增強從而增大訓練資料量以提升模型泛化能力
- 可以嘗試採用更深的神經網路，如 Resnet、VGG