e96031413

# LabelSmoothing.py

# https://www.aiuai.cn/aifarm1333.html 示例 3
# From: Github - NVIDIA/DeepLearningExamples/PyTorch/Classification
# smoothing.py
import torch
import torch.nn as nn

# 一般版本LabelSmoothing
class LabelSmoothing(nn.Module):

e96031413

# https://stackoverflow.com/a/45925049/13369757
import matplotlib.pyplot as plt 
import numpy as np

fig, host = plt.subplots(figsize=(12,10)) # (width, height) in inches
    
par1 = host.twinx()
par2 = host.twinx()

host.set_xlabel("Threshold")

e96031413

for i, (img, target,_) in enumerate(tqdm(dataloader)):
    
    feat_list = []
    def hook(module, input, output): 
        
        # 由於MobileNetv2不像ResNet18有宣告self.avgpool()，因此我的作法是將模型卷積層的最後一層的輸出手動進行adaptive_avg_pool2d
        # 接著將它加到feature_list中

        feat_list.append(nn.functional.adaptive_avg_pool2d(output.clone(), (1, 1)).reshape(output.clone().shape[0], -1).detach())                
    

e96031413

import numpy as np
import torch
import pickle
from torch.utils.data import Dataset, DataLoader

# 提取特徵
def extract(loader):
    featureExtractor.eval()
    features = None
    with torch.no_grad():

e96031413

import os
import shutil

fileListingFile = "missclassified_file_path.txt"
outputDir = "./"

# sample path in txt: /root/notebooks/XXXXX/datasets/XXXXXXXX.jpg

with open(fileListingFile, "r") as file:
    fileNames = [file.strip() for file in file.readlines()]

e96031413

batch_iterator = iter(dataloader)

images, targets = next(batch_iterator)
try:
	images, targets = next(batch_iterator)
except StopIteration:
	batch_iterator = iter(data_loader)
	images, targets = next(batch_iterator)

e96031413

# https://github.com/rajatkoner08/oodformer/blob/master/networks/resnet.py
'''ResNet in PyTorch.
BasicBlock and Bottleneck module is from the original ResNet paper:
[1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
    Deep Residual Learning for Image Recognition. arXiv:1512.03385
PreActBlock and PreActBottleneck module is from the later paper:
[2] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
    Identity Mappings in Deep Residual Networks. arXiv:1603.05027
Original code is from https://github.com/kuangliu/pytorch-cifar/blob/master/models/resnet.py
'''

e96031413

from torchvision.utils import save_image

features, imgs, labels, image_path = get_features_trained_weight(model, transform_dataset)

img1 = imgs[1] #第1張圖片
img1 = torch.from_numpy(img1) #格式轉換
save_image(img1, 'img1.png')  #保存成'img1.png'

e96031413

#初始化
$ heroku login #登入heroku
$ git init     #設定git
$ heroku git:remote -a your-app-name #填入在Heroku中設定好的名字
$ heroku config:set DISABLE_COLLECTSTATIC=1

#Push到Heroku
$ git add .
$ git commit -am "make it better"
$ git push heroku master

e96031413

def get_features_trained_weight(model, transform_dataset):
    if torch.cuda.is_available():
        device = 'cuda'
    else:
        device = 'cpu'
    if isinstance(model,torch.nn.DataParallel):
        model = model.module

    model.eval()
    model.to(device)