test_single.py

#Import Libraries
print("Importing libraries...")
import os, os.path
import numpy as np
import torch
import torch.nn as nn 
import torchvision
from torchvision import datasets, models, transforms
from sklearn.svm import SVC 
from sklearn.svm import LinearSVC
from sklearn.metrics import confusion_matrix
from PIL import Image
import pickle
print("Libraries imported")

#Directories
HOME_DIR = os.getcwd()
img_name = str(input("Enter complete path to image: \n"))
OBJECT_REC = "modak" #Name of object to be recognized. Replace it with required object name

#Similar transforms as ImageNet on which model has been trained
from torchvision import transforms
transform = transforms.Compose([            #[1]
 transforms.Resize(224),                    #[2]
 transforms.ToTensor(),                     #[3]
 transforms.Normalize(                      #[4]
 mean=[0.485, 0.456, 0.406],                #[5]
 std=[0.229, 0.224, 0.225]                  #[6]
 )])


#Pretrained VGG19 without final softmax layer
model_conv = torchvision.models.vgg19(pretrained=True)
for param in model_conv.features.parameters():
    param.requires_grad = False
for param in model_conv.classifier.parameters():
    param.requires_grad = False
new_classifier = nn.Sequential(*list(model_conv.classifier.children())[:-1])
model_conv.classifier = new_classifier
model_conv.eval()

#Load the classifier
with open('linearSVM.pkl', 'rb') as fid:
    svm_model_linear = pickle.load(fid)

    image = Image.open(img_name)
    image_t = transform(image)
    batch_t = torch.unsqueeze(image_t, 0)
    out = model_conv(batch_t)
    test_x = np.asarray(out)

svm_predictions = svm_model_linear.predict(test_x)
if(svm_predictions == 1):
    print("Image is of " + OBJECT_REC)
else:
    print("Image is not of " + OBJECT_REC)