Inconsistent predictions for pascal voc dataset

I am currently working on pascal voc dataset. I used Unets for getting segmentation masks from the given inputs. Here is github gist of my notebook

As we can see from the notebook for given input image I am getting correct segmentation map. In this notebook I am using fastai’s load() to load my pretrained model. I have used data.export() to export all configurations of my data in file called as export.pkl.

Now I want to deploy this as a web-app. I created flask-app for this deployment. I placed export.pkl file in my flask-app and pretrained model stage-1_224 in models folder. Here, since I need to make predictions on a single image, so I am creating an empty databunch using ImageDataBunch.load_empty() and passing fname=export1.pkl (export1.pkl is alias to export.pkl). Here is code snippet of my model in flask-app:

from fastai import *
from import *
# from utils import SegmentationProcessor,SegmentationLabelList,SegmentationItemList
import pathlib
import matplotlib.pyplot as plt
from PIL import Image

class UnetSeg():

    def __init__(self): = ImageDataBunch.load_empty(pathlib.PosixPath('.'),fname='export1.pkl') 

        self.learn = unet_learner(,models.resnet34,metrics=self.custom_acc,loss_func=self.custom_loss).load('stage_1_224',strict=False)

        # self.learn.load('stage-1_224_opt_f',strict=False)

    def custom_loss(self,y_hat,y):
        y = y.squeeze(1)
        loss = nn.CrossEntropyLoss(ignore_index=255)
        return loss(y_hat,y)
        # data.c = 1    

    def custom_acc(self,input,targs):
        targs = targs.squeeze(dim=1)
        input = input.argmax(dim=1)
        return (input.flatten()==targs.flatten()).float().mean()

    def get_seg(self,img_path):
        img = open_image(img_path)
        pred = self.learn.predict(img)
        return pred

u1 = UnetSeg()

print("Empty databunch initialization done....\n\n",

print("Enter the path of input image = \n")
path = input()
y = u1.get_seg(path)

print("Plotting input image = \n\n")
# plt1 = plt.figure()
# plt.imshow(

print("Plotting predicted seg_mask = \n\n")
temp = np.array(y[0].data)
temp  = temp.squeeze()

After running this script, I get inconsistent predictions i.e. predictions from notebook doesn’t match with predictions from the above script.
Output of above script:

Why is there is mismatch in predictions? Are my pretrained weights not loading properly?