torch dnn batchnorm

batch.py

# colab - install latest
# !pip install opencv-python==4.1.0.25


import torch
import torch.nn as nn
import torchvision
import numpy as np
import cv2


def convnet(in_channels, out_channels=64, kernel_size=3):
    block = torch.nn.Sequential(
        torch.nn.Conv2d(kernel_size=kernel_size, in_channels=in_channels,
                        out_channels=out_channels, padding=1),
        # torch.nn.BatchNorm2d(out_channels),
        torch.nn.ReLU(),
        torch.nn.MaxPool2d(kernel_size=2)
    )
    return block

# input
torch.manual_seed(7973847)
x = torch.randn((1, 3, 28, 28), dtype=torch.float32)

# torch

tn = convnet(3)
y1 = tn.forward(x).detach().numpy()

outname = "min_proto"
torch.onnx.export(tn, x, outname+".onnx", verbose=False, input_names=["input1"], output_names=["output1"])

# opencv

on = cv2.dnn.readNet(outname+".onnx")
on.setInput(x.numpy())
y2 = on.forward()
print(on.getLayerNames())

print("diff", y1.shape[1]*y1.shape[2]*y1.shape[3], np.sum(np.abs(y1)), np.sum(np.abs(y2)), ":", np.sum(np.abs(y1 - y2)))

""" 
# colab
from google.colab.patches import cv2_imshow
def show(arr):
  r1 = cv2.resize(arr,(128,128))
  cv2_imshow(127+r1*64)
show(y1[0,1])
show(y2[0,1])
show(y1[0,1]-y2[0,1])
"""

out.txt

diff 12544 7313.5576 7313.5576 : 0.00060441886 # w/o bn

diff 12544 6494.3984 3735.4143 : 2794.75       # using bn