Dear @kitstar,
As previously stated I have used above-mentioned sequence of commands for converting the ResNet152-11k from MXNet to PyTorch via your fantastic repository. I hope my conversion commands is correct. My used sequence of commands are as follows:
1- python -m mmdnn.conversion._script.convertToIR -f mxnet -n resnet-152-symbol.json -w resnet-152-0000.params -d resnet152 --inputShape 3 224 224
2- python -m mmdnn.conversion._script.IRToCode -f pytorch --IRModelPath resnet152.pb --dstModelPath kit_imagenet.py --IRWeightPath resnet152.npy -dw kit_pytorch.npy
3- python -m mmdnn.conversion.examples.pytorch.imagenet_test --dump resnet152Full.pth -n kit_imagenet.py -w kit_pytorch.npy

Now, I want to compare the results of the original MXNet Model (i.e., resnet-152-symbol.json & resnet-152-0000.params) vs the converted PyTorch one (i.e., resnet152Full.pth model) visually.
This is my used script on MXNet:

import mxnet as mx
import matplotlib.pyplot as plt
import cv2
import numpy as np
from collections import namedtuple


# Load & Download the ResNet152 trained of the Full ImageNet:

# path='http://data.mxnet.io/models/imagenet-11k/'
# [mx.test_utils.download(path+'resnet-152/resnet-152-symbol.json'),
#  mx.test_utils.download(path+'resnet-152/resnet-152-0000.params'),
#  mx.test_utils.download(path+'synset.txt')]


# Load the ResNet152 trained of the Full ImageNet:

sym, arg_params, aux_params = mx.model.load_checkpoint('resnet-152', 0)
mod = mx.mod.Module(symbol=sym, context=mx.gpu(), label_names=None)
mod.bind(for_training=False, data_shapes=[('data', (1,3,224,224))],
         label_shapes=mod._label_shapes)
mod.set_params(arg_params, aux_params, allow_missing=True)
with open('synset.txt', 'r') as f:
    labels = [l.rstrip() for l in f]


# ** Predicting:

# define a simple data batch
Batch = namedtuple('Batch', ['data'])

def get_image(imageAddress, show=True):
    # load and show the image
    img = cv2.cvtColor(cv2.imread(imageAddress), cv2.COLOR_BGR2RGB)
    if img is None:
         return None
    if show:
         plt.imshow(img)
         plt.axis('off')
         plt.show()
    # convert into format (batch, RGB, width, height)
    img = cv2.resize(img, (224, 224))
    img = np.swapaxes(img, 0, 2)
    img = np.swapaxes(img, 1, 2)
    img = img[np.newaxis, :]
    return img

def predict(imageAddress):
    img = get_image(imageAddress, show=True)
    # compute the predict probabilities
    mod.forward(Batch([mx.nd.array(img)]))
    prob = mod.get_outputs()[0].asnumpy()
    # print the top-5
    prob = np.squeeze(prob)
    a = np.argsort(prob)[::-1]
    for i in a[0:5]:
        print(i)
        print('probability=%f, class=%s' %(prob[i], labels[i]))


# *** Test Model:
sampleInputImage = 'seagull.jpg'
predict(sampleInputImage)

And this is the produced Results:

1278
probability=0.536169, class=n02041246 gull, seagull, sea gull
1277
probability=0.104755, class=n02041085 larid
282
probability=0.098345, class=n01517966 carinate, carinate bird, flying bird
1280
probability=0.081313, class=n02041875 black-backed gull, great black-backed gull, cob, Larus marinus
1281
probability=0.051755, class=n02042046 herring gull, Larus argentatus

And this my used PyTorch script:

import torch
from torch.autograd import Variable
import torchvision.transforms as transforms
from torch.nn import functional as F
from PIL import Image
import matplotlib.pyplot as plt
from scipy.misc import imread

# Parameter:
num_predictions = 5


# Load Model
bestModelAddress = 'resnet152Full.pth' # for loading models
model = torch.load(bestModelAddress).cuda()
model.eval()


# Load One Input Image
test_image_address = 'seagull.jpg'

# ***** Show the original image
img = imread(test_image_address)
plt.subplot(1, 1, 1)
plt.axis('off') # turn off axis

# ***** Image pre-processing transforms
normalize = transforms.Normalize(
    mean=[0.485, 0.456, 0.406],
    std=[0.229, 0.224, 0.225]
)
preprocess = transforms.Compose([
    transforms.Scale(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    normalize
])


img_pil = Image.open(test_image_address)
img_tensor = preprocess(img_pil)
img_tensor.unsqueeze_(0)
img_tensor = Variable(img_tensor).cuda()


# Load Full-Imagenet Synset:
# *** Note: You can download it from:
# http://data.mxnet.io/models/imagenet-11k/synset.txt
with open('synset.txt', 'r') as f:
    labels = [l.rstrip() for l in f]


# Make predictions
output = model(img_tensor) # size(1, 11,221)
max, argmax = output.data.squeeze().max(0)
class_id = argmax[0]
classname = labels[class_id]
# print(class_id)

# print the top-5
h_x = F.softmax(output).data.squeeze()
probs, idx = h_x.sort(0, True)

print('Result: ')
# output the prediction
for i in range(0, num_predictions):
    print('{:.3f} -> {}'.format(100 * probs[i], labels[idx[i]]))



print('The Image is a', classname)

StrLabel = 'The Image is a ' + classname

# ***** Show the result & Image:
ax = plt.subplot(1, 1, 1)
# str_confidence = str("{0:.2f}".format(class_probabilities_cpu[0][max_index]))
ax.set_title(StrLabel)
plt.imshow(img)

plt.show()

And this is the PyTorch produced Results:

0.013 -> n04960277 black, blackness, inkiness
0.012 -> n04960582 coal black, ebony, jet black, pitch black, sable, soot black
0.009 -> n13896217 crescent
0.009 -> n03728437 match, lucifer, friction match
0.009 -> n02846141 black

So It is obviously clear that the results are significantly different from each other. It is worth noting that I want to use the resnet152Full.pth model on PyTorch. And I also tested some other images, but the results are very different.
I hope that my conversion commands would be correct. If you can help me to tackle this issue I will be very appreciated.
Thank you.

Convert Model from MXNet to PyTorch #6

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions