Predict with pre-trained models

This tutorial explains how to recognize objects in an image with a pre-trained model, and how to perform feature extraction.


To complete this tutorial, we need:

$ pip install matplotlib


We first download a pre-trained ResNet 18 model that is trained on the ImageNet dataset with over 1 million images and one thousand classes. A pre-trained model contains two parts, a json file containing the model definition and a binary file containing the parameters. In addition, there may be a synset.txt text file for the labels.

import mxnet as mx

Next, we load the downloaded model.

# set the context on CPU, switch to GPU if there is one available
ctx = mx.cpu()
sym, arg_params, aux_params = mx.model.load_checkpoint('resnet-18', 0)
mod = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
mod.bind(for_training=False, data_shapes=[('data', (1,3,224,224))], 
mod.set_params(arg_params, aux_params, allow_missing=True)
with open('synset.txt', 'r') as f:
    labels = [l.rstrip() for l in f]


We first define helper functions for downloading an image and performing the prediction:

%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np
# define a simple data batch
from collections import namedtuple
Batch = namedtuple('Batch', ['data'])

def get_image(url, show=False):
    # download and show the image. Remove query string from the file name.
    fname =, fname=url.split('/')[-1].split('?')[0])
    img = mx.image.imread(fname)
    if img is None:
        return None
    if show:
    # convert into format (batch, RGB, width, height)
    img = mx.image.imresize(img, 224, 224) # resize
    img = img.transpose((2, 0, 1)) # Channel first
    img = img.expand_dims(axis=0) # batchify
    img = img.astype('float32') # for gpu context
    return img

def predict(url):
    img = get_image(url, show=True)
    # compute the predict probabilities
    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('probability=%f, class=%s' %(prob[i], labels[i]))

Now, we can perform prediction with any downloadable URL:


probability=0.249607, class=n02119022 red fox, Vulpes vulpes

probability=0.172868, class=n02119789 kit fox, Vulpes macrotis


probability=0.873920, class=n02110958 pug, pug-dog

probability=0.102659, class=n02108422 bull mastiff

Feature extraction

By feature extraction, we mean presenting the input images by the output of an internal layer rather than the last softmax layer. These outputs, which can be viewed as the feature of the raw input image, can then be used by other applications such as object detection.

We can use the get_internals method to get all internal layers from a Symbol.

# list the last 10 layers
all_layers = sym.get_internals()

An often used layer for feature extraction is the one before the last fully connected layer. For ResNet, and also Inception, it is the flattened layer with name flatten0 which reshapes the 4-D convolutional layer output into 2-D for the fully connected layer. The following source code extracts a new Symbol which outputs the flattened layer and creates a model.

fe_sym = all_layers['flatten0_output']
fe_mod = mx.mod.Module(symbol=fe_sym, context=ctx, label_names=None)
fe_mod.bind(for_training=False, data_shapes=[('data', (1,3,224,224))])
fe_mod.set_params(arg_params, aux_params)

We can now invoke forward to obtain the features:

img = get_image('')
features = fe_mod.get_outputs()[0]
assert features.shape == (1, 512)