Contents

• CPCNN-Net Description
• Model Architecture
• Dataset
• Environment Requirements
• Code Description
  • File List
  • Code Parameters
    • Parameters Configuration
  • Training Process
    • Training
    • Training Result
  • Evaluation Process
    • Evaluation
    • Evaluation Result
• Description of Random Situation

CPCNN-Net Description

Paper: Cross-Part Learning for Fine-Grained Image Classification Authors: Man Liu, Chunjie Zhang, Huihui Bai, Riquan Zhang, Yao Zhao. Published in: IEEE Transactions on Image Processing ( Volume: 31) Page(s): 748 - 758 Date of Publication: 20 December 2021 DOI: 10.1109/TIP.2021.3135477

Model Architecture

In order to improve the performance of fine-grained image classification (FGIC) without boundary box or dense component labeling, this paper proposes a weakly supervised cross-component convolution neural network model (CP-CNN). Firstly, the proposed feature enhancement module (FEB) improves the positioning ability of the part proposal generator (PPG); After locating the parts, realize the joint feature learning between components through the context transformer; Finally, Pyramid Context Integration Block (PCIB) is used to integrate the fused features.

Dataset

Note that you can run the scripts based on the dataset mentioned in original paper or widely used in relevant domain/network architecture. In the following sections, we will introduce how to run the scripts using the related dataset below.

Dataset used: Caltech-UCSD Birds-200-2011

Please download the datasets [CUB_200_2011.tgz] and unzip it, then put all training images into a directory named "train", put all testing images into a directory named "test".

The directory structure is as follows:

.
└─cub_200_2011
  ├─train
  └─test

Environment Requirements

• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Code Description

Many thanks for NTS-NET.A part of the code is borrowed from it.

File List

.
└─ntsnet
  ├─README.md                             # README
  ├─src
    ├─config.py                           # network configuration
    ├─dataset.py                          # dataset utils
    ├─lr_generator.py                     # leanring rate generator
    ├─network.py                          # network define for ntsnet
    └─my_resnet.py                        # resnet.py
  ├─eval.py                               # evaluation scripts
  └─train.py                              # training scripts

Code Parameters

Parameters Configuration

"img_width": 448,           # width of the input images
"img_height": 448,          # height of the input images


# LR
"base_lr": 0.001,                                                              # base learning rate
"num_epochs": 200,                                                             # total epoch in lr generator
"momentum": 0.9,                                                               # sgd momentum in optimizer

# train
"topK": 6,                                                                     # topK parts
"batch_size": 8,
"num_classes": 200，
"learning_rate": 0.001,
"weight_decay": 1e-4,
"epoch_size": 200,                                                             # total epoch size

# checkpoint
"save_checkpoint": True,                                                       # whether save checkpoint or not
"save_checkpoint_epochs": 5,                                                   # save checkpoint interval
"keep_checkpoint_max": 10,
"num_train_images": 5994,                                                      # train images number in dataset
"num_test_images": 5794,                                                       # test images number in dataset

Training Process

• Set options in config.py, including learning rate, output filename and network hyperparameters.

Training

• Run train.py for training of CPCNN-Net model.

• Notes

1. As for PRETRAINED_MODEL，it should be a trained ResNet50 checkpoint.

Training Result

Training result will be stored in train_url path. You can find checkpoint file together with result. The directory structure is as follows:

.
└─cache
  ├─ckpt_0
    ├─cpcnnnet-101_100.ckpt
  ├─summary_dir
    ├─events.out.events.summary.1676516343.0.hostname_lineage
    ├─events.out.events.summary.1676516343.0.hostname_MS

After training, start MindInsight to visualize the collected data.For more information, please check the resources MindSpore summary

Evaluation Process

Evaluation

• Run eval.py for evaluation.

Evaluation Result

Inference result will be stored in the train_url path. Under this, you can find result like the following in eval.log.

ckpt file name: cpcnnnet-180_749.ckpt
accuracy:  0.916 

We provide the model parameters trained on CUB_200_2011. You can download it here

Description of Random Situation

We use random seed in train.py and eval.py for weight initialization.