Skip to content

Latest commit

 

History

History
114 lines (87 loc) · 6.95 KB

README.md

File metadata and controls

114 lines (87 loc) · 6.95 KB
Raw

DARTS-ADP

We propose three networks for Computational Pathology (CPath) applications. The network architectures are searched on ADP using Differentiable Architecture Search (DARTS) and can be well transfered to other CPath datasets.

Probeable DARTS with Application to Computational Pathology,
Sheyang Tang, Mahdi S. Hosseini, Lina Chen, Sonal Varma, Corwyn Rowsell, Savvas Damaskinos, Konstantinos N. Plataniotis, Zhou Wang
In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV2021-CDPath)

Architectures

  • Overall architecture Network structure Above is the overall structure of the searched networks, where the normal and reduction cells are searched. A cell can be represented as a directed acyclic graph with nodes and edges. Each node is a feature map, and each edge belongs to one of the candidate operations, including 3x3 and 5x5 seperable convolutions, 3x3 and 5x5 dilated separable convolutions, 3x3 max pooling, 3x3 average pooling, and skip-connection.

We alter the number of nodes and search for the optimum architectures in each configuration. Here we present three best-performing architectures with different number of nodes.

  • DARTS_ADP_N2: each cell has 2 nodes. DARTS_ADP_N2

  • DARTS_ADP_N3: each cell has 3 nodes. DARTS_ADP_N3

  • DARTS_ADP_N4: each cell has 4 nodes. DARTS_ADP_N4

Datasets

We search the architectures on the ADP dataset and transfer them to three more datasets including BCSS, BACH, and Osteosarcoma.

  • ADP: a multi-label histological tissue type dataset. This is where the architectures are searched. More details can be found in the ADP Website and this paper.

  • BCSS: a multi-label breast cancer tissue dataset. More details can be found here.

  • BACH: a single-label breast cancer histology image dataset. More details can be found in this paper.

  • Osteosarcoma: contains osteosarcoma histology images and is available through this website.

Searching procedure

Overview The searching is based on DARTS. We improve the existing DARTS framework with a probing metric and an adaptive optimizer.

Probing metric

We apply stable rank to monitor the learning process of convolutional layers during searching, and show that the default DARTS lacks proper learning rate tuning.

These are experiments on CIFAR-100. Each column corresponds to a different initial learning rate.

knowledge_gain_weight_compare

With larger initial learning rate:

  • More layers generate higher stable rank, meaning they are learning better.
  • The preferance over skip-connection (a common issue when searched on CV datasets) is suppressed, meaning that the resulting network has more learnable parameters and thus performs better.

We show that the default DARTS (left column) lacks proper learning rate tuning. For more details, please refer to the paper.

Adaptive optimizer

We change the default SGD optimizer to Adas, an adaptive optimizer that automatically tunes the learning rates for each layer based on their stable rank evolution. Adas Adas helps reduce the gap between training and validation error during searching, leading to more generalizable architectures with higher test accuracy. For more details, please refer to the paper.

Performance

The searched networks are trained in 4 datasets and compared with multiple state-of-the-art networks. Results show their superioty in prediction accuracy and computation complexity. Performance

Usage

Pretrained models

Pretrained model weights are provided in the /pretrained folder, where four subfolders contains pretrained weights for three architectures on each dataset. The script test.py demonstrates how to load the pretrained weights of a network and test its performance.

First, download the dataset you want to train on and store them to a local directory.

Then, open test_demo.sh and edit the following:

  • Change the path of --data to where you store the downloaded data.
  • Change the name of --dataset accordingly. Valid names are ADP, BCSS, BACH, and OS.
  • Select the architecture you want to train. Valid architectures are DARTS_ADP_N2, DARTS_ADP_N3, and DARTS_ADP_N4.
  • Change --model_path according to the chosen dataset and architect name. E.g., ./pretrained/ADP/darts_adp_n4.pt if testing DARTS_ADP_N4 on ADP.

Now, simply run

cd path/to/this-repo
sh test_demo.sh

Training

First, download the dataset you want to train on and store them to a local directory.

Then, open train_demo.sh and edit the following:

  • Change the path of --data to where you store the downloaded data.
  • Change the name of --dataset correspondingly. Valid names are ADP, BCSS, BACH, and OS.
  • Select the architecture you want to train. Valid architectures are DARTS_ADP_N2, DARTS_ADP_N3, and DARTS_ADP_N4.
  • Other hyperparameters including learning rate, batch size and epochs, etc.

You can open train.py to see full details of hyperparameters.

Now, run the demo script to start training

cd path/to/this-repo
sh train_demo.sh