This is an unofficial PyTorch implementation of DeepLab v2 [1] with a ResNet-101 backbone.

• COCO-Stuff dataset [2] and PASCAL VOC dataset [3] are supported.
• The official Caffe weights provided by the authors can be used without building the Caffe APIs.
• DeepLab v3/v3+ models with the identical backbone are also included (not tested).
• torch.hub is supported.

† Images and labels are pre-warped to square-shape 513x513
‡ Note for SPADE followers: The provided COCO-Stuff 164k weight has been kept intact since 2019/02/23.

Required Python packages are listed in the Anaconda configuration file configs/conda_env.yaml. Please modify the listed cudatoolkit=10.2 and python=3.6 as needed and run the following commands.

# Set up with Anaconda
conda env create -f configs/conda_env.yaml
conda activate deeplab-pytorch

• COCO-Stuff 10k/164k
• PASCAL VOC 2012

Caffemodels pre-trained on COCO and PASCAL VOC datasets are released by the DeepLab authors. In accordance with the papers [1,2], this repository uses the COCO-trained parameters as initial weights.

1. Run the follwing script to download the pre-trained caffemodels (1GB+).

$ bash scripts/setup_caffemodels.sh

2. Convert the caffemodels to pytorch compatibles. No need to build the Caffe API!

# Generate "deeplabv1_resnet101-coco.pth" from "init.caffemodel"
$ python convert.py --dataset coco
# Generate "deeplabv2_resnet101_msc-vocaug.pth" from "train2_iter_20000.caffemodel"
$ python convert.py --dataset voc12

To train DeepLab v2 on PASCAL VOC 2012:

python main.py train \
    --config-path configs/voc12.yaml

To evaluate the performance on a validation set:

python main.py test \
    --config-path configs/voc12.yaml \
    --model-path data/models/voc12/deeplabv2_resnet101_msc/train_aug/checkpoint_final.pth

Note: This command saves the predicted logit maps (.npy) and the scores (.json).

To re-evaluate with a CRF post-processing:

python main.py crf \
    --config-path configs/voc12.yaml

Execution of a series of the above scripts is equivalent to bash scripts/train_eval.sh.

To monitor a loss, run the following command in a separate terminal.

tensorboard --logdir data/logs

Please specify the appropriate configuration files for the other datasets.

Note: Although the label indices range from 0 to 181 in COCO-Stuff 10k/164k, only 171 classes are supervised.

Common settings:

• Model: DeepLab v2 with ResNet-101 backbone. Dilated rates of ASPP are (6, 12, 18, 24). Output stride is 8.
• GPU: All the GPUs visible to the process are used. Please specify the scope with CUDA_VISIBLE_DEVICES=.
• Multi-scale loss: Loss is defined as a sum of responses from multi-scale inputs (1x, 0.75x, 0.5x) and element-wise max across the scales. The unlabeled class is ignored in the loss computation.
• Gradient accumulation: The mini-batch of 10 samples is not processed at once due to the high occupancy of GPU memories. Instead, gradients of small batches of 5 samples are accumulated for 2 iterations, and weight updating is performed at the end (batch_size * iter_size = 10). GPU memory usage is approx. 11.2 GB with the default setting (tested on the single Titan X). You can reduce it with a small batch_size.
• Learning rate: Stochastic gradient descent (SGD) is used with momentum of 0.9 and initial learning rate of 2.5e-4. Polynomial learning rate decay is employed; the learning rate is multiplied by (1-iter/iter_max)**power at every 10 iterations.
• Monitoring: Moving average loss (average_loss in Caffe) can be monitored in TensorBoard.
• Preprocessing: Input images are randomly re-scaled by factors ranging from 0.5 to 1.5, padded if needed, and randomly cropped to 321x321.

Processed images and labels in COCO-Stuff 164k:

You can use the pre-trained models, the converted models, or your models.

To process a single image:

python demo.py single \
    --config-path configs/voc12.yaml \
    --model-path deeplabv2_resnet101_msc-vocaug-20000.pth \
    --image-path image.jpg

To run on a webcam:

python demo.py live \
    --config-path configs/voc12.yaml \
    --model-path deeplabv2_resnet101_msc-vocaug-20000.pth

To run a CRF post-processing, add --crf. To run on a CPU, add --cpu.

Model setup with two lines

import torch.hub
model = torch.hub.load("kazuto1011/deeplab-pytorch", "deeplabv2_resnet101", pretrained='cocostuff164k', n_classes=182)

• While the official code employs 1/16 bilinear interpolation (Interp layer) for downsampling a label for only 0.5x input, this codebase does for both 0.5x and 0.75x inputs with nearest interpolation (PIL.Image.resize, related issue).
• Bilinear interpolation on images and logits is performed with the align_corners=False.

This codebase only supports DeepLab v2 training which freezes batch normalization layers, although v3/v3+ protocols require training them. If training their parameters on multiple GPUs as well in your projects, please install the extra library below.

pip install torch-encoding

Batch normalization layers in a model are automatically switched in libs/models/resnet.py.

try:
    from encoding.nn import SyncBatchNorm
    _BATCH_NORM = SyncBatchNorm
except:
    _BATCH_NORM = nn.BatchNorm2d

1. L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, A. L. Yuille. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. IEEE TPAMI, 2018.
   Project / Code / arXiv paper

2. H. Caesar, J. Uijlings, V. Ferrari. COCO-Stuff: Thing and Stuff Classes in Context. In CVPR, 2018.
   Project / arXiv paper

3. M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, A. Zisserman. The PASCAL Visual Object Classes (VOC) Challenge. IJCV, 2010.
   Project / Paper