This repository contains code for the deepLab-ResNet architecture used in my paper "SketchParse: Towards rich descriptions for poorly drawn sketches using multi-task hierarchical deep networks" published at ACM MM 2017. This architecture calculates losses on input images over multiple scales ( 1x, 0.75x, 0.5x ). Losses are calculated individually over these 3 scales. In addition to these 3 losses, one more loss is calculated after merging the output score maps on the 3 scales. These 4 losses are added to calculate the total loss.
24 June 2017
- Now, weights over the 3 scales ( 1x, 0.75x, 0.5x ) are shared as in the caffe implementation. Previously, each of the 3 scales had seperate weights. Results are almost same after making this change (more in the results section). However, the size of the trained .pth model has reduced significantly. Memory occupied on GPU(11.9 GB) and time taken (~3.5 hours) during training are same as before. Links to corresponding .pth files have been updated.
- Custom data can be used to train pytorch-deeplab-resnet using train.py, flag --NoLabels (total number of labels in training data) has been added to train.py and evalpyt.py for this purpose. Please note that labels should be denoted by contiguous values (starting from 0) in the ground truth images. For eg. if there are 7 (no_labels) different labels, then each ground truth image must have these labels as 0,1,2,3,...6 (no_labels-1).
Note that this repository has been tested with python 2.7 only.
Step 1: Convert init.caffemodel to a .pth file: init.caffemodel contains MS COCO trained weights. We use these weights as initilization for all but the final layer of our model. For the last layer, we use random gaussian with a standard deviation of 0.01 as the initialization.
To convert init.caffemodel to a .pth file, run (or download the converted .pth here)
python init_net_surgery.py
To run init_net_surgery .py, deeplab v2 caffe and pytorch (python 2.7) are required.
Step 2: Now that we have our initialization, we can train deeplab-resnet by running,
python train.py
To get a description of each command-line arguments, run
python train.py -h
To run train.py, pytorch (python 2.7) is required.
By default, snapshots are saved in every 1000 iterations in the data/snapshots. The following features have been implemented in this repository -
- Training regime is the same as that of the caffe implementation - SGD with momentum is used, along with the
polylr decay policy. A weight decay has been used. The last layer has10times the learning rate of other layers. - The iter_size parameter of caffe has been implemented, effectively increasing the batch_size to batch_size times iter_size
- Random flipping and random scaling of input has been used as data augmentation. The caffe implementation uses 4 fixed scales (0.5,0.75,1,1.25,1.5) while in the pytorch implementation, for each iteration scale is randomly picked in the range - [0.5,1.3].
- The boundary label (255 in ground truth labels) has not been ignored in the loss function in the current version, instead it has been merged with the background. The ignore_label caffe parameter would be implemented in the future versions. Post processing using CRF has not been implemented.
- Batchnorm parameters are kept fixed during training. Also, caffe setting
use_global_stats = Trueis reproduced during training. Running mean and variance are not calculated during training.
When run on a Nvidia Titan X GPU, train.py occupies about 11.9 GB of memory.
Evaluation of the saved models can be done by running
python evalpyt.py
To get a description of each command-line arguments, run
python evalpyt.py -h
When trained on VOC augmented training set (with 10582 images) using MS COCO pretrained initialization in pytorch, we get a validation performance of 72.40%(evalpyt2.py, on VOC). The corresponding .pth file can be downloaded here. This is in comparision to 75.54% that is acheived by using train_iter_20000.caffemodel released by authors, which can be replicated by running this file . The .pth model converted from .caffemodel using the first section also gives 75.54% mean IOU.
A previous version of this file reported mean IOU of 78.48% on the pytorch trained model which is caclulated in a different way (evalpyt.py, Mean IOU is calculated for each image and these values are averaged together. This way of calculating mean IOU is different than the one used by authors).
To replicate this performance, run
train.py --lr 0.00025 --wtDecay 0.0005 --maxIter 20000 --GTpath <train gt images path here> --IMpath <train images path here> --LISTpath data/list/train_aug.txt
The model presented in the results section was trained using the augmented VOC train set which was released by this paper. You may download this augmented data directly from here.
Note that this code can be used to train pytorch-deeplab-resnet model for other datasets also.
A part of the code has been borrowed from https://github.com/ry/tensorflow-resnet.