Skip to content

Latest commit

 

History

History
87 lines (64 loc) · 4.01 KB

faq.md

File metadata and controls

87 lines (64 loc) · 4.01 KB
Raw

FAQ

Q1: What if I want to use other network backbones, such as ResNet [1], instead of only those provided ones (e.g., Xception)?

A: The users could modify the provided core/feature_extractor.py to support more network backbones.

Q2: What if I want to train the model on other datasets?

A: The users could modify the provided dataset/build_{cityscapes,voc2012}_data.py and dataset/segmentation_dataset.py to build their own dataset.

Q3: Where can I download the PASCAL VOC augmented training set?

A: The PASCAL VOC augmented training set is provided by Bharath Hariharan et al. [2] Please refer to their website for details and consider citing their paper if using the dataset.

Q4: Why the implementation does not include DenseCRF [3]?

A: We have not tried this. The interested users could take a look at Philipp Krähenbühl's website and paper for details.

Q5: What if I want to train the model and fine-tune the batch normalization parameters?

A: If given the limited resource at hand, we would suggest you simply fine-tune from our provided checkpoint whose batch-norm parameters have been trained (i.e., train with a smaller learning rate, set fine_tune_batch_norm = false, and employ longer training iterations since the learning rate is small). If you really would like to train by yourself, we would suggest

  1. Set output_stride = 16 or maybe even 32 (remember to change the flag atrous_rates accordingly, e.g., atrous_rates = [3, 6, 9] for output_stride = 32).

  2. Use as many GPUs as possible (change the flag num_clones in train.py) and set train_batch_size as large as possible.

  3. Adjust the train_crop_size in train.py. Maybe set it to be smaller, e.g., 513x513 (or even 321x321), so that you could use a larger batch size.

  4. Use a smaller network backbone, such as MobileNet-v2.

Q6: How can I train the model asynchronously?

A: In the train.py, the users could set num_replicas (number of machines for training) and num_ps_tasks (we usually set num_ps_tasks = num_replicas / 2). See slim.deployment.model_deploy for more details.

Q7: I could not reproduce the performance even with the provided checkpoints.

A: Please try running

# Run the simple test with Xception_65 as network backbone.
sh local_test.sh

or

# Run the simple test with MobileNet-v2 as network backbone.
sh local_test_mobilenetv2.sh

First, make sure you could reproduce the results with our provided setting. After that, you could start to make a new change one at a time to help debug.

Q8: What value of eval_crop_size should I use?

A: Our model uses whole-image inference, meaning that we need to set eval_crop_size equal to output_stride * k + 1, where k is an integer and set k so that the resulting eval_crop_size is slightly larger the largest image dimension in the dataset. For example, we have eval_crop_size = 513x513 for PASCAL dataset whose largest image dimension is 512. Similarly, we set eval_crop_size = 1025x2049 for Cityscapes images whose image dimension is all equal to 1024x2048.

Q9: Why multi-gpu training is slow?

A: Please try to use more threads to pre-process the inputs. For, example change num_readers = 4.

References

  1. Deep Residual Learning for Image Recognition
    Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
    [link], In CVPR, 2016.

  2. Semantic Contours from Inverse Detectors
    Bharath Hariharan, Pablo Arbelaez, Lubomir Bourdev, Subhransu Maji, Jitendra Malik
    [link], In ICCV, 2011.

  3. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials
    Philipp Krähenbühl, Vladlen Koltun
    [link], In NIPS, 2011.