Skip to content

Latest commit

 

History

History
68 lines (50 loc) · 3.62 KB

model_zoo.md

File metadata and controls

68 lines (50 loc) · 3.62 KB
Raw

Pre-trained ICNet and PSPNet Models

This project provided pre-trained model weights for both ICNet and the PSPNet50 baseline model. Both Tensorflow checkpoints and frozen graphs are provided for each model variant, allowing you to either deploy or fine-tune on your own dataset.

Each downloadable archive will contain the following:

  • a graph proto (graph.pbtxt)
  • a checkpoint (model.ckpt.data-00000-of-00001, model.ckpt.index, model.ckpt.meta)
  • a frozen graph proto (frozen_inference_graph.pb) with folded batch norm parameters and weights included in the graph as constants. To be used for out of the box inference
  • a config file (pipeline.config) which was used to generate the graph.
  • a text file with per-layer benchmarking information (benchmark.txt)

Cityscapes-trained models

The table below contains a download link as well as information about each model. All models were trained with the Cityscapes training set with fine data only (as opposed to fine and coarse data). The listed information is:

  • mean Intersection over Union (mIoU) on the Cityscapes validation set ‡
  • compression factor indicating the percentage of filters removed during pruning
  • number of parameters in the model
  • estimate of floating point operations per second
Model Val. Set mIoU Compress. Factor Parameters FLOPs
0818_pspnet50_1.0_713_resnet_v1 73.8% 1.0 46.53M 2942.59B
0818_icnet_1.0_1025_resnet_v1 69.8% 1.0 26.82M 246.95B
0818_icnet_0.5_1025_resnet_v1 65.8% 0.5 6.71M 63.20B

‡ The mIoU here is based on whole-image inference. This means that evaluation is done with one forward pass of the network. In the case of Cityscapes, that means the 1024x2048 resolution images are evaluated without cropping and without evaluating subdivided patches (sliding evaluation).

Validating Pre-trained Models

Listed here is an example of checking the accuracy and inference results of a model downloaded from the above list. We will use the PSPNet50 as an example. Start by downlaoding and then extracting the model files with

# Download using Google Drive link provided then extract
tar -zxvf 0818_pspnet_1.0_713_resnet_v1.tar.gz

You will have multiple formats of the model which are listed above. Edit the configuration file pipeline.config with the paths to your Cityscapes train and val TFRecords. Then, to evaluate the checkpoint run

python eval.py \
	--train_dir 0818_pspnet_1.0_713_resnet_v1 \
	--eval_dir 0818_pspnet_1.0_713_resnet_v1 \
	--config_path 0818_pspnet_1.0_713_resnet_v1/pipeline.config \
	--evaluate_all_from_checkpoint model.ckpt \
	--verbose

The resulting accuracy measure in will be almost identical if you were to run the official Cityscapes per-pixel evaluation script.

To use the model out the box for inference, we must specify either a directory with images or a single image path. We use the inference script like so for Cityscapes size images

python inference.py \
	--input_shape 1024,2048,3 \
	--pad_to_shape 1025,2049 \
	--input_path /tmp/DIR_WITH_IMAGES_OR_IMAGE_PATH \
	--config_path 0818_pspnet_1.0_713_resnet_v1/pipeline.config \
	 --trained_checkpoint 0818_pspnet_1.0_713_resnet_v1/model.ckpt \
	--output_dir /tmp/RESULTS_DIR \
	--label_ids # set this if you need labelID outputs, else will be colour