tool-readme

tools/README.md

@@ -5,10 +5,28 @@
 * th_warp: threshold in the propagation process. Modifying the th_warp will lead to different results.
 If you do not satisfy the result, pls try to modify this argument.
 
-## Training the Refined Stage
+## Training Guidance
 To get the better results, you can use the stacked multi-scale(MS) DFC-Net.
 When training third stage, you'd better use the second stage's weights as initialization.
 
+For fixed-region(mid-bbox) inpainting, we give further guidance for your reference.
+
+1. For initial stage(stage 1). Firstly, you should generate the data list which is much more similar with the testing list.
+However, you should remove the output dir in the data list which is the second to last item. And then you can use the following command:
+```
+CUDA_VISIBLE_DEVICES=xxxxx python tools/train_initial.py --model_name stage1 --FIX_MASK --MASK_MODE mid-bbox --TRAIN_LIST path_to_train_datalist --DATA_ROOT path_to_input_flow --GT_FLOW_ROOT path_to_gt_flow --INITIAL_HOLE
+```
+2. For Refine stage(stage2 and stage3). For convenient and quick loading data, we first extract the flow by using stage1 network and then use 
+these results as the input of the next stage.
+```
+CUDA_VISIBLE_DEVICES=xxxxx python tools/train_refine.py --model_name stagex --FIX_MASK --MASK_MODE mid-bbox --TRAIN_LIST path_to_train_datalist --DATA_ROOT path_to_input_flow --GT_FLOW_ROOT path_to_gt_flow
+```
+Some tips in training refinement stage:
+* The image shape, result shape and (MASK_HEIGHT, MASK_WIDTH) should be set according to your need and MASK_HEIGHT = img_height // 4, MASK_WIDTH = img_width // 4. 
+In our experiment, we set `--IMAGE_SHAPE 320 600 --RES_SAHPE 320 600` in stage 2 and `--IMAGE_SHAPE 480 840 --RES_SAHPE 480 840` in stage 3.
+* We highly recommend you to keep the batch_size of 32. However, if you do not have enough gpu memory, you can decrease the learning rate and the weight of HFEM ('--LAMBDA_HARD').
+* If you cannot train the following stage with that resolution, you can pre-train the refinement stage with low resolution and large batch_size and then finetune in the large resolution.
+
 ## Image Inpainting Model
 The contextual attention module can be substituted with better implementation such as no-local module.
 We provide this version just for testing DeepFillv1 with the same result as the original Tensorflow version.