【修改说明】README已上传，bug待修理

README.md

@@ -1,20 +1,19 @@
 # HowTOs
 
-***善用加粗、倾斜***
-
 ### 📖 **RESIN: A self-supervised framework for enhancing axial resolution of volumetric imaging data**
 > [![arXiv](https://img.shields.io/badge/arXiv-Paper-<COLOR>.svg)](https://baidu.com)<br>
 > [Author](https://github.com)
 
-Based on BasicSR.
-
 ### 🚩 Updates
-* **2024.9.16**: XXX
+* **2024.9.23**: 【修改说明】代码整改完毕，README待上传，bug待修理
+* **2024.9.24**: 【修改说明】README已上传，bug待修理
+
 
 ## Web Demo and API
 
 [![Replicate](https://replicate.com/cjwbw/animesr/badge)]() 
-*coming soon*
+
+The web demo and API (Huggingface) is *coming soon*.
 
 ## Video Demos
 
@@ -50,39 +49,143 @@ Based on BasicSR.
     git checkout RESIN
     ```
 
+## 💻 Training
+
+### **Overview**
+
+The training pipeline contains 2 steps, Dataset Preparation and Training. The training params is in `./options`，you can create yourself's yml file for your training pipeline. More params' explanation can be found in [here](https://github.com/XPixelGroup/BasicSR-docs).
+
+### **Dataset Preparation**
+
+1. Get a big ROI regin from your data, the format of the chosen data should be h5/ims (or tif/tiff).
+2. Crop the above big single image to sub-images TIFFs, size 128×128, for example. Note that you can apply filtering rules such as filtering background or low-quality data. The purpsoe of this step is ensuring the cleanliness of the training dataset.
+
+    Then The scripts is as follows:
+    ```bash
+    # Get sub-images from a single tifffile
+    python scripts/get_subdatasets_from_bigtif.py --size=128 --percentiles_lower_bound=0.75 --percentiles_upper_bound=0.99 --minmax=450 --maxmin=60000 --input_folder="path" --output_folder="path" --input_tif_name="name" --crop_size=50 --rotated_flag=False
+
+    # Get sub-images from a single h5file
+    python scripts/get_subdatasets_from_ims.py --size=128 --percentiles_lower_bound=0 --percentiles_upper_bound=0.9999 --minmax=400 --maxmin=60000 --input_folder="path" --output_folder="path" --input_tif_name="name" --x_floor=0 --x_upper=8700 --y_floor=0 --y_upper=8400 --z_floor=0 --z_upper=900 --channel=1 --rotated_flag=False
+    ```
+
+    The py_script will generate 2 (or 3 if rotated_flag is True) datasets, as `frontground_datasets`, `background_datasets` (and `rotated_frontground_datatsets`). Note that the `background_datasets` is just for check. If your params setting is good, all dirty data should should be in the `background_datasets` folder, and all the `background_datasets` folder's data should be dirty.
+
+### **Training**
+
+Before the training, you should modify the [yaml file](options/RESIN.yml) accordingly. For example, you should modify the `datasets_cube` to your own anisotropic dataset. The training instruction is as follows:
+
+```bash
+# Using slurm
+CUDA_VISIBLE_DEVICES=0,1,2,3 srun\
+--ntasks-per-node=4 --mpi=pmi2 python sr_3dunet/train.py \
+-opt options/RESIN.yml --launcher="slurm" # --auto_resume
+```
+
 ## ⚡ **Inference**
 
-We provided 3 ways to inference images, include inference **tif files in a given folder_path**, **tif files in a given folder_path**, **tif files in a given folder_path**:
+RESIN support both TIFFs and ims/h5 as input for inference. We have provided corresponding scripts. Otherwise, we supply a debug inference script, for developers check training results quickly.
+
+You can execute them in following manner:
+
+**Inference TIFFs**
 
-讲清楚区别在于分开的原因在于norm方式
+The script supports batch-inference for TIFF files in the given specified path. If your images are too large to fit in the GPU memory (it is common in 3D inference), you can set the `--piece*` parameter to process the images in samll pieces.
+
+For some images with a slanted blur direction (like VISoR), if your model is not trained with slanted orientations, you can set the 'rotated_flag' parameter. The script will automatically inference in rotated manner.
 
 ```bash
-# inference tif files in folder
-bash inference_from_folder.sh
+# bash inference_from_folderTIFFs.sh
+CUDA_VISIBLE_DEVICES=0 python scripts/inference_from_folder.py \
+    -i /share/home/wangwb/workspace/sr_3dunet/datasets/rm009_labeled/bg \
+    -o /share/home/wangwb/workspace/sr_3dunet/datasets/rm009_labeled/bg_sr \
+    --model_path /share/home/wangwb/workspace/sr_3dunet/weights/MPCN_VISoR_oldbaseline_256_net_g_A_140000.pth \
+    --piece_flag True --piece_size 128 --piece_overlap 16 --rotated_flag False
+```
 
-# inference a single h5 file
-bash inference_from_h5.sh
+```console
+Usage:
+    -i, --input:        The folder path containing your test TIFF files.
+    -o, --output:       The folder path for the model's output TIFF files; this folder may not already exist.
+    --model_path:       The path of your Restoration model file (.pth).
+    --piece_flag:       Set to True if you wants to processing large TIFF files by splitting them into smaller segments before inference.
+    --piece_size:       Applicable when "--piece_flag" is enabled, defines the dimensions of the smaller TIFF segments.
+    --piece_overlap:    Applicable when "--piece_flag" is enabled, indicates the overlap area between adjacent smaller TIFF segments.
+    --rotated_flag:     Set to True if your model expects horizontal data but the test data contains oblique angles (e.g., in VISoR).
 ```
 
-We also supply a simple scripts for you to quickly inference your 
+**Inference ims/h5**
 
-## 💻 Training
+Different from TIFFs-Inference script, the Ims/H5-Inference scripts only supports single file inference and no rotated data, cause the Ims/H5 files are usually single and big. If you wants to do batch inference, you can write inference scripts yourself.
 
-### **Dataset Preparation**
+```bash
+# bash inference_from_h5.sh
+CUDA_VISIBLE_DEVICES=0 python scripts/inference_from_h5.py \
+    -i /share/data/VISoR_Reconstruction/SIAT_ION/LiuCiRong/20230910_CJ004/CJ4-1um-ROI1/CJ4ROI2.ims \
+    -o datasets/NISSL/CJ4ROI2_out_h5_piece64/output_res0.h5 \
+    --model_path weights/MPCN_VISoR_NISSL_net_g_A_110000.pth \
+    --piece_size 64 --piece_overlap 16
+```
 
-1. Get a big ROI regin from your data, the format of your data should be h5/ims (or tif/tiff).
-2. Crop the above big single image to sub-images, 128×128, for example. Note that you can apply filtering rules (using args arguments or write yourself's code) when selecting data subsets, as ensuring the cleanliness of the dataset is essential for successful training.
-    
-    The scripts is as follows:
-    ```bash
-    # Get sub-images from a single tifffile
-    python scripts/get_subdatasets_from_bigtif.py --size=128 --percentiles_lower_bound=0.75 --percentiles_upper_bound=0.99 --minmax=450 --maxmin=60000 --input_folder="path" --output_folder="path" --input_tif_name="name" --crop_size=50
-    # Get sub-images from a single h5file
-    python scripts/get_subdatasets_from_ims.py --size=128 --percentiles_lower_bound=0 --percentiles_upper_bound=0.9999 --minmax=400 --maxmin=60000 --input_folder="path" --output_folder="path" --input_tif_name="name" --x_floor=0 --x_upper=8700 --y_floor=0 --y_upper=8400 --z_floor=0 --z_upper=900 --channel=1
-    ```
-    If the direction of anisotropy is not axial but 45° orientation, add `--rotated_flag` at the end of the command. The py_script will generate 2 (or 3 if rotated_flag is True) datasets, as `frontground_datasets`, `background_datasets` (and `rotated_frontground_datatsets`). Note that the `background_datasets` is just for check.
+```console
+Usage:
+    -i, --input:        Your test h5/ims file's path.
+    -o, --output:       Your model\'s output h5/ims file's path.
+    --model_path:       The path of your Restoration model file (.pth).
+    --piece_size:       Defines the dimensions of the smaller image segments.
+    --piece_overlap:    Indicates the overlap area between adjacent smaller image segments.
+```
+
+**Inference for debug**
+
+This scripts is written for developers and users to check training results quickly, so this scripts's parameters are more than above. It will create a new directory, and put input/output/reconstruct data in it. The `--piece*` parameters is setting as above. Otherwise, you can set `--remove_size` params to cut the specified most edge pixels, cause the `conv` operation is usually performs bad in image's edge, which obstructed practical visualization.
 
-### **Start Training**
-Let's take the de-anisotropy pipeline with VISoR's NISSL datasets for example.
+```bash
+# bash inference_debug.sh
+branch_names=("RESIN_concat_40X")
+iters=("110000")
+
+for branch_name in "${branch_names[@]}"
+do
+    for iter in "${iters[@]}"
+    do 
+        CUDA_VISIBLE_DEVICES=0 python scripts/inference_debug.py \
+            -i /share/home/wangwb/workspace/sr_3dunet/datasets/40X/val_datasets\
+            -o /share/home/wangwb/workspace/sr_3dunet/results/${branch_name}_net_g_${iter}\
+            --model_path /share/home/wangwb/workspace/sr_3dunet/experiments/${branch_name}/models/net_g_A_${iter}.pth\
+            --model_back_path /share/home/wangwb/workspace/sr_3dunet/experiments/${branch_name}/models/net_g_B_${iter}.pth\
+            --piece_flag True --piece_size 64 --piece_overlap 16 --remove_size 4 --rotated_flag True
+    done
+done
+```
+
+```console
+Usage:
+    -i, --input:        The folder path containing your test TIFF files.
+    -o, --output:       The folder path for the model's output TIFF files; this folder may not already exist.
+    --model_path:       The path of your Restoration model file (.pth).
+    --model_back_path:  The path of the simulated degradation model file obtained through cycle-consistency training, which aids in visualizing more training results.
+    --piece_flag:       Set to True if you wants to processing large TIFF files by splitting them into smaller segments before inference.
+    --piece_size:       Applicable when "--piece_flag" is enabled, defines the dimensions of the smaller TIFF segments.
+    --piece_overlap:    Applicable when "--piece_flag" is enabled, indicates the overlap area between adjacent smaller TIFF segments.
+    --remove_size:      The number of pixels to be trimmed from the outermost edge.
+    --rotated_flag:     Set to True if your model expects horizontal data but the test data contains oblique angles (e.g., in VISoR).
+```
+After run the above command, you will get the de-anisotropy images in specified output path.
+
+## Acknowledgement
+This project is build based on BasicSR.
+
+##  Citation
+If you find this project useful for your research, please consider citing our paper:
+```bibtex
+@InProceedings{wang2024RESIN,
+  author={XXX},
+  title={RESIN: A self-supervised framework for enhancing axial resolution of volumetric imaging data},
+  booktitle={XXX},
+  year={2024}
+}
+```
 
-支持neuron，nissl，dapi等
+## 📧 Contact
+If you have any question, please email `[email protected]`.

inference_debug.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-# -i: The folder path containing your test TIFF files.
-# -o: The folder path for the model's output TIFF files; this folder may not already exist.
+# -i, --input: The folder path containing your test TIFF files.
+# -o, --output: The folder path for the model's output TIFF files; this folder may not already exist.
 # --model_path: The path of your Restoration model file (.pth).
 # --model_back_path: The path of the simulated degradation model file obtained through cycle-consistency training, which aids in visualizing more training results.
 # --piece_flag: Set to True if you wants to processing large TIFF files by splitting them into smaller segments before inference.

inference_from_folderTIFFs.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-# -i: The folder path containing your test TIFF files.
-# -o: The folder path for the model's output TIFF files; this folder may not already exist.
+# -i, --input: The folder path containing your test TIFF files.
+# -o, --output: The folder path for the model's output TIFF files; this folder may not already exist.
 # --model_path: The path of your Restoration model file (.pth).
 # --piece_flag: Set to True if you wants to processing large TIFF files by splitting them into smaller segments before inference.
 # --piece_size: Applicable when "--piece_flag" is enabled, defines the dimensions of the smaller TIFF segments.

inference_from_h5.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-# -i: Your test h5/ims file's path.
-# -o: Your model\'s output h5/ims file's path.
+# -i, --input: Your test h5/ims file's path.
+# -o, --output: Your model\'s output h5/ims file's path.
 # --model_path: The path of your Restoration model file (.pth).
 # --piece_size: Defines the dimensions of the smaller image segments.
 # --piece_overlap: Indicates the overlap area between adjacent smaller image segments.

options/RESIN.yml

@@ -4,11 +4,6 @@ model_type: RESIN_Model
 num_gpu: auto  # set num_gpu: 0 for cpu mode
 manual_seed: 0
 
-# # USM the ground-truth
-# l1_gt_usm: False
-# percep_gt_usm: False
-# gan_gt_usm: False
-
 # dataset and data loader settings
 datasets:
   train:

options/RESIN_rotated.yml

@@ -4,11 +4,6 @@ model_type: RESIN_Aux_Model
 num_gpu: auto  # set num_gpu: 0 for cpu mode
 manual_seed: 0
 
-# # USM the ground-truth
-# l1_gt_usm: False
-# percep_gt_usm: False
-# gan_gt_usm: False
-
 # dataset and data loader settings
 datasets:
   train:

scripts/get_subdatasets_from_bigtif.py

@@ -5,7 +5,7 @@
 import math
 import argparse
 
-from sr_3dunet.utils.data_utils import  get_rotated_img
+from sr_3dunet.utils.data_utils import  get_rotated_img, str2bool
 
 def img_loader(img, start_x, start_y, start_z, size):
     return img[start_x:start_x+size, start_y:start_y+size, start_z:start_z+size]
@@ -35,7 +35,7 @@ def judge_img(img, percentiles, minmax, maxmin):
     parser.add_argument('--output_folder', type=str, default="/share/home/wangwb/workspace/sr_3dunet/datasets/40X")
     parser.add_argument('--input_tif_name ', type=str, default="40x20x030p26sp1.tif")
     parser.add_argument('--crop_size ', type=int, default=50)
-    parser.add_argument("--rotated_flag", action= "store_true")
+    parser.add_argument('--rotated_flag', type=str2bool, default=False)
 
     args = parser.parse_args()
     args.stride = args.size//2

scripts/get_subdatasets_from_ims.py

@@ -6,7 +6,7 @@
 import math
 import argparse
 
-from sr_3dunet.utils.data_utils import get_rotated_img
+from sr_3dunet.utils.data_utils import get_rotated_img, str2bool
 
 def img_loader(img, start_x, start_y, start_z, size):
     return img[start_x:start_x+size, start_y:start_y+size, start_z:start_z+size]
@@ -34,7 +34,7 @@ def judge_img(img, percentiles, minmax):
     parser.add_argument('--input_folder', type=str, default="/share/data/VISoR_Reconstruction/SIAT_ION/LiuCiRong/20230910_CJ004/CJ4-1um-ROI1")
     parser.add_argument('--output_folder', type=str, default="/share/home/wangwb/workspace/sr_3dunet/datasets/NISSL")
     parser.add_argument('--input_ims_name ', type=str, default="CJ4ROI1.ims")
-    parser.add_argument("--rotated_flag", action= "store_true")
+    parser.add_argument('--rotated_flag', type=str2bool, default=False)
 
     parser.add_argument('--x_floor', type=int, default=0)
     parser.add_argument('--x_upper', type=int, default=8700)

train.sh

@@ -11,4 +11,4 @@ source activate RESIN
 
 CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 srun\
  --ntasks-per-node=8 --mpi=pmi2 python sr_3dunet/train.py \
- -opt options/MPCN_simulation.yml --launcher="slurm" # --auto_resume
+ -opt options/RESIN.yml --launcher="slurm" # --auto_resume