add code for feature distribution analysis

paper_figure/visual_feature_distribution/README.md

@@ -0,0 +1,81 @@
+# Feature Extraction and Distance Matrix Calculation
+
+This Python script is designed for performing feature extraction and distance matrix calculation for a given dataset of images. It supports various models for feature extraction, norm types for distance calculation, and allows customization of parameters through command-line arguments.
+
+## Table of Contents
+
+- [Prerequisites](#prerequisites)
+- [Usage](#usage)
+- [Customization](#customization)
+- [Folder Structure](#folder-structure)
+- [Output](#output)
+- [Acknowledgments](#acknowledgments)
+
+## Prerequisites
+
+Before using this script, ensure you have the following dependencies installed:
+
+- Python 3.x
+- PyTorch
+- torchvision
+- NumPy
+- PIL (Python Imaging Library)
+- tqdm
+
+You can typically install these dependencies using `pip`:
+
+```bash
+pip install torch torchvision numpy pillow tqdm
+```
+
+## Usage
+To use this script, follow these steps:
+
+1. Clone this repository or download the script to your local machine.
+    ```
+    git clone xxx
+    ```
+2. Navigate to the directory where the script is located:
+    ```
+    cd xxx
+    ```
+
+3. Run the script with the desired parameters. Here's the basic usage:
+    ```
+    python script_name.py --model [model] --image_type [image_type] --norm [norm_type] --csv [csv_file] --batch_size [batch_size] --data_root [data_root_directory]
+    ```
+
+    For example, 
+
+    ```
+    python main.py --model resnet18 --image_type texture --norm l2 --csv "CSV/256_20x.csv" 
+    ```
+4. Wait for the script to perform feature extraction and distance matrix calculation.
+5. The results, including the distance matrix, will be saved in the data folder in the current directory.
+
+
+## Customization
+
+* Model Selection (--model): You can choose from the following models for feature extraction:
+    - resnet18
+    - alexnet
+    - convnext
+    - vgg11
+    - vit
+    - dinov2
+* Image Type (--image_type): Specify the type of images in your dataset. Choose between "heightmap" and "texture."
+
+* Norm Type (--norm): Select the norm type for distance calculation. Currently, only "l2" (Euclidean distance) is supported.
+
+* CSV File (--csv): Provide the path to the CSV file containing image information.
+
+* Batch Size (--batch_size): Set the batch size for feature extraction. The default is 10.
+
+* Data Root Directory (--data_root): Specify the root directory of your dataset.
+
+## Output
+The script will generate the following output:
+
+- A distance matrix saved as a NumPy array in the data folder. The filename format is csv_filename_image_type_model_norm.npy.
+
+- Information about the image closest to the mean feature vector will be saved in a file called "closest_to_mean.txt."

paper_figure/visual_feature_distribution/closest_to_mean.txt

@@ -0,0 +1,48 @@
+256_20x	texture	resnet18	l2	44.933701	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/20x/texture/fern/TRIBO_0C1_FLINT_FERN_9h_20X_005_021.bmp
+256_20x	heightmap	resnet18	l2	44.933701	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/20x/texture/fern/TRIBO_0C1_FLINT_FERN_9h_20X_005_021.bmp
+256_50x	texture	resnet18	l2	54.821190	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/50x/texture/beforeuse/TRIBO_023_FLINT_BEFOREUSE_0h_50X_002_016.bmp
+256_50x	heightmap	resnet18	l2	54.821190	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/50x/texture/beforeuse/TRIBO_023_FLINT_BEFOREUSE_0h_50X_002_016.bmp
+512_20x	texture	resnet18	l2	44.804939	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/20x/texture/fern/TRIBO_0C1_FLINT_FERN_9h_20X_020_003.bmp
+512_20x	heightmap	resnet18	l2	44.804939	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/20x/texture/fern/TRIBO_0C1_FLINT_FERN_9h_20X_020_003.bmp
+512_50x	texture	resnet18	l2	51.226322	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/50x/texture/beforeuse/TRIBO_023_FLINT_BEFOREUSE_0h_50X_002_001.bmp
+512_50x	heightmap	resnet18	l2	51.226322	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/50x/texture/beforeuse/TRIBO_023_FLINT_BEFOREUSE_0h_50X_002_001.bmp
+865_20x	texture	resnet18	l2	45.037403	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/20x/texture/bone/TRIBO_008_FLINT_BONE_5h_20X_065_001.bmp
+865_20x	heightmap	resnet18	l2	45.037403	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/20x/texture/bone/TRIBO_008_FLINT_BONE_5h_20X_065_001.bmp
+865_50x	texture	resnet18	l2	56.155617	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/50x/texture/barley/TRIBO_0A2_FLINT_BARLEY_9h_50X_010_001.bmp
+865_50x	heightmap	resnet18	l2	56.155617	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/50x/texture/barley/TRIBO_0A2_FLINT_BARLEY_9h_50X_010_001.bmp
+256_20x	texture	vgg11	l2	24.920313	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/20x/texture/beforeuse/TRIBO_022_FLINT_BEFOREUSE_0h_20X_002_001.bmp
+256_20x	heightmap	vgg11	l2	24.920313	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/20x/texture/beforeuse/TRIBO_022_FLINT_BEFOREUSE_0h_20X_002_001.bmp
+256_50x	texture	vgg11	l2	28.182840	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/50x/texture/beforeuse/TRIBO_008_FLINT_BEFOREUSE_0h_50X_003_005.bmp
+256_50x	heightmap	vgg11	l2	28.182840	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/50x/texture/beforeuse/TRIBO_008_FLINT_BEFOREUSE_0h_50X_003_005.bmp
+512_20x	texture	vgg11	l2	21.526806	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/20x/texture/ivory/TRIBO_014_FLINT_IVORY_5h_20X_017_003.bmp
+512_20x	heightmap	vgg11	l2	21.526806	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/20x/texture/ivory/TRIBO_014_FLINT_IVORY_5h_20X_017_003.bmp
+512_50x	texture	vgg11	l2	25.522821	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/50x/texture/beforeuse/TRIBO_022_FLINT_BEFOREUSE_0h_50X_003_002.bmp
+512_50x	heightmap	vgg11	l2	25.522821	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/50x/texture/beforeuse/TRIBO_022_FLINT_BEFOREUSE_0h_50X_003_002.bmp
+865_20x	texture	vgg11	l2	22.694120	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/20x/texture/barley/TRIBO_0A1_FLINT_BARLEY_9h_20X_009_001.bmp
+865_20x	heightmap	vgg11	l2	22.694120	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/20x/texture/barley/TRIBO_0A1_FLINT_BARLEY_9h_20X_009_001.bmp
+865_50x	texture	vgg11	l2	27.138252	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/50x/texture/barley/TRIBO_0A2_FLINT_BARLEY_9h_50X_014_001.bmp
+865_50x	heightmap	vgg11	l2	27.138252	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/50x/texture/barley/TRIBO_0A2_FLINT_BARLEY_9h_50X_014_001.bmp
+256_20x	texture	convnext	l2	21.231182	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/20x/texture/beforeuse/TRIBO_012_FLINT_BEFOREUSE_0h_20X_001_021.bmp
+256_20x	heightmap	convnext	l2	21.231182	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/20x/texture/beforeuse/TRIBO_012_FLINT_BEFOREUSE_0h_20X_001_021.bmp
+256_50x	texture	convnext	l2	26.675819	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/50x/texture/beforeuse/TRIBO_012_FLINT_BEFOREUSE_0h_50X_003_002.bmp
+256_50x	heightmap	convnext	l2	26.675819	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/50x/texture/beforeuse/TRIBO_012_FLINT_BEFOREUSE_0h_50X_003_002.bmp
+512_20x	texture	convnext	l2	20.582735	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/20x/texture/bone/TRIBO_008_FLINT_BONE_5h_20X_004_002.bmp
+512_20x	heightmap	convnext	l2	20.582735	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/20x/texture/bone/TRIBO_008_FLINT_BONE_5h_20X_004_002.bmp
+512_50x	texture	convnext	l2	23.087261	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/50x/texture/beforeuse/TRIBO_008_FLINT_BEFOREUSE_0h_50X_002_003.bmp
+512_50x	heightmap	convnext	l2	23.087261	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/50x/texture/beforeuse/TRIBO_008_FLINT_BEFOREUSE_0h_50X_002_003.bmp
+865_20x	texture	convnext	l2	20.701147	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/20x/texture/beforeuse/TRIBO_016_FLINT_BEFOREUSE_0h_20X_002_001.bmp
+865_20x	heightmap	convnext	l2	20.701147	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/20x/texture/beforeuse/TRIBO_016_FLINT_BEFOREUSE_0h_20X_002_001.bmp
+865_50x	texture	convnext	l2	24.126169	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/50x/texture/beforeuse/TRIBO_004_FLINT_BEFOREUSE_0h_50X_001_001.bmp
+865_50x	heightmap	convnext	l2	24.126169	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/50x/texture/beforeuse/TRIBO_004_FLINT_BEFOREUSE_0h_50X_001_001.bmp
+256_20x	texture	dinov2	l2	20.723244	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/20x/texture/sprucewood/TRIBO_012_FLINT_SPRUCEWOOD_5h_20X_010_001.bmp
+256_20x	heightmap	dinov2	l2	20.723244	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/20x/texture/sprucewood/TRIBO_012_FLINT_SPRUCEWOOD_5h_20X_010_001.bmp
+256_50x	texture	dinov2	l2	26.677032	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/50x/texture/bone/TRIBO_020_FLINT_BONE_5h_50X_004_013.bmp
+256_50x	heightmap	dinov2	l2	26.677032	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/256/50x/texture/bone/TRIBO_020_FLINT_BONE_5h_50X_004_013.bmp
+512_20x	texture	dinov2	l2	21.638500	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/20x/texture/fern/TRIBO_0C1_FLINT_FERN_9h_20X_020_006.bmp
+512_20x	heightmap	dinov2	l2	21.638500	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/20x/texture/fern/TRIBO_0C1_FLINT_FERN_9h_20X_020_006.bmp
+512_50x	texture	dinov2	l2	27.832876	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/50x/texture/beforeuse/TRIBO_025_FLINT_BEFOREUSE_0h_50X_004_001.bmp
+512_50x	heightmap	dinov2	l2	27.832876	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/512/50x/texture/beforeuse/TRIBO_025_FLINT_BEFOREUSE_0h_50X_004_001.bmp
+865_20x	texture	dinov2	l2	21.984348	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/20x/texture/barley/TRIBO_0A2_FLINT_BARLEY_9h_20X_014_001.bmp
+865_20x	heightmap	dinov2	l2	21.984348	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/20x/texture/barley/TRIBO_0A2_FLINT_BARLEY_9h_20X_014_001.bmp
+865_50x	texture	dinov2	l2	28.085300	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/50x/texture/beforeuse/TRIBO_025_FLINT_BEFOREUSE_0h_50X_004_001.bmp
+865_50x	heightmap	dinov2	l2	28.085300	/mnt/SSD_SATA_0/DATASET/LUA_Dataset/865/50x/texture/beforeuse/TRIBO_025_FLINT_BEFOREUSE_0h_50X_004_001.bmp

paper_figure/visual_feature_distribution/main.py

@@ -0,0 +1,190 @@
+import os, sys
+import csv
+import numpy as np
+import torch
+from torchvision import transforms
+from torchvision.models import resnet50, ResNet50_Weights
+from PIL import Image
+from tqdm import tqdm
+import argparse
+from models import *
+
+
+def read_image_tensor(image_path):
+    image = Image.open(image_path)
+
+    transform = transforms.ToTensor()
+    tensor_image = transform(image)
+
+    return tensor_image
+
+def read_csv_file(file_path):
+    data = []
+    with open(file_path, mode='r', newline='') as file:
+        reader = csv.DictReader(file)
+        next(reader)  # Skip the header row
+        for row in reader:
+            data.append(row)
+    return data
+
+def csv_to_path(file_path, image_type="texture"):
+    path = []
+    # read csv file
+    csv_data = read_csv_file(file_path)
+    # check image type
+    assert image_type in ["texture", "heightmap"]
+    # use loop to generate path
+    RES = os.path.basename(file_path)[:-4].split("_")[0]
+    ZOOM = os.path.basename(file_path)[:-4].split("_")[1]
+    for row in csv_data:
+        path.append(os.path.join(DATA_FOLDER_ROOT, RES, ZOOM, image_type, row['Worked Material'].lower(), row['Image Name']))
+        assert os.path.exists(path[-1]), f"ERROR: CAN NOT LOCATED FILE: {path[-1]}"
+
+    return path
+
+def batch_feature_extractor(feature_extractor, preprocess, image_path_list, batch_size, device):
+
+    feature_tensor_all = []
+    print(f"\nBatch Feature Extraction")
+    for i in tqdm(range(0, len(image_path_list), batch_size)):
+        image_tensor_batch = []
+        for i in image_path_list[i:i+batch_size]:
+            image_tensor_batch.append(read_image_tensor(i))
+
+        image_tensor_batch = torch.stack(image_tensor_batch)
+        batch = preprocess(image_tensor_batch.to(device))
+        feature = feature_extractor(batch)
+
+        # feature_tensor_all.append(feature)
+        feature_tensor_all.append(feature.detach().cpu().numpy())
+
+
+    # feature_tensor_all = torch.cat(feature_tensor_all, dim=0)
+    feature_tensor_all = np.vstack(feature_tensor_all)
+    feature_tensor_all = feature_tensor_all.reshape(feature_tensor_all.shape[0], -1)
+
+
+    return feature_tensor_all
+
+
+def distance_matrix_gpu(feature_tensor, norm_type, device):
+
+    num_vectors = feature_tensor.shape[0]
+    distance_matrix = torch.zeros((num_vectors, num_vectors)).to(device)
+
+    if norm_type == "l2":
+        distance_matrix = torch.cdist(feature_tensor, feature_tensor, p=2)
+    elif norm_type == "l1":
+        distance_matrix = torch.cdist(feature_tensor, feature_tensor, p=1)
+    elif norm_type == "inf":
+        distance_matrix = torch.cdist(feature_tensor, feature_tensor, p=0)
+
+    return distance_matrix
+
+def find_cloest_to_mean(feature_array, image_path_list, csv_path):
+    # calculate the mean
+    mean_feature = np.mean(feature_tensor_all, axis=0)
+
+    # Calculate the Euclidean distance between each feature tensor and the mean
+    distances = np.linalg.norm(feature_array - mean_feature, axis=1)
+
+    # Find the index of the feature tensor that is closest to the mean
+    closest_idx = np.argmin(distances)
+
+    filename = image_path_list[closest_idx]
+
+    with open(f'closest_to_mean.txt', 'a') as f:
+        content = f"{os.path.basename(csv_path[:-4])}\t{IMAGE_TYPE}\t{MODEL}\t{NORM_TYPE}\t{distances[closest_idx]:4f}\t{filename}"
+
+        f.write(content + '\n')
+
+if __name__ == "__main__":
+
+    # change working dir to script location
+    os.chdir(sys.path[0])
+
+    print(f"\n##################################################")
+    print(f"ML Toolkits: {os.path.basename(os.getcwd())}")
+
+    parser = argparse.ArgumentParser(description="Feature extraction and distance matrix calculation")
+
+    parser.add_argument("--model", choices=["resnet18", "alexnet", "convnext", "vgg11", "vit", "dinov2"], default="resnet18", help="Choose the model (default: resnet18)")
+    parser.add_argument("--image_type", choices=["heightmap", "texture"], default="texture", help="Choose the image type (default: texture)")
+    parser.add_argument("--norm", choices=["l2"], default="l2", help="Choose the norm type (default: l2)")
+    parser.add_argument("--csv", required=True, help="Path to the CSV file containing image information")
+    parser.add_argument("--batch_size", type=int, default=10, help="Batch size for feature extraction (default: 10)")
+    parser.add_argument("--data_root", default="/mnt/SSD_SATA_0/DATASET/LUA_Dataset", help="Root folder of the dataset (LUA_Dataset)")
+
+    args = parser.parse_args()
+    MODEL = args.model
+    IMAGE_TYPE = args.image_type
+    NORM_TYPE = args.norm
+    CSV_PATH = args.csv
+    BATCH_SIZE = args.batch_size
+    DATA_FOLDER_ROOT = args.data_root
+
+    # Check for GPU availability
+    if torch.cuda.is_available():
+        device = torch.device("cuda")
+        print(f"Using GPU: {torch.cuda.is_available()}")
+    else:
+        device = torch.device("cpu")
+        print(f"Using GPU: {torch.cuda.is_available()}")
+
+    print(f"Model:\t\t{MODEL.upper()}")
+    print(f"Image Type:\t{IMAGE_TYPE.upper()}")
+    print(f'CSV FILE:\t{CSV_PATH}')
+    print(f"Batch Size:\t{BATCH_SIZE}\n")
+
+
+
+    # read csv file
+    CSV_PATH = os.path.join(DATA_FOLDER_ROOT, CSV_PATH)
+    image_path_list = csv_to_path(CSV_PATH)
+    # image_path_list = [os.path.join(DATA_FOLDER_ROOT, i) for i in image_path_list]
+    print(f"Number of Images: {len(image_path_list)}")
+
+    ################## CHANGE THIS PART FOR DIFFERENT MODEL ##################
+    # feature extractor initialization
+    if MODEL == "resnet18":
+        feature_extractor, preprocess = feature_extractor_resnet18()
+    elif MODEL == "alexnet":
+        feature_extractor, preprocess = feature_extractor_alexnet()
+    elif MODEL == "convnext":
+        feature_extractor, preprocess = feature_extractor_convnext()
+    elif MODEL == "vgg11":
+        feature_extractor, preprocess = feature_extractor_vgg11()
+    elif MODEL == "vit":
+        feature_extractor, preprocess = feature_extractor_vit()
+    elif MODEL == "dinov2":
+        feature_extractor, preprocess = feature_extractor_dinov2()
+    else:
+        print("ERROR: MODEL NOT SUPPORTED")
+
+    ##########################################################################
+
+    # perform batch feature extraction
+    with torch.no_grad():
+        feature_tensor_all = batch_feature_extractor(feature_extractor.to(device), preprocess, image_path_list, BATCH_SIZE, device)
+
+    # find the image that is cloest to the mean
+    find_cloest_to_mean(feature_tensor_all, image_path_list, CSV_PATH)
+
+    print(f"Feature Tensor: {feature_tensor_all.shape}")
+    # feature_tensor_all = feature_tensor_all.reshape(feature_tensor_all.size(0), -1)
+    feature_tensor_all = torch.from_numpy(feature_tensor_all).to(device)
+
+    # calculate distance matrix
+    distance_matrix = distance_matrix_gpu(feature_tensor_all, NORM_TYPE, device)
+    print(f"Distance Matrix: {distance_matrix.shape}")
+
+    # save distance matrix
+    filename = f"{os.path.basename(CSV_PATH[:-4])}_{IMAGE_TYPE}_{MODEL}_{NORM_TYPE}.npy"
+    print(f"\nSaving distance matrix to {filename}...")
+    np.save(f"data/{filename}", distance_matrix.cpu().numpy())
+
+    print("Done\n")
+
+
+
+