Merge pull request #7 from jangsoopark/v2.1.0

V2.1.0

README.md

@@ -41,8 +41,7 @@ The proposed model only consists of **sparsely connected layers** without any fu
 ## Training
 For training, this implementation fixes the random seed to `12321` for `reproducibility`.
 
-The experimental conditions are same as in the paper, except for `data augmentation` and `learning rate`. 
-The `learning rate` is initialized with `1e-3` and decreased by a factor of 0.1 **after 26 epochs**.
+The experimental conditions are same as in the paper, except for `data augmentation`.
 You can see the details in `src/model/_base.py` and `experiments/config/AConvNet-SOC.json`
 
 ### Data Augmentation
@@ -52,10 +51,9 @@ You can see the details in `src/model/_base.py` and `experiments/config/AConvNet
 
 - However, for SOC, this repository does not use random shifting tue to accuracy issue.
     - You can see the details in `src/data/generate_dataset.py` and `src/data/mstar.py`
-    - This implementation failed to achieve higher than 98% accuracy when using random sampling.
     - The implementation details for data augmentation is as: 
-        - Crop the center of 94 x 94 size image on 128 x 128 SAR image chip (49 patches per image chip).
-        - Extract 88 x 88 patches with stride 1 from 94 x 94 image.
+        - Crop the center of 94 x 94 size image on 100 x 100 SAR image chip (49 patches per image chip).
+        - Extract 88 x 88 patches with stride 1 from 94 x 94 image with random cropping.
 
 
 ## Experiments
@@ -141,14 +139,14 @@ MSTAR-PublicMixedTargets-CD1/MSTAR_PUBLIC_MIXED_TARGETS_CD1
     - Place the two directories (`train` and  `test`) to the `dataset/raw`.
 ```shell
 $ cd src/data 
-$ python3 generate_dataset.py --is_train=True --use_phase=True --chip_size=94 --dataset=soc 
-$ python3 generate_dataset.py --is_train=False --use_phase=True --dataset=soc
+$ python3 generate_dataset.py --is_train=True --use_phase=True --chip_size=100 --patch_size=94 --use_phase=True --dataset=soc 
+$ python3 generate_dataset.py --is_train=False --use_phase=True --chip_size=128 --patch_size=128  --use_phase=True --dataset=soc
 $ cd ..
-$ python3 train.py
+$ python3 train.py --config_name=config/AConvNet-SOC.json
 ```
 
 #### Results of SOC
-- Final Accuracy is **99.18%** (The official accuracy is 99.13%)
+- Final Accuracy is **99.13%** at epoch 26 (The official accuracy is 99.13%)
 - You can see the details in `notebook/experiments-SOC.ipynb`
 
 - Visualization of training loss and test accuracy
@@ -165,10 +163,10 @@ $ python3 train.py
 
 | Noise | 1% | 5% | 10% | 15%|
 | :---: | :---: | :---: | :---: | :---: |
-| AConvNet-PyTorch | 98.56 | 94.39 | 85.03 | 73.65 |
+| AConvNet-PyTorch | 98.60 | 95.18 | 85.36 | 73.24 |
 | AConvNet-Official | 91.76 | 88.52 | 75.84 | 54.68 |
 
-
+<!--
 ### Extended Operating Conditions (EOC)
 
 #### EOC-1 (Large depression angle change)
@@ -216,15 +214,28 @@ MSTAR-PublicMixedTargets-CD2/MSTAR_PUBLIC_MIXED_TARGETS_CD2
 └ ...
 
 ```
-- Train Target: 2S1, BRDM2, T72, ZSU234 with depression angle 17$\degree$
-- Test Target: 2S1, BRDM2, T72, ZSU234 with depression angle 30$\degree$
+
+#### Quick Start Guide for Training
+
+- Dataset Preparation
+    - Download the [soc-dataset.zip](https://github.com/jangsoopark/AConvNet-pytorch/releases/download/V2.0.0/soc-raw.zip) 
+    - After extracting it, you can find `train` and  `test` directories inside `raw` directory.
+    - Place the two directories (`train` and  `test`) to the `dataset/raw`.
+```shell
+$ cd src/data 
+$ python3 generate_dataset.py --is_train=True --use_phase=True --chip_size=96 --dataset=eoc-1 
+$ python3 generate_dataset.py --is_train=False --use_phase=True --dataset=soc
+$ cd ..
+$ python3 train.py --config_name=config/AConvNet-EOC-1.json
+```
+
 
 #### EOC-2 (Target configuration and version variants)
 
 ### Outlier Rejection
 
 ### End-to-End SAR-ATR Cases
-
+-->
 ## Details about the specific environment of this repository
 
 | | |

experiments/config/AConvNet-EOC-1.json

@@ -4,10 +4,10 @@
   "num_classes": 4,
   "channels": 2,
   "batch_size": 100,
-  "epochs": 50,
+  "epochs": 100,
   "momentum": 0.9,
-  "lr": 1e-3,
-  "lr_step": [14],
+  "lr": 1e-4,
+  "lr_step": [50],
   "lr_decay": 0.1,
   "weight_decay": 4e-3,
   "dropout_rate": 0.5

experiments/config/AConvNet-SOC.json

@@ -4,10 +4,10 @@
   "num_classes": 10,
   "channels": 2,
   "batch_size": 100,
-  "epochs": 50,
+  "epochs": 100,
   "momentum": 0.9,
   "lr": 1e-3,
-  "lr_step": [26],
+  "lr_step": [50],
   "lr_decay": 0.1,
   "weight_decay": 4e-3,
   "dropout_rate": 0.5

requirements.txt

@@ -6,4 +6,5 @@ tqdm==4.61.2
 torchvision==0.10.0+cu111
 matplotlib
 scikit-learn
-seaborn
+seaborn
+Pillow

src/data/generate_dataset.py

@@ -3,6 +3,7 @@
 from absl import app
 
 from multiprocessing import Pool
+from PIL import Image
 import numpy as np
 
 import json
@@ -13,23 +14,34 @@
 
 flags.DEFINE_string('image_root', default='dataset', help='')
 flags.DEFINE_string('dataset', default='soc', help='')
-flags.DEFINE_boolean('is_train', default=True, help='')
-flags.DEFINE_boolean('use_phase', default=False, help='')
-flags.DEFINE_integer('chip_size', default=94, help='')
+flags.DEFINE_boolean('is_train', default=False, help='')
+flags.DEFINE_integer('chip_size', default=100, help='')
+flags.DEFINE_integer('patch_size', default=94, help='')
+flags.DEFINE_boolean('use_phase', default=True, help='')
+
 FLAGS = flags.FLAGS
 
 project_root = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
 
 
-def generate(src_path, dst_path, is_train, use_phase, chip_size, dataset):
+def data_scaling(chip):
+    r = chip.max() - chip.min()
+    return (chip - chip.min()) / r
+
+
+def log_scale(chip):
+    return np.log10(np.abs(chip) + 1)
+
+
+def generate(src_path, dst_path, is_train, chip_size, patch_size, use_phase, dataset):
     if not os.path.exists(src_path):
         return
     if not os.path.exists(dst_path):
         os.makedirs(dst_path, exist_ok=True)
     print(f'Target Name: {os.path.basename(dst_path)}')
 
     _mstar = mstar.MSTAR(
-        name=dataset, is_train=is_train, use_phase=use_phase, chip_size=chip_size, patch_size=88, stride=1
+        name=dataset, is_train=is_train, chip_size=chip_size, patch_size=patch_size, use_phase=use_phase, stride=1
     )
 
     image_list = glob.glob(os.path.join(src_path, '*'))
@@ -40,7 +52,10 @@ def generate(src_path, dst_path, is_train, use_phase, chip_size, dataset):
             name = os.path.splitext(os.path.basename(path))[0]
             with open(os.path.join(dst_path, f'{name}-{i}.json'), mode='w', encoding='utf-8') as f:
                 json.dump(label, f, ensure_ascii=False, indent=2)
+
+            # _image = log_scale(_image)
             np.save(os.path.join(dst_path, f'{name}-{i}.npy'), _image)
+            # Image.fromarray(data_scaling(_image)).convert('L').save(os.path.join(dst_path, f'{name}-{i}.bmp'))
 
 
 def main(_):
@@ -57,7 +72,7 @@ def main(_):
         (
             os.path.join(raw_root, mode, target),
             os.path.join(output_root, target),
-            FLAGS.is_train, FLAGS.use_phase, FLAGS.chip_size, FLAGS.dataset
+            FLAGS.is_train, FLAGS.chip_size, FLAGS.patch_size, FLAGS.use_phase, FLAGS.dataset
         ) for target in mstar.target_name_soc
     ]
 

src/data/loader.py

@@ -1,6 +1,6 @@
 import numpy as np
 
-import torchvision
+from skimage import io
 import torch
 import tqdm
 

src/data/preprocess.py

@@ -35,8 +35,13 @@ def __call__(self, sample):
 
         h, w, _ = _input.shape
         oh, ow = self.size
-        y = np.random.randint(0, h - oh)
-        x = np.random.randint(0, w - ow)
+
+        dh = h - oh
+        dw = w - ow
+        y = np.random.randint(0, dh) if dh > 0 else 0
+        x = np.random.randint(0, dw) if dw > 0 else 0
+        oh = oh if dh > 0 else h
+        ow = ow if dw > 0 else w
 
         return _input[y: y + oh, x: x + ow, :]
 

src/model/network.py

@@ -12,7 +12,7 @@ def __init__(self, **params):
         self.classes = params.get('classes', 10)
         self.channels = params.get('channels', 1)
 
-        _w_init = params.get('w_init', lambda x: nn.init.kaiming_uniform_(x, nonlinearity='relu'))
+        _w_init = params.get('w_init', lambda x: nn.init.kaiming_normal_(x, nonlinearity='relu'))
         _b_init = params.get('b_init', lambda x: nn.init.constant_(x, 0.1))
 
         self._layer = nn.Sequential(
@@ -34,7 +34,7 @@ def __init__(self, **params):
             ),
             nn.Dropout(p=self.dropout_rate),
             _blocks.Conv2DBlock(
-                shape=[3, 3, 128, self.classes], stride=3, padding='valid',
+                shape=[3, 3, 128, self.classes], stride=1, padding='valid',
                 w_init=_w_init, b_init=nn.init.zeros_
             ),
             nn.Flatten()

src/train.py

@@ -14,6 +14,7 @@
 import json
 import os
 
+from data import preprocess
 from data import loader
 from utils import common
 import model
@@ -22,14 +23,17 @@
 flags.DEFINE_string('config_name', 'config/AConvNet-SOC.json', help='')
 FLAGS = flags.FLAGS
 
-#common.set_random_seed(12321)
 
+common.set_random_seed(12321)
 
-def load_dataset(path, is_train, name, batch_size):
 
+def load_dataset(path, is_train, name, batch_size):
+    transform = [preprocess.CenterCrop(88), torchvision.transforms.ToTensor()]
+    if is_train:
+        transform = [preprocess.RandomCrop(88), torchvision.transforms.ToTensor()]
     _dataset = loader.Dataset(
         path, name=name, is_train=is_train,
-        transform=torchvision.transforms.Compose([torchvision.transforms.ToTensor()])
+        transform=torchvision.transforms.Compose(transform)
     )
     data_loader = torch.utils.data.DataLoader(
         _dataset, batch_size=batch_size, shuffle=is_train, num_workers=1
@@ -99,7 +103,9 @@ def run(epochs, dataset, classes, channels, batch_size,
 
         accuracy = validation(m, valid_set)
 
-        logging.info(f'Epoch: {epoch + 1:03d}/{epochs:03d} | loss={np.mean(_loss):.4f} | lr={lr} | accuracy={accuracy}')
+        logging.info(
+            f'Epoch: {epoch + 1:03d}/{epochs:03d} | loss={np.mean(_loss):.4f} | lr={lr} | accuracy={accuracy:.2f}'
+        )
 
         history['loss'].append(np.mean(_loss))
         history['accuracy'].append(accuracy)