no message

Author: ronekko

demo.py

@@ -5,7 +5,6 @@
 @author: yamane
 """
 
-from path import Path
 import os
 import numpy as np
 import matplotlib.pyplot as plt
@@ -19,13 +18,13 @@
 import make_html
 
 
-def lossfun(model, stream):
+def lossfun(model, stream, preprocess):
     loss = []
     loss_abs = []
     target = []
     predict = []
     for it in stream.get_epoch_iterator():
-        x, t_l = load_datasets.data_crop(it[0])
+        x, t_l = load_datasets.data_crop(it[0], preprocess=preprocess)
         y_l = model.predict(x, True)
         e_l = t_l - y_l
         e_l_abs = np.abs(t_l - y_l)
@@ -51,9 +50,12 @@ def show_and_save(stream, target, predict, save_path_f, save_path_d,
             fix_img = utility.change_aspect_ratio(dis_img, 1/y_r[batch][i], 1)
 
             print('[test_data]:', i+1)
-            print('[t_l]:', np.round(target[batch][i], 4), '\t[t_r]:', np.round(t_r[batch][i], 4))
-            print('[y_l]:', np.round(predict[batch][i], 4), '\t[y_r]:', np.round(y_r[batch][i], 4))
-            print('[e_l]:', np.round(e_l[i], 4), '\t[e_r]:', np.round(e_r[i], 4))
+            print('[t_l]:', np.round(target[batch][i], 4),
+                  '\t[t_r]:', np.round(t_r[batch][i], 4))
+            print('[y_l]:', np.round(predict[batch][i], 4),
+                  '\t[y_r]:', np.round(y_r[batch][i], 4))
+            print('[e_l]:', np.round(e_l[i], 4), '\t[e_r]:',
+                  np.round(e_r[i], 4))
 
             plt.figure(figsize=(16, 16))
             plt.subplot(131)
@@ -105,89 +107,89 @@ def draw_graph(loss, loss_abs, success_asp, num_test, save_root):
     else:
         max_value = np.abs(min(error))
 
-    plt.figure(figsize=(16, 12))
+    plt.rcParams["font.size"] = 18
+    plt.figure(figsize=(8, 3))
     plt.plot(error_abs)
     plt.plot(base_line, 'r-')
-    plt.title('absolute Error for each test data', fontsize=28)
-    plt.legend(["Error", "log(1.1)"], loc="upper right")
-    plt.xlabel('Order of test data number', fontsize=28)
-    plt.ylabel('Error(|t-y|) in log scale', fontsize=28)
+    plt.legend(["Error", "log(1.1303)"], loc="upper left")
+    plt.xlabel('Order of test data number', fontsize=24)
+    plt.ylabel('Error(|t-y|) in log scale', fontsize=24)
     plt.ylim(0, max(error_abs)+0.01)
     plt.grid()
     plt.savefig(loss_abs_file+'.jpg', format='jpg', bbox_inches='tight')
     plt.show()
 
-    plt.figure(figsize=(16, 12))
+    plt.figure(figsize=(8, 3))
     plt.plot(error, label='Error')
-    plt.plot(base_line, label="log(1.1)")
-    plt.plot(-base_line, label="log(1.1^-1)")
+    plt.plot(base_line, label="log(1.1303)")
+    plt.plot(-base_line, label="log(1.1303^-1)")
     plt.title('Error for each test data', fontsize=28)
-    plt.legend(loc="upper right")
+    plt.legend(loc="upper left")
     plt.xlabel('Order of test data number', fontsize=28)
     plt.ylabel('Error(t-y) in log scale', fontsize=28)
     plt.ylim(-max_value-0.01, max_value+0.01)
     plt.grid()
     plt.savefig(loss_file+'.jpg', format='jpg', bbox_inches='tight')
     plt.show()
 
-    fig = plt.figure(figsize=(16, 12))
+    fig = plt.figure(figsize=(8, 3))
     ax = fig.add_subplot(1, 1, 1)
-    ax.hist(error, bins=25)
-    ax.set_title('Error histogram', fontsize=28)
-    ax.set_xlabel('Error(t-y) in log scale', fontsize=28)
-    ax.set_ylabel('Percentage', fontsize=28)
-    plt.xlim(-1, 1)
+    ax.hist(error, bins=22, range=(-1.0, 1.0))
+    ax.set_xlabel('Error in log scale', fontsize=20)
+    ax.set_ylabel('Percentage', fontsize=20)
+    plt.grid()
+#    plt.xlim(-1, 1)
     plt.savefig(loss_hist+'.jpg', format='jpg', bbox_inches='tight')
     fig.show()
 
     count = 0
     for i in range(num_test):
         if loss_abs[0][i] < threshold:
             count += 1
-    print('under log(1.1) =', count, '%')
+    print('under log(1.1303) =', count, '%')
     print('[mean]:', np.mean(loss_abs))
 
 
 if __name__ == '__main__':
     # テスト結果を保存する場所
-#    save_root = r'E:\demo'
-    save_root = r'E:\yamane'
+    save_root = r'demo'
     # テストに使うモデルのnpzファイルの場所
-#    model_file = r'voc2012_regression_max_pooling\1489665734.69_asp_max_4.0\voc2012_regression_max_pooling.npz'
-    model_file = 'dog_data_regression_ave_pooling.npz'
+    model_file = r'npz\dog_data_regression_ave_pooling.npz'
     num_train = 16500  # 学習データ数
     num_valid = 500  # 検証データ数
     num_test = 100  # テストデータ数
     asp_r_max = 3.0  # 歪み画像の最大アスペクト比
-    success_asp = 1.1  # 修正成功とみなす修正画像のアスペクト比の最大値
+    success_asp = np.exp(0.12247601469)  # 修正成功とみなす修正画像のアスペクト比の最大値
     batch_size = 100
+    preprocesses = [None, 'edge', 'blur']  # specify None or 'edge' or 'blur'
 
     # モデルのファイル名をフォルダ名にする
-#    folder_name = model_file.split('\\')[-2]
-    folder_name = model_file.split('.')[-2]
-
-    # テスト結果を保存するフォルダを作成
-    test_folder_path = utility.create_folder(save_root, folder_name)
-    fix_folder_path = utility.create_folder(test_folder_path, 'fix')
-    dis_folder_path = utility.create_folder(test_folder_path, 'distorted')
-    ori_folder_path = utility.create_folder(test_folder_path, 'original')
-    model_path = Path(save_root) / model_file
-
-    # モデル読み込み
-    model = voc2012_regression_max_pooling.Convnet().to_gpu()
-    serializers.load_npz(model_path, model)
-
-    # streamを取得
-    streams = load_datasets.load_voc2012_stream(
-        batch_size, num_train, num_valid, num_test)
-    train_stream, valid_stream, test_stream = streams
-
-    # 歪み画像の修正を実行
-    with chainer.using_config('train', False):
-        loss, loss_abs, target, predict = lossfun(model, test_stream)
-
-    show_and_save(test_stream, target, predict, fix_folder_path,
-                  dis_folder_path, ori_folder_path)
-
-    # 修正結果の誤差を描画
-    draw_graph(loss, loss_abs, success_asp, num_test, test_folder_path)
+    folder_name = model_file.split('\\')[-2]
+
+    for preprocess in preprocesses:
+        # テスト結果を保存するフォルダを作成
+        test_folder_path = utility.create_folder(
+            save_root, folder_name + '_' + str(preprocess))
+        fix_folder_path = utility.create_folder(test_folder_path, 'fix')
+        dis_folder_path = utility.create_folder(test_folder_path, 'distorted')
+        ori_folder_path = utility.create_folder(test_folder_path, 'original')
+
+        # モデル読み込み
+        model = voc2012_regression_max_pooling.Convnet().to_gpu()
+        serializers.load_npz(model_file, model)
+
+        # streamを取得
+        streams = load_datasets.load_voc2012_stream(
+            batch_size, num_train, num_valid, num_test)
+        train_stream, valid_stream, test_stream = streams
+
+        # 歪み画像の修正を実行
+        with chainer.no_backprop_mode(), chainer.using_config('train', False):
+            loss, loss_abs, target, predict = lossfun(model, test_stream,
+                                                      preprocess=preprocess)
+
+    #    show_and_save(test_stream, target, predict, fix_folder_path,
+    #                  dis_folder_path, ori_folder_path)
+
+        # 修正結果の誤差を描画
+        draw_graph(loss, loss_abs, success_asp, num_test, test_folder_path)

load_datasets.py

@@ -71,14 +71,26 @@ def load_toy_stream(batch_size, ):
 
 
 def data_crop(X_batch, aspect_ratio_max=3.0, output_size=256, crop_size=224,
-              random=True, t=0):
+              random=True, t=0, preprocess=None):
+    if preprocess == 'edge':
+        kernel = np.array([[-1, -1, -1],
+                           [-1,  8, -1],
+                           [-1, -1, -1]
+                           ], np.float32)
+    elif preprocess == 'blur':
+        kernel = np.ones((5, 5), np.float32) / 25
+
     images = []
     ts = []
 
     for b in range(X_batch.shape[0]):
         # 補間方法を乱数で設定
         u = np.random.randint(5)
         image = X_batch[b]
+
+        if preprocess is not None:
+            image = cv2.filter2D(image, -1, kernel)
+
         if random is False:
             t = t
         else: