更新ＣＮＮ参数

tools/TrainCNN/backward.py

@@ -54,7 +54,7 @@ def save_para(folder, paras):
         save_bias(fp, paras[7])
 
 
-STEPS = 100000
+STEPS = 60000
 BATCH = 50
 LEARNING_RATE_BASE  = 0.001
 LEARNING_RATE_DECAY = 0.99
@@ -101,16 +101,31 @@ def train(dataset, show_bar=False):
 
             _, loss_value, step = sess.run(
                 [train_op, loss, global_step],
-                feed_dict={x: images_samples, y_: labels_samples, keep_rate:0.3}
+                feed_dict={x: images_samples, y_: labels_samples, keep_rate:0.2}
             )
 
-            if i % 100 == 0:
+            if (i-1) % 100 == 0:
-                if i % 500 == 0:
+                if (i-1) % 500 == 0:
-                    test_images, test_labels = dataset.sample_test_sets(10000)
+                    test_images, test_labels = dataset.sample_test_sets(5000)
-                    acc = sess.run(accuracy, feed_dict={x: test_images, y_: test_labels, keep_rate:1.0})
+                    test_acc, output = sess.run([accuracy, y], feed_dict={x: test_images, y_: test_labels, keep_rate:1.0})
-                bar.set_postfix({"loss": loss_value, "acc": acc})
+                    output = np.argmax(output, axis=1)
+                    real = np.argmax(test_labels, axis=1)
+                    print("=============test-set===============")
+                    for n in range(forward.OUTPUT_NODES):
+                        print("label: %d, precise: %f, recall: %f" %
+                            (n, np.mean(real[output==n]==n), np.mean(output[real==n]==n)))
 
+                    train_images, train_labels = dataset.sample_train_sets(5000)
+                    train_acc, output = sess.run([accuracy, y], feed_dict={x: train_images, y_: train_labels, keep_rate:1.0})
+                    output = np.argmax(output, axis=1)
+                    real = np.argmax(train_labels, axis=1)
+                    print("=============train-set===============")
+                    for n in range(forward.OUTPUT_NODES):
+                        print("label: %d, precise: %f, recall: %f" %
+                              (n, np.mean(real[output==n]==n), np.mean(output[real==n]==n)))
+                    print("\n")
 
+                bar.set_postfix({"loss": loss_value, "train_acc": train_acc, "test_acc": test_acc})
 
         vars_val = sess.run(vars)
         save_para("/home/xinyang/Workspace/RM_auto-aim/tools/para", vars_val)
@@ -206,9 +221,9 @@ def train(dataset, show_bar=False):
 
 
 if __name__ == "__main__":
-    import os
+    # import os
-    os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+    # os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
-    os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
+    # os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
-    dataset = generate.DataSet("/home/xinyang/Workspace/box_cut")
+    dataset = generate.DataSet("/home/xinyang/Workspace/box_resize")
     train(dataset, show_bar=True)
     input("press enter to continue...")

tools/TrainCNN/forward.py

@@ -29,16 +29,16 @@ def max_pool_2x2(x):
 CONV1_KERNAL_SIZE = 5
 
 # 第一层卷积输出通道数
-CONV1_OUTPUT_CHANNELS = 6
+CONV1_OUTPUT_CHANNELS = 8
 
 # 第二层卷积核大小
 CONV2_KERNAL_SIZE = 3
 
 # 第二层卷积输出通道数
-CONV2_OUTPUT_CHANNELS = 12
+CONV2_OUTPUT_CHANNELS = 16
 
 # 第一层全连接宽度
-FC1_OUTPUT_NODES = 30
+FC1_OUTPUT_NODES = 100
 
 # 第二层全连接宽度（输出标签类型数）
 FC2_OUTPUT_NODES = 15

tools/para/conv1_b

@@ -1,7 +1,9 @@
-6
+8
--0.07201801
+-0.19897088
-2.4805095
+2.2773967
--0.03871701
+-0.07212669
--0.0016837252
+-0.22893764
--0.0016489439
+-0.022769619
--0.11918187
+0.9122422
+1.3221853
+-0.21709026

tools/para/conv2_b

@@ -1,13 +1,17 @@
-12
+16
-0.011819358
+0.032270256
-2.0857728
+2.3110154
--0.00558426
+0.4078679
-0.571528
+0.30516425
--0.5205912
+-0.027841559
-1.1178641
+0.80007833
-0.12018313
+0.10999302
-0.35316235
+0.83167756
-2.4709818
+-0.09058099
--0.030677324
+0.67901427
--0.07035873
+0.2736649
-2.3538203
+0.08881351
+0.10975416
+0.25698876
+0.076515704
+-0.017468728

tools/para/fc1_b

@@ -1,31 +1,101 @@
-30
+100
--0.004537243
+-0.39040318
--0.7359988
+-0.031096617
--0.014530569
+-0.06425226
-1.9556539
+0.24911235
--0.005490738
+-0.002578787
--0.0075725853
+-0.086705275
-1.0653863
+-0.0322658
--0.42296687
+-0.017816741
--0.33600682
+-0.11621032
--0.7384473
+0.21196772
-0.7910003
+-0.004639828
--0.00820433
+-0.023076132
--0.025636531
+-0.50997764
--0.02046858
+-0.04299724
--0.0068890513
+-0.01989839
--0.26979277
+-0.011238396
--0.004196582
+-0.003221448
--0.7438407
+-0.019384952
--0.5907334
+-0.0007764693
--0.38981074
+-0.015599826
-0.13757768
+0.16373938
--0.027325256
+-0.0027049272
-1.1313668
+-0.18095633
-0.5019343
+-0.050923813
-1.462373
+0.12674743
-0.6535602
+-0.064153716
-0.8464444
+-0.028386148
-0.059315335
+-0.059802737
--0.030320408
+-0.036068685
--0.329629
+-0.004065791
+-0.03783843
+-0.16458924
+-0.0328307
+-0.032716025
+-0.020594684
+-0.042352736
+-0.084991984
+-0.028080234
+-0.001538593
+-0.10711875
+-0.024680987
+-0.008004385
+-0.5063542
+-0.09158748
+-0.08181085
+-0.22574262
+-0.075171836
+0.28233245
+-0.024944687
+-0.0029645876
+-0.041441295
+-0.08904015
+0.30993482
+-0.06328518
+-0.0075723003
+-0.005151164
+-0.0021952058
+-0.013833341
+-0.023337327
+-0.01824665
+-0.025177158
+-0.067239
+-0.02126352
+0.11769418
+-0.64603645
+-0.014887376
+-0.14686602
+-0.020528413
+-0.018256638
+-0.0017088759
+-0.018110225
+-0.003289471
+1.0441891
+0.30619404
+-0.001282074
+-0.09424017
+-0.24455559
+-0.026046017
+-0.004658401
+-0.022633847
+-0.022873487
+0.4393057
+-0.033948973
+-0.042779494
+-0.0059623853
+0.6859317
+-0.19052452
+-0.020080235
+-0.010588832
+0.012147919
+-0.002949453
+0.41500625
+-0.16353038
+-0.023607356
+-0.38747007
+-0.014272043
+-0.0033837124
+-0.1627222
+-0.055108108
+0.74174875

tools/para/fc2_b

@@ -1,16 +1,16 @@
 15
-1.4582382
+0.7158619
-0.23248857
+0.074666105
-0.04220094
+-0.20262705
--0.46074286
+-0.32018387
--1.0683035
+-0.35891113
--0.18858828
+-0.021645868
-0.7637116
+0.2373232
-0.50826854
+0.17633525
--0.99831194
+-0.4311263
--0.7008356
+-0.2179705
--0.7342338
+-0.21293324
-0.08691424
+0.20182535
--0.80010396
+-0.8007338
--0.11393161
+-0.02198011
-0.1028082
+0.30222887