new scriptsHEADmaster

7 files changed, 99 insertions, 24 deletions

diff --git a/.gitignore b/.gitignore
index ac7b520..4b44bdb 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,15 +1,8 @@
+__pycache__
+checkpoints
+summary
+psnrs
+*.pyc
+Data/
+tfpy3
 
-.idea/webServers.xml
-Codes/__pycache__/unet.cpython-36.pyc
-Codes/__pycache__/pix2pix.cpython-36.pyc
-.idea/deployment.xml
-.idea/ano_pred_cvpr2018.iml
-.idea/codeStyles/codeStyleConfig.xml
-.idea/misc.xml
-.idea/vcs.xml
-.idea/workspace.xml
-.idea/modules.xml
-.idea/codeStyles/codeStyleConfig.xml
-.idea/codeStyles/codeStyleConfig.xml
-.idea/codeStyles/codeStyleConfig.xml
-.idea/codeStyles/codeStyleConfig.xml
diff --git a/Codes/evaluate.py b/Codes/evaluate.py
index cd0fcef..4909e50 100644
--- a/Codes/evaluate.py
+++ b/Codes/evaluate.py
@@ -114,8 +114,10 @@ class GroundTruthLoader(object):
             gt = self.__load_shanghaitech_gt()
         elif dataset == GroundTruthLoader.TOY_DATA:
             gt = self.__load_toydata_gt()
+        #else:
+        #    gt = self.__load_ucsd_avenue_subway_gt(dataset)
         else:
-            gt = self.__load_ucsd_avenue_subway_gt(dataset)
+            gt = self.__load_no_anomaly_gt(dataset)
         return gt
 
     def __load_ucsd_avenue_subway_gt(self, dataset):
@@ -200,6 +202,15 @@ class GroundTruthLoader(object):
         return gt
 
     @staticmethod
+    def __load_no_anomaly_gt():
+        gt = []
+        for i in xrange(27):
+            length = 64
+            video_gt = np.zeros((length,), dtype=np.uint8)
+            gt.append(video_gt)
+        return gt
+
+    @staticmethod
     def get_pixel_masks_file_list(dataset):
         # pixel mask folder
         pixel_mask_folder = os.path.join(DATA_DIR, dataset, 'pixel_masks')
@@ -573,4 +584,4 @@ if __name__ == '__main__':
     if eval_type == 'test_func':
         test_func()
     else:
-        evaluate(eval_type, file_path)
\ No newline at end of file
+        evaluate(eval_type, file_path)
diff --git a/Codes/inference-test.py b/Codes/inference-test.py
index b16a56c..83a4fd4 100644
--- a/Codes/inference-test.py
+++ b/Codes/inference-test.py
@@ -10,6 +10,7 @@ from utils import DataLoader, load, save, psnr_error
 from constant import const
 import evaluate
 
+from PIL import Image
 
 slim = tf.contrib.slim
 
@@ -66,6 +67,7 @@ with tf.Session(config=config) as sess:
         load(loader, sess, ckpt)
 
         output_records = []
+        outputs = []
         videos_info = data_loader.videos
         num_videos = len(videos_info.keys())
         total = 0
@@ -78,15 +80,19 @@ with tf.Session(config=config) as sess:
 
             for i in range(num_his, length):
                 video_clip = data_loader.get_video_clips(video_name, i - num_his, i + 1)
-                output, psnr = sess.run([test_outputs, test_psnr_error]
+                output, psnr = sess.run([test_outputs, test_psnr_error],
                                 feed_dict={test_video_clips_tensor: video_clip[np.newaxis, ...]})
-                outputs[i] = output
+                outputs.append(output)
 
-                tf.image.encode_png(
-                    output,
-                    compression=-1,
-                    name=None
-                )
+                #tf.image.encode_png(
+                #    output,
+                #    compression=-1,
+                #    name=None
+                #)
+                ohhh = (output[0,...] + 1.0) * 127.5
+                print(ohhh.shape, np.amin(ohhh), np.amax(ohhh))
+                out = Image.fromarray(ohhh, 'RGB')
+                out.save(os.path.join(os.getcwd(), '{}_{:05d}.png'.format(video_name, i)))
 
                 print('video = {} / {}, i = {} / {}, psnr = {:.6f}'.format(
                     video_name, num_videos, i, length, psnr))
@@ -108,6 +114,7 @@ with tf.Session(config=config) as sess:
         # print(results)
 
 
+    print(os.path.isdir(snapshot_dir))
     if os.path.isdir(snapshot_dir):
         def check_ckpt_valid(ckpt_name):
             is_valid = False
diff --git a/Codes/test.sh b/Codes/test.sh
new file mode 100644
index 0000000..a683a5a
--- /dev/null
+++ b/Codes/test.sh
@@ -0,0 +1,12 @@
+#python inference-test.py --dataset kaulsdorf_gray \
+#  --train_folder /home/lens/Desktop/thumbs/kaulsdorf_gray/Train/ \
+#  --test_folder /home/lens/Desktop/thumbs/kaulsdorf_gray/Test/ \
+#  --snapshot_dir ./checkpoints/kaulsdorf_gray_l_2_alpha_1_lp_1.0_adv_0.05_gdl_1.0_flow_0.01/model.ckpt-140000 \
+#  --gpu 0
+
+python inference-test.py --dataset kaulsdorf_single \
+  --train_folder /home/lens/code/Adversarial_Video_Generation/Data/kaulsdorf_single/Train/ \
+  --test_folder /home/lens/code/Adversarial_Video_Generation/Data/kaulsdorf_single/Test/ \
+  --snapshot_dir ./checkpoints/kaulsdorf_single_l_2_alpha_1_lp_0.001_adv_0.05_gdl_1.0_flow_2.0/model.ckpt-19000 \
+  --gpu 0
+
diff --git a/Codes/train.sh b/Codes/train.sh
new file mode 100644
index 0000000..3145aa9
--- /dev/null
+++ b/Codes/train.sh
@@ -0,0 +1,6 @@
+python train.py --dataset kaulsdorf_gray \
+  --train_folder /home/lens/Desktop/thumbs/kaulsdorf_gray/Train/ \
+  --test_folder /home/lens/Desktop/thumbs/kaulsdorf_gray/Test/ \
+  --gpu 0  \
+  --iters 500000
+
diff --git a/Codes/training_hyper_params/hyper_params.ini b/Codes/training_hyper_params/hyper_params.ini
index f7868f1..963902c 100644
--- a/Codes/training_hyper_params/hyper_params.ini
+++ b/Codes/training_hyper_params/hyper_params.ini
@@ -41,6 +41,52 @@ LRATE_D = [0.00001, 0.000001]
 LRATE_D_BOUNDARIES = [40000]
 
 
+[kaulsdorf_gray]
+# for lp loss. e.g, 1 or 2 for l1 and l2 loss, respectively)
+L_NUM = 2
+# the power to which each gradient term is raised in GDL loss
+ALPHA_NUM = 1
+# the percentage of the adversarial loss to use in the combined loss
+LAM_ADV = 0.05
+# the percentage of the lp loss to use in the combined loss
+LAM_LP = 1
+# the percentage of the GDL loss to use in the combined loss
+LAM_GDL = 1
+# the percentage of the different frame loss, LAM_FLOW = 2 is also ok, but LAM_FLOW = 0.01 is slightly better.
+LAM_FLOW = 0.01
+
+# For gray scale video, such as Ped2 and Ped1, learning rate of G and D star from 1e-4 and 1e-5, respectively.
+LRATE_G = [0.0001, 0.00001]
+LRATE_G_BOUNDARIES = [40000]
+
+LRATE_D = [0.00001, 0.000001]
+LRATE_D_BOUNDARIES = [40000]
+
+
+[kaulsdorf_single]
+# for lp loss. e.g, 1 or 2 for l1 and l2 loss, respectively)
+L_NUM = 2
+# the power to which each gradient term is raised in GDL loss
+ALPHA_NUM = 1
+# the percentage of the adversarial loss to use in the combined loss
+LAM_ADV = 0.05
+# the percentage of the lp loss to use in the combined loss,
+# we found in smaller lp is slightly better in avenue, but not too much difference
+# LAM_LP = 1 is 84.9, LAM_LP = 0.001 may arrive to 85.1
+LAM_LP = 0.001
+# the percentage of the GDL loss to use in the combined loss
+LAM_GDL = 1
+# the percentage of the different frame loss
+LAM_FLOW = 2
+
+# For rgb color scale video, such as Ped2 and Ped1, learning rate of G and D star from 2e-4 and 2e-5, respectively.
+LRATE_G = [0.0002, 0.00002]
+LRATE_G_BOUNDARIES = [100000]
+
+LRATE_D = [0.00002, 0.000002]
+LRATE_D_BOUNDARIES = [100000]
+
+
 [avenue]
 # for lp loss. e.g, 1 or 2 for l1 and l2 loss, respectively)
 L_NUM = 2
diff --git a/Codes/utils.py b/Codes/utils.py
index 8466d08..7f54fe7 100644
--- a/Codes/utils.py
+++ b/Codes/utils.py
@@ -76,7 +76,7 @@ class DataLoader(object):
             video_name = video.split('/')[-1]
             self.videos[video_name] = {}
             self.videos[video_name]['path'] = video
-            self.videos[video_name]['frame'] = glob.glob(os.path.join(video, '*.jpg'))
+            self.videos[video_name]['frame'] = glob.glob(os.path.join(video, '*.png'))
             self.videos[video_name]['frame'].sort()
             self.videos[video_name]['length'] = len(self.videos[video_name]['frame'])