Towards a bigger C3D

Author: pratikgujjar

code/autoencoder_model/gifs/vid_17.gif

@@ -79,14 +79,16 @@
 # KL coeff damages learning
 KL_COEFF = 0
 CLASSIFIER = True
-BUF_SIZE = 10
-LOSS_WEIGHTS = [1, 1]
-A_TRAIN_RATIO = 1
-C_TRAIN_RATIO = 1
-
-ped_actions = ['slow down', 'standing', 'moving fast', 'speed up', 'look', 'nod', 'unknown', 'moving slow',
-               'flasher signal', 'looking' , 'handwave', 'clear path', 'stopped', 'slowing down',
-               'crossing', 'speeding up']
+RAM_DECIMATE = True
+RETRAIN_CLASSIFIER = True
+CLASS_TARGET_INDEX = 24
+
+ped_actions = ['slow down', 'moving slow', 'standing', 'stopped',
+               'speed up', 'moving fast', 'look', 'looking',  'clear path',
+               'crossing', 'nod', 'handwave', 'unknown']
+
+simple_ped_set = ['crossing', 'stopped', 'looking', 'clear path', 'unknown']
+
 driver_actions = ['moving slow', 'slowing down', 'standing', 'speeding up', 'moving fast']
 simple_driver_set = ['slow down', 'stop', 'speed up']
 
@@ -107,22 +109,32 @@
 print ("Loading network/training configuration...")
 print ("Config file: " + str(__name__))
 
-BATCH_SIZE = 10
+BATCH_SIZE = 25
 NB_EPOCHS_AUTOENCODER = 0
 NB_EPOCHS_CLASS = 100
 
 OPTIM_A = Adam(lr=0.0001, beta_1=0.5)
-OPTIM_G = Adam(lr=0.0001, beta_1=0.5)
-OPTIM_D = Adam(lr=0.000001, beta_1=0.5)
-OPTIM_C = Adam(lr=0.0001, beta_1=0.5)
-
-lr_schedule = [10, 20, 30]  # epoch_step
-
-def schedule(epoch_idx):
-    if (epoch_idx + 1) < lr_schedule[0]:
-        return 0.00000002
-    elif (epoch_idx + 1) < lr_schedule[1]:
-        return 0.000000002  # lr_decay_ratio = 10
-    elif (epoch_idx + 1) < lr_schedule[2]:
-        return 0.0000000002
-    return 0.0000000002
+# OPTIM_C = Adam(lr=0.0000002, beta_1=0.5)
+OPTIM_C = SGD(lr=0.0001, momentum=0.9, nesterov=True)
+
+
+auto_lr_schedule = [25, 30, 35]  # epoch_step
+def auto_schedule(epoch_idx):
+    if (epoch_idx + 1) < auto_lr_schedule[0]:
+        return 0.0001
+    elif (epoch_idx + 1) < auto_lr_schedule[1]:
+        return 0.00001  # lr_decay_ratio = 10
+    elif (epoch_idx + 1) < auto_lr_schedule[2]:
+        return 0.00001
+    return 0.00001
+
+
+clas_lr_schedule = [50, 55, 60]  # epoch_step
+def clas_schedule(epoch_idx):
+    if (epoch_idx + 1) < clas_lr_schedule[0]:
+        return 0.0001
+    elif (epoch_idx + 1) < clas_lr_schedule[1]:
+        return 0.00001  # lr_decay_ratio = 10
+    elif (epoch_idx + 1) < clas_lr_schedule[2]:
+        return 0.000001
+    return 0.000001

code/autoencoder_model/gifs/vid_18.gif

@@ -83,7 +83,9 @@
 LOSS_WEIGHTS = [1, 1]
 A_TRAIN_RATIO = 1
 C_TRAIN_RATIO = 1
-RAM_DECIMATE = True
+RAM_DECIMATE = False
+RETRAIN_CLASSIFIER = True
+CLASS_TARGET_INDEX = 8
 
 ped_actions = ['slow down', 'moving slow', 'standing', 'stopped',
                'speed up', 'moving fast', 'look', 'looking',  'clear path',
@@ -129,16 +131,7 @@
 # OPTIM_C = Adam(lr=0.0000002, beta_1=0.5)
 OPTIM_C = SGD(lr=0.0001, momentum=0.9, nesterov=True)
 
-lr_schedule = [5, 10, 15]  # epoch_step
-
-# def schedule(epoch_idx):
-#     if (epoch_idx + 1) < lr_schedule[0]:
-#         return 0.0000001
-#     elif (epoch_idx + 1) < lr_schedule[1]:
-#         return 0.000000001  # lr_decay_ratio = 10
-#     elif (epoch_idx + 1) < lr_schedule[2]:
-#         return 0.0000000001
-#     return 0.0000000001
+lr_schedule = [50, 55, 60]  # epoch_step
 
 def schedule(epoch_idx):
     if (epoch_idx + 1) < lr_schedule[0]:
@@ -147,4 +140,13 @@ def schedule(epoch_idx):
         return 0.00001  # lr_decay_ratio = 10
     elif (epoch_idx + 1) < lr_schedule[2]:
         return 0.000001
-    return 0.000001
+    return 0.000001
+
+# def schedule(epoch_idx):
+#     if (epoch_idx + 1) < lr_schedule[0]:
+#         return 0.000001
+#     elif (epoch_idx + 1) < lr_schedule[1]:
+#         return 0.0000001  # lr_decay_ratio = 10
+#     elif (epoch_idx + 1) < lr_schedule[2]:
+#         return 0.00000001
+#     return 0.00000001

code/autoencoder_model/gifs/vid_25.gif

@@ -24,13 +24,13 @@ def strip(image, img_size, vid_len, vid_num=1):
     for i in range(n_vertical_imgs):
         for j in range(n_horizontal_imgs):
             img = image[i*img_size:(i+1)*img_size, j*img_size:(j+1)*img_size]
-            if (j>=10):
-                # img = cv2.blur(img, (5, 5), 0)
-                img = cv2.medianBlur(img, 7)
-                # img = cv2.bilateralFilter(img, 9, 150, 25)
-                img = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)  # convert it to hsv
-                img[:, :, 2] -= 2
-                img = cv2.cvtColor(img, cv2.COLOR_HSV2BGR)
+            # if (j>=10):
+            #     # img = cv2.blur(img, (5, 5), 0)
+            #     img = cv2.medianBlur(img, 7)
+            #     # img = cv2.bilateralFilter(img, 9, 150, 25)
+            #     img = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)  # convert it to hsv
+            #     img[:, :, 2] -= 2
+            #     img = cv2.cvtColor(img, cv2.COLOR_HSV2BGR)
             filename = "vid_" + str(vid_num) + "_frame_" + str(frame_num) + ".png"
             cv2.imwrite(os.path.join(GIF_IMG_DIR, filename), img)
             filenames.append(os.path.join(GIF_IMG_DIR, filename))