From 6e6b9b5cd6f27f626616844e1ec31088515de9e8 Mon Sep 17 00:00:00 2001
From: Danila Fedorin <danila.fedorin@gmail.com>
Date: Sun, 22 Dec 2019 23:17:28 -0800
Subject: [PATCH] Change model, still broken

---
 model.py | 20 +++++++++++---------
 1 file changed, 11 insertions(+), 9 deletions(-)

diff --git a/model.py b/model.py
index 79304a9..bb90caf 100644
--- a/model.py
+++ b/model.py
@@ -42,6 +42,8 @@ class SimpleCNN(object):
                             for i in range(2)])
 
             return tf.map_fn(per_output_tensor, output_tensor)
+        truth_tensor = tf.transpose(truth_tensor, perm=(1,2,0))
+        output_tensor = tf.transpose(output_tensor, perm=(1,2,0))
 
         # Compute per object IOU values for each square, for each box.
         ious = input_output_tensor(self.iou)
@@ -57,7 +59,7 @@ class SimpleCNN(object):
         eq_min_box = tf.map_fn(lambda iou:
             tf.convert_to_tensor([equal(iou[j], min_class_ious) for j in range(2)]), ious, dtype='bool')
         # Same as above, but per-square rather than per-box.
-        eq_min_square= any(eq_min_box, axis=1)
+        eq_min_square = any(eq_min_box, axis=1)
 
         # Whether each box of each square is responsible
         # for the maximum IOU. This is used for penalizing
@@ -65,7 +67,7 @@ class SimpleCNN(object):
         eq_max_box = tf.map_fn(lambda iou:
             tf.convert_to_tensor([equal(iou[j], max_class_ious) for j in range(2)]), ious, dtype='bool')
         # Same as above, but per-square. Penalizes bad class guesses.
-        eq_max_square= any(eq_max_box, axis=1)
+        eq_max_square = any(eq_max_box, axis=1)
 
         # The cost of incorrect coordinate guesses per box.
         coord_cost = input_output_tensor(
@@ -79,7 +81,7 @@ class SimpleCNN(object):
         # The cost of incorrect class guesses, per square.
         class_cost = tf.map_fn(lambda output:
                 tf.map_fn(lambda truth:
-                    tf.norm(output[2*5:2*5+6]-truth[5:12]), truth_tensor), output_tensor)
+                    sum(pow(output[2*5:2*5+6]-truth[5:12],2), axis=0), truth_tensor), output_tensor)
 
         # Weights from the YOLO paper.
         coord_weight = 5
@@ -92,15 +94,15 @@ class SimpleCNN(object):
         noobj_cost= noobj_weight * confidence_cost
         # Cost per box, selecting only "responsible" entries.
         box_cost = (
-            obj_cost* cast(eq_max_box, 'float32') + 
-            noobj_cost* cast(eq_min_box, 'float32')
+            obj_cost* tf.cast(eq_max_box, tf.float32) + 
+            noobj_cost* tf.cast(eq_min_box, tf.float32)
         )
 
         # Cost per square, penalizing only "responsible" squares.
-        square_cost= class_cost * cast(eq_max_square, 'float32')
+        square_cost = class_cost * tf.cast(eq_max_square, tf.float32)
 
         # Total cost
-        cost = sum(sum(sum(box_cost))) + sum(sum(square_cost))
+        cost = sum(box_cost) + sum(square_cost)
         return cost
         
     def build_model(self):
@@ -118,7 +120,7 @@ class SimpleCNN(object):
         self.model.add(Flatten())
         self.model.add(Dense(units=self.squares*self.squares*(self.boxes*5+self.classes)))
         self.model.add(Reshape((64,-1)))
-        self.model.compile(loss=self.loss, optimizer=Adam(lr=0.5e-4, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0))
+        self.model.compile(loss=self.cost, optimizer=Adam(lr=0.5e-5, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0))
 
     def maxpool_layer(self, *args, **kwargs):
         self.model.add(MaxPooling2D(*args, **kwargs))
@@ -143,7 +145,7 @@ def image_generator(source, n):
 
 inputs = []
 outputs = []
-for (inp, out) in image_generator('data/test', 100):
+for (inp, out) in image_generator('data/test', 200):
     inputs.append(inp)
     outputs.append(out)