104061213林倢愷HW1

data.py

@@ -0,0 +1,98 @@
+import sys, os
+import numpy as np
+import pdb
+import warnings
+
+def load_one_label_seq(path):
+    npy = np.load(path)
+    return npy
+
+
+def load_label_seqs(path, mode, index):#index are generated by gen_index
+    labels=[]
+    for i in range(len(index)): #eg index = [(lab,1),(lab,2),(lab,3),...]
+        loc = index[i][0] #eg lab office house
+        idx = index[i][1] #eg 1,2,3
+        labelnpy = os.path.join(path,loc,mode+"_left"+str(idx)+'.npy')
+        labels.append(load_one_label_seq(labelnpy))
+        #labelnpy = os.path.join(path,loc,mode+"_right"+str(idx)+'.npy')
+        #labels.append(load_one_label_seq(labelnpy))
+    for i in range(len(index)):
+        loc = index[i][0]
+        idx = index[i][1]
+        #labelnpy = os.path.join(path,loc,mode+"_left"+str(idx)+'.npy')
+        #labels.append(load_one_label_seq(labelnpy))
+        labelnpy = os.path.join(path,loc,mode+"_right"+str(idx)+'.npy')
+        labels.append(load_one_label_seq(labelnpy))
+    return labels #left1~3 right1~3
+
+def gen_index(setting_index):
+    train_index=[]
+    test_index =[]
+    if setting_index == 0:
+        #order: 
+            #train : lab1~4  off1~3  house1~3
+            #test  : lab5~8  off4~6  house4~6
+        for i in range(1,7):
+            if i <= 3:
+                train_index.append(('house',i))
+            else:
+                test_index.append(('house',i))
+        for i in range(1,9):
+            if i <= 4:
+                train_index.append(('lab',i))
+            else:
+                test_index.append(('lab',i))
+        for i in range(1,7):
+            if i <= 3:
+                train_index.append(('office',i))
+            else:
+                test_index.append(('office',i))
+
+    elif setting_index == 1:
+        for i in range(1,9):
+            train_index.append(('lab',i))
+        for i in range(1,7):
+            train_index.append(('office',i))
+        for i in range(1,7):
+            test_index.append(('house',i))
+    else:
+        raise ValueError ('error setting index')
+
+    return train_index, test_index
+
+
+
+def gen_index_process(index=None, setting_index=None):
+    if index == None:
+        if setting_index==None:
+            raise ValueError('Setting index can not be none')
+        else:
+            train_index, test_index = gen_index(setting_index)
+    return train_index, test_index
+
+
+def load_train_labels(path, mode, index=None, setting_index=None):
+    if index == None:
+        index,_ = gen_index_process(index,setting_index)
+    else:
+        if setting_index != None:
+            warnings.warn('setting_index has no effect when given particular index')
+    labels = load_label_seqs(path, mode, index)
+    return labels
+
+def load_test_labels(path, mode, index=None, setting_index=None):
+    if index == None:
+        _,index = gen_index_process(index,setting_index)
+    else:
+        if setting_index != None:
+            warnings.warn('setting_index has no effect when given particular index')
+    labels = load_label_seqs(path, mode, index)
+    return labels
+
+
+def load_all_labels(path, mode, setting_index):
+    train_index, test_index = gen_index(setting_index)
+    train_labels = load_train_labels(path, mode,train_index)
+    test_labels = load_train_labels(path, mode,test_index)
+    return train_labels, test_labels

results/index.md

@@ -1,47 +1,43 @@
-# Your Name <span style="color:red">(id)</span>
+# Your Name <span style="color:red">104061213 林倢愷</span>
 
-#Project 5: Deep Classification
+#Project 1: Deep Classification
 
 ## Overview
-The project is related to 
-> quote
-
+The project is related to finetuning VGG16. 
+Reference 
+https://gist.github.com/omoindrot/dedc857cdc0e680dfb1be99762990c9c
+https://github.com/bgshih/vgg16.tf
 
 ## Implementation
-1. One
-	* item
-	* item
+1. One load in data 
 2. Two
-
-```
-Code highlights
-```
+	load in pretrained VGG16
+	restore VGG16 weight
+	delete FC8 layer cause for our own classification task.
+3. Three
+	finetune from pretrained VGG16
+
+VGG architecture 
+![](http://book.paddlepaddle.org/03.image_classification/image/vgg16.png)
+original VGG paper : https://arxiv.org/pdf/1409.1556.pdf
+
+finetune 
 
 ## Installation
-* Other required packages.
-* How to compile from source?
+Tensorflow, numpy, scipy...etc
 
 ### Results
+test accuracy : 0.48 (I will train a pretrain resnet50 to surpass 0.6)
+
+### discussion
+原本預計使用pretrained model就可以達到還不錯的成績，
+但是最後的結果很慘，
+所以來分析一下，
+大部分純classification的task用pretrained model+finetune都可以達到還不錯的結果，
+我結果差我認為有兩個可能。
+1. 本身dataset跟imagenet差異太大，之前看過一個分析是目前用imagenet來train的model都有明顯的缺點在小的物體、薄的物體、半透明的物體．．．等，而老師的dataset大部分的物體都很小，因此model辨識不出來，所以導致result很差。
+2. finetune不夠遠，我freeze住除了FC8跟最後兩個conv以外的所有層數，根據上一點，應該要finetune更遠，可能一半的network都finetune才可以達到比較好的result，原本finetune不遠的考量是dataset不是很大，所以才決定只finetune最後幾層。
+
 
-<table border=1>
-<tr>
-<td>
-<img src="placeholder.jpg" width="24%"/>
-<img src="placeholder.jpg"  width="24%"/>
-<img src="placeholder.jpg" width="24%"/>
-<img src="placeholder.jpg" width="24%"/>
-</td>
-</tr>
-
-<tr>
-<td>
-<img src="placeholder.jpg" width="24%"/>
-<img src="placeholder.jpg"  width="24%"/>
-<img src="placeholder.jpg" width="24%"/>
-<img src="placeholder.jpg" width="24%"/>
-</td>
-</tr>
-
-</table>