extra experiment for omniglot and cross_char

backbone.py

@@ -32,7 +32,7 @@ def forward(self, x):
         L_norm = torch.norm(self.L.weight.data, p=2, dim =1).unsqueeze(1).expand_as(self.L.weight.data)
         self.L.weight.data = self.L.weight.data.div(L_norm + 0.00001)
         cos_dist = self.L(x_normalized) #matrix product by forward function
-        scores = 2* (cos_dist) #a fixed scale factor to scale the output of cos value into a reasonably large input for softmax
+        scores = 10* (cos_dist) #a fixed scale factor to scale the output of cos value into a reasonably large input for softmax
 
         return scores
 

configs.py

@@ -1,4 +1,6 @@
-save_dir                    = './record/'
+save_dir                    = '/work/newriver/wyharveychen/CloserLookFewShot/'
 data_dir = {}
 data_dir['CUB']             = './filelists/CUB/' 
 data_dir['miniImagenet']    = './filelists/miniImagenet/' 
+data_dir['omniglot']        = './filelists/omniglot/' 
+data_dir['emnist']          = './filelists/emnist/' 

filelists/emnist/download_emnist.sh

@@ -0,0 +1,5 @@
+#!/usr/bin/env bash
+wget https://github.com/NanqingD/DAOSL/raw/master/data/emnist.zip 
+unzip emnist.zip
+python invert_emnist.py
+python write_cross_char_valnovel_filelist.py

filelists/emnist/invert_emnist.py

@@ -0,0 +1,33 @@
+import numpy as np
+from os import listdir
+from os.path import isfile, isdir, join
+import os
+import json
+import random
+from PIL import Image
+import PIL.ImageOps
+
+cwd = os.getcwd() 
+data_path = join(cwd,'emnist')
+inv_data_path = join(cwd,'inv_emnist')
+savedir = './'
+
+#if not os.path.exists(savedir):
+#    os.makedirs(savedir)
+if not os.path.exists(inv_data_path):
+    os.makedirs(inv_data_path)
+
+character_folder_list = [str(i) for i in range(62)] #lazy_hack 
+
+classfile_list_all = []
+
+for character_folder in character_folder_list:
+    character_folder_path = join(data_path, character_folder)        
+    inv_character_folder_path = join(inv_data_path, character_folder)        
+    image_list =  [ img for img in listdir(character_folder_path) if (isfile(join(character_folder_path,img)) and img[0] != '.')]
+    if not os.path.exists(inv_character_folder_path):
+        os.makedirs(inv_character_folder_path)
+    for img in image_list:                
+        inverted_img =PIL.ImageOps.invert(Image.open(join(character_folder_path,img)))
+        inverted_img.save(join(inv_character_folder_path ,img))
+

filelists/emnist/write_cross_char_valnovel_filelist.py

@@ -0,0 +1,59 @@
+import numpy as np
+from os import listdir
+from os.path import isfile, isdir, join
+import os
+import json
+import random
+
+cwd = os.getcwd() 
+data_path = join(cwd,'inv_emnist')
+savedir = './'
+dataset_list = ['val','novel']
+
+#if not os.path.exists(savedir):
+#    os.makedirs(savedir)
+
+folder_list = [str(i) for i in range(62)] #lazy_hack
+label_dict = dict(zip(folder_list,range(0,len(folder_list))))
+
+classfile_list_all = []
+
+for i, folder in enumerate(folder_list):
+    folder_path = join(data_path, folder)
+    classfile_list_all.append( [ join(folder_path, cf) for cf in listdir(folder_path) if (isfile(join(folder_path,cf)) and cf[0] != '.')])
+    random.shuffle(classfile_list_all[i])
+
+for dataset in dataset_list:
+    file_list = []
+    label_list = []
+    for i, classfile_list in enumerate(classfile_list_all):
+        if 'val' in dataset:
+            if (i%2 == 0):
+                file_list = file_list + classfile_list
+                label_list = label_list + np.repeat(i, len(classfile_list)).tolist()
+        if 'novel' in dataset:
+            if (i%2 == 1):
+                file_list = file_list + classfile_list
+                label_list = label_list + np.repeat(i, len(classfile_list)).tolist()
+
+    fo = open(savedir + dataset + ".json", "w")
+    fo.write('{"label_names": [')
+    fo.writelines(['"%s",' % item  for item in folder_list])
+    fo.seek(0, os.SEEK_END) 
+    fo.seek(fo.tell()-1, os.SEEK_SET)
+    fo.write('],')
+
+    fo.write('"image_names": [')
+    fo.writelines(['"%s",' % item  for item in file_list])
+    fo.seek(0, os.SEEK_END) 
+    fo.seek(fo.tell()-1, os.SEEK_SET)
+    fo.write('],')
+
+    fo.write('"image_labels": [')
+    fo.writelines(['%d,' % item  for item in label_list])
+    fo.seek(0, os.SEEK_END) 
+    fo.seek(fo.tell()-1, os.SEEK_SET)
+    fo.write(']}')
+
+    fo.close()
+    print("%s -OK" %dataset)

filelists/omniglot/download_omniglot.sh

@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+wget https://raw.githubusercontent.com/jakesnell/prototypical-networks/master/data/omniglot/splits/vinyals/train.txt 
+wget https://raw.githubusercontent.com/jakesnell/prototypical-networks/master/data/omniglot/splits/vinyals/val.txt
+wget https://raw.githubusercontent.com/jakesnell/prototypical-networks/master/data/omniglot/splits/vinyals/test.txt
+
+DATADIR=./images
+mkdir -p $DATADIR
+wget -O images_background.zip https://github.com/brendenlake/omniglot/blob/master/python/images_background.zip?raw=true
+wget -O images_evaluation.zip https://github.com/brendenlake/omniglot/blob/master/python/images_evaluation.zip?raw=true
+unzip images_background.zip -d $DATADIR
+unzip images_evaluation.zip -d $DATADIR
+mv $DATADIR/images_background/* $DATADIR/
+mv $DATADIR/images_evaluation/* $DATADIR/
+rmdir $DATADIR/images_background
+rmdir $DATADIR/images_evaluation
+
+python rot_omniglot.py
+python write_omniglot_filelist.py
+python write_cross_char_base_filelist.py

filelists/omniglot/rot_omniglot.py

@@ -0,0 +1,37 @@
+import numpy as np
+from os import listdir
+from os.path import isfile, isdir, join
+import os
+import json
+import random
+from PIL import Image
+
+cwd = os.getcwd() 
+data_path = join(cwd,'images')
+savedir = './'
+
+#if not os.path.exists(savedir):
+#    os.makedirs(savedir)
+
+language_folder_list = [f for f in listdir(data_path) if isdir(join(data_path, f))]
+language_folder_list.sort()
+
+classfile_list_all = []
+
+for language_folder in language_folder_list:
+    language_folder_path = join(data_path, language_folder)
+    character_folder_list = [cf for cf in listdir(language_folder_path) if isdir(join(language_folder_path, cf))]
+    character_folder_list.sort()
+    for character_folder in character_folder_list:
+        character_folder_path = join(language_folder_path, character_folder)        
+        image_list =  [ img for img in listdir(character_folder_path) if (isfile(join(character_folder_path,img)) and img[0] != '.')]
+        for deg in [0,90,180,270]:
+            rot_str = "rot%03d"%deg
+            rot_character_path = join(character_folder_path, rot_str)
+            print(rot_character_path)
+            if not os.path.exists(rot_character_path):
+                os.makedirs(rot_character_path)
+            for img in image_list:                
+                rot_img = Image.open(join(character_folder_path,img)).rotate(deg)
+                rot_img.save(join(character_folder_path,rot_str,img))
+

filelists/omniglot/write_cross_char_base_filelist.py

@@ -0,0 +1,64 @@
+import numpy as np
+from os import listdir
+from os.path import isfile, isdir, join
+import os
+import json
+import random
+import re
+
+cwd = os.getcwd() 
+data_path = join(cwd,'images')
+savedir = './'
+
+#if not os.path.exists(savedir):
+#    os.makedirs(savedir)
+
+cl = -1
+folderlist = []
+
+language_folder_list = [f for f in listdir(data_path) if isdir(join(data_path, f))]
+language_folder_list.sort()
+
+filelists = {}
+
+for language_folder in language_folder_list:
+    if language_folder == 'Latin':
+        continue
+    language_folder_path = join(data_path, language_folder)
+    character_folder_list = [cf for cf in listdir(language_folder_path) if isdir(join(language_folder_path, cf))]
+    character_folder_list.sort()
+    for character_folder in character_folder_list:
+        character_folder_path = join(language_folder_path, character_folder)
+        label = join(language_folder,character_folder)
+        folderlist.append(label)
+        filelists[label] =  [ join(character_folder_path,img) for img in listdir(character_folder_path) if (isfile(join(character_folder_path,img)) and img[-3:] == 'png')]
+
+filelists_flat = []
+labellists_flat = []
+for key, filelist in filelists.items():
+    cl += 1
+    random.shuffle(filelist)
+    filelists_flat += filelist
+    labellists_flat += np.repeat(cl, len(filelist)).tolist() 
+
+fo = open(join(savedir, "noLatin.json"), "w")
+fo.write('{"label_names": [')
+fo.writelines(['"%s",' % item  for item in folderlist])
+fo.seek(0, os.SEEK_END) 
+fo.seek(fo.tell()-1, os.SEEK_SET)
+fo.write('],')
+
+fo.write('"image_names": [')
+fo.writelines(['"%s",' % item  for item in filelists_flat])
+fo.seek(0, os.SEEK_END) 
+fo.seek(fo.tell()-1, os.SEEK_SET)
+fo.write('],')
+
+fo.write('"image_labels": [')
+fo.writelines(['%d,' % item  for item in labellists_flat])
+fo.seek(0, os.SEEK_END) 
+fo.seek(fo.tell()-1, os.SEEK_SET)
+fo.write(']}')
+
+fo.close()
+print("noLatin -OK")

filelists/omniglot/write_omniglot_filelist.py

@@ -0,0 +1,59 @@
+import numpy as np
+from os import listdir
+from os.path import isfile, isdir, join
+import os
+import json
+import random
+import re
+
+cwd = os.getcwd() 
+data_path = join(cwd,'images')
+savedir = './'
+dataset_list = ['base', 'val', 'novel']
+
+#if not os.path.exists(savedir):
+#    os.makedirs(savedir)
+
+cl = -1
+folderlist = []
+
+datasetmap = {'base':'train','val':'val','novel':'test'};
+filelists = {'base':{},'val':{},'novel':{} }
+filelists_flat = {'base':[],'val':[],'novel':[] }
+labellists_flat = {'base':[],'val':[],'novel':[] }
+
+for dataset in dataset_list:
+    with open(datasetmap[dataset] + ".txt", "r") as lines:
+        for i, line in enumerate(lines):
+            label = line.replace('\n','')
+            folderlist.append(label)
+            filelists[dataset][label] = [ join(data_path,label, f) for f in listdir( join(data_path, label))] 
+
+    for key, filelist in filelists[dataset].items():
+        cl += 1
+        random.shuffle(filelist)
+        filelists_flat[dataset] += filelist
+        labellists_flat[dataset] += np.repeat(cl, len(filelist)).tolist() 
+
+for dataset in dataset_list:
+    fo = open(savedir + dataset + ".json", "w")
+    fo.write('{"label_names": [')
+    fo.writelines(['"%s",' % item  for item in folderlist])
+    fo.seek(0, os.SEEK_END) 
+    fo.seek(fo.tell()-1, os.SEEK_SET)
+    fo.write('],')
+
+    fo.write('"image_names": [')
+    fo.writelines(['"%s",' % item  for item in filelists_flat[dataset]])
+    fo.seek(0, os.SEEK_END) 
+    fo.seek(fo.tell()-1, os.SEEK_SET)
+    fo.write('],')
+
+    fo.write('"image_labels": [')
+    fo.writelines(['%d,' % item  for item in labellists_flat[dataset]])
+    fo.seek(0, os.SEEK_END) 
+    fo.seek(fo.tell()-1, os.SEEK_SET)
+    fo.write(']}')
+
+    fo.close()
+    print("%s -OK" %dataset)

io_utils.py

@@ -6,6 +6,7 @@
 
 model_dict = dict(
             Conv4 = backbone.Conv4,
+            Conv4S = backbone.Conv4S,
             Conv6 = backbone.Conv6,
             ResNet10 = backbone.ResNet10,
             ResNet18 = backbone.ResNet18,
@@ -15,16 +16,16 @@
 
 def parse_args(script):
     parser = argparse.ArgumentParser(description= 'few-shot script %s' %(script))
-    parser.add_argument('--dataset'     , default='CUB',        help='CUB/ miniImagenet/cross')
+    parser.add_argument('--dataset'     , default='CUB',        help='CUB/miniImagenet/cross/omniglot/cross_char')
     parser.add_argument('--model'       , default='Conv4',      help='model: Conv{4|6} / ResNet{10|18|34|50|101}') # 50 and 101 are not used in the paper
     parser.add_argument('--method'      , default='baseline',   help='baseline/baseline++/protonet/matchingnet/relationnet{_softmax}/maml{_approx}') #relationnet_softmax replace L2 norm with softmax to expedite training, maml_approx use first-order approximation in the gradient for efficiency
-    parser.add_argument('--train_n_way' , default=5, type=int,  help='class num to classify for training')
-    parser.add_argument('--test_n_way'  , default=5, type=int,  help='class num to classify for testing (validation) ')
-    parser.add_argument('--n_shot'      , default=5, type=int,  help='number of labeled data in each class, same as n_support')
-    parser.add_argument('--train_aug'   , action='store_true',  help='perform data augmentation or not during training ')
+    parser.add_argument('--train_n_way' , default=5, type=int,  help='class num to classify for training') #baseline and baseline++ would ignore this parameter
+    parser.add_argument('--test_n_way'  , default=5, type=int,  help='class num to classify for testing (validation) ') #baseline and baseline++ only use this parameter in finetuning
+    parser.add_argument('--n_shot'      , default=5, type=int,  help='number of labeled data in each class, same as n_support') #baseline and baseline++ only use this parameter in finetuning
+    parser.add_argument('--train_aug'   , action='store_true',  help='perform data augmentation or not during training ') #still required for save_features.py and test.py to find the model path correctly
 
     if script == 'train':
-        parser.add_argument('--num_classes' , default=200, type=int, help='total number of classes in softmax, only used in baseline')
+        parser.add_argument('--num_classes' , default=200, type=int, help='total number of classes in softmax, only used in baseline') #make it larger than the maximum label value in base class
         parser.add_argument('--save_freq'   , default=50, type=int, help='Save frequency')
         parser.add_argument('--start_epoch' , default=0, type=int,help ='Starting epoch')
         parser.add_argument('--stop_epoch'  , default=400, type=int, help ='Stopping epoch') # for meta-learning methods, each epoch contains 100 episodes
@@ -35,7 +36,7 @@ def parse_args(script):
         parser.add_argument('--save_iter', default=-1, type=int,help ='save feature from the model trained in x epoch, use the best model if x is -1')
     elif script == 'test':
         parser.add_argument('--split'       , default='novel', help='base/val/novel') #default novel, but you can also test base/val class accuracy if you want 
-        parser.add_argument('--save_iter', default=-1, type=int,help ='save feature from the model trained in x epoch, use the best model if x is -1') #please match the one used in save_features
+        parser.add_argument('--save_iter', default=-1, type=int,help ='saved feature from the model trained in x epoch, use the best model if x is -1')
         parser.add_argument('--adaptation'  , action='store_true', help='further adaptation in test time or not')
     else:
        raise ValueError('Unknown script')

save_features.py

@@ -45,16 +45,28 @@ def save_features(model, data_loader, outfile ):
     assert params.method != 'maml' and params.method != 'maml_approx', 'maml do not support save_feature and run'
 
     if 'Conv' in params.model:
-        image_size = 84 
+        if params.dataset in ['omniglot', 'cross_char']:
+            image_size = 28
+        else:
+            image_size = 84 
     else:
         image_size = 224
 
+    if params.dataset in ['omniglot', 'cross_char']:
+        assert params.model == 'Conv4' and not params.train_aug ,'omniglot only support Conv4 without augmentation'
+        params.model = 'Conv4S'
+
     split = params.split
     if params.dataset == 'cross':
         if split == 'base':
             loadfile = configs.data_dir['miniImagenet'] + 'all.json' 
         else:
             loadfile   = configs.data_dir['CUB'] + split +'.json' 
+    elif params.dataset == 'cross_char':
+        if split == 'base':
+            loadfile = configs.data_dir['omniglot'] + 'noLatin.json' 
+        else:
+            loadfile  = configs.data_dir['emnist'] + split +'.json' 
     else:
         loadfile = configs.data_dir[params.dataset] + split + '.json'
 
@@ -84,6 +96,8 @@ def save_features(model, data_loader, outfile ):
             model = backbone.Conv4NP()
         elif params.model == 'Conv6': 
             model = backbone.Conv6NP()
+        elif params.model == 'Conv4S': 
+            model = backbone.Conv4SNP()
         else:
             model = model_dict[params.model]( flatten = False )
     elif params.method in ['maml' , 'maml_approx']: