Create femnist.py

femnist.py

@@ -0,0 +1,133 @@
+import argparse
+import json
+import matplotlib.pyplot as plt
+import math
+import numpy as np
+import os
+
+from scipy import io
+from scipy import stats
+from tqdm import tqdm 
+import math
+import PIL.Image as Img
+import uuid
+
+def show_img(image):
+    image = np.multiply(image, 255)
+    image = image.reshape(int(math.sqrt(len(image))), int(math.sqrt(len(image))))
+    image = Img.fromarray(image)
+    image = image.convert('L')
+    image.save('outfile.png')
+    
+def gen_data_dict(data_dict, raw_data, group_name):
+    data_dict[group_name] = []
+    for _, value in raw_data.items():
+        for img_data, img_label in zip(value['x'], value['y']):
+            data_dict[group_name].append({
+                'image': np.multiply(img_data, 255),
+                'label': img_label
+            })
+
+def construct_data_dirs_clientwise(raw_data, group_name, data_dir, data_type='train'):
+    data_group_dir = os.path.join(data_dir, 'huggingface', 'clientwise', group_name, data_type)
+    if not os.path.exists(data_group_dir):
+        os.makedirs(data_group_dir)
+
+    for _, value in raw_data.items():
+        for img_data, img_label in zip(value['x'], value['y']):
+            sample_dir = os.path.join(data_group_dir, str(img_label))
+            if not os.path.exists(sample_dir):
+                os.makedirs(sample_dir)
+            sampel_path = os.path.join(sample_dir, str(uuid.uuid1())[:8]+'.png')
+
+            image = np.multiply(img_data, 255)
+            image = image.reshape(int(math.sqrt(len(image))), int(math.sqrt(len(image))))
+            image = Img.fromarray(image)
+            image = image.convert('L')
+            image.save(sampel_path)
+
+def construct_data_dirs(raw_data, data_dir, data_type='train'):
+    data_group_dir = os.path.join(data_dir, 'huggingface', 'centralized', data_type)
+    if not os.path.exists(data_group_dir):
+        os.makedirs(data_group_dir)
+
+    for _, value in raw_data.items():
+        for img_data, img_label in zip(value['x'], value['y']):
+            sample_dir = os.path.join(data_group_dir, str(img_label))
+            if not os.path.exists(sample_dir):
+                os.makedirs(sample_dir)
+            sampel_path = os.path.join(sample_dir, str(uuid.uuid1())[:8]+'.png')
+
+            image = np.multiply(img_data, 255)
+            image = image.reshape(int(math.sqrt(len(image))), int(math.sqrt(len(image))))
+            image = Img.fromarray(image)
+            image = image.convert('L')
+            image.save(sampel_path)
+
+def load_data(name):
+    train_users = []
+    train_num_samples = []
+    train_data = {}
+
+    test_users = []
+    test_num_samples = []
+    test_data = {}
+    
+    parent_path = os.path.dirname(os.path.realpath(__file__))
+    data_dir = os.path.join(parent_path, 'dataset', name)
+    train_subdir = os.path.join(data_dir, 'train')
+    test_subdir = os.path.join(data_dir, 'test')
+
+    # load train
+    train_files = os.listdir(train_subdir)
+    train_files = [f for f in train_files if f.endswith('.json')]
+
+    for index, f in tqdm(enumerate(train_files), desc='Training Data Generating', total=len(train_files)):
+        group_name = 'client_' + str(index)
+        file_dir = os.path.join(train_subdir, f)
+
+        with open(file_dir) as inf:
+            data = json.load(inf)
+
+        # show image and print label
+        # show_img(data['user_data']['f3242_19']['x'][0])
+        # print("Label: " + str(data['user_data']['f3242_19']['y'][0]))
+
+        # train_users.extend(data['users'])
+        train_num_samples.extend([sum(data['num_samples'])])
+        gen_data_dict(train_data, data['user_data'], group_name)
+        construct_data_dirs_clientwise(data['user_data'], group_name, data_dir, data_type='train')
+        construct_data_dirs(data['user_data'], data_dir, data_type='train')
+
+    # load test
+    test_files = os.listdir(test_subdir)
+    test_files = [f for f in test_files if f.endswith('.json')]
+
+    for index, f in tqdm(enumerate(test_files), desc='Testing Data Generating', total=len(test_files)):
+        group_name = 'client_' + str(index)
+        file_dir = os.path.join(test_subdir, f)
+
+        with open(file_dir) as inf:
+            data = json.load(inf)
+
+        test_users.extend(data['users'])
+        test_num_samples.extend([sum(data['num_samples'])])
+        gen_data_dict(test_data, data['user_data'], group_name)
+        construct_data_dirs_clientwise(data['user_data'], group_name, data_dir, data_type='test')
+        construct_data_dirs(data['user_data'], data_dir, data_type='test')
+
+    return train_num_samples, train_data, test_num_samples, test_data
+
+if __name__ == '__main__':
+    name = 'femnist-small'
+    train_num_samples, train_data, \
+    test_num_samples, test_data = load_data(name)
+
+    print('####################################')
+    print('DATASET: %s' % name)
+    print('%d train samples (total)' % np.sum(train_num_samples))
+    print('%d test samples (total)' % np.sum(test_num_samples))
+    print('%.2f train samples per user (mean)' % np.mean(train_num_samples))
+    print('%.2f test samples per user (mean)' % np.mean(test_num_samples))
+
+