revise fine tuning and cifar

chapter_computer-vision/fine-tuning.md

@@ -28,29 +28,29 @@
 我们首先将数据下载到`../data`。在当前目录解压后得到`hotdog/train`和`hotdog/test`这两个文件夹。每个下面有`hotdog`和`not-hotdog`这两个类别文件夹，里面是对应的图片文件。
 
 ```{.python .input  n=4}
-%matplotlib inline
 import sys
 sys.path.insert(0, '..')
 import zipfile
 import gluonbook as gb
-from mxnet import nd, image, gluon, init
+from mxnet import nd, gluon, init
+from mxnet.gluon import data as gdata, loss as gloss, model_zoo, utils as gutils
 from mxnet.gluon.data.vision import transforms
 
 data_dir = '../data/'
 base_url = 'https://apache-mxnet.s3-accelerate.amazonaws.com/'
-fname = gluon.utils.download(
+fname = gutils.download(
     base_url+'gluon/dataset/hotdog.zip',
     path=data_dir, sha1_hash='fba480ffa8aa7e0febbb511d181409f899b9baa5')
 
-with zipfile.ZipFile(fname, 'r') as f:
-    f.extractall(data_dir)
+with zipfile.ZipFile(fname, 'r') as z:
+    z.extractall(data_dir)
 ```
 
 我们使用使用`ImageFolderDataset`类来读取数据。它将每个类别文件夹当做一个类，并读取下面所有的图片。
 
 ```{.python .input  n=6}
-train_imgs = gluon.data.vision.ImageFolderDataset(data_dir+'/hotdog/train')
-test_imgs = gluon.data.vision.ImageFolderDataset(data_dir+'/hotdog/test')
+train_imgs = gdata.vision.ImageFolderDataset(data_dir+'/hotdog/train')
+test_imgs = gdata.vision.ImageFolderDataset(data_dir+'/hotdog/test')
 ```
 
 下面画出前8张正例图片和最后的8张负例图片，可以看到他们性质和高宽各不相同。
@@ -113,15 +113,15 @@ finetune_net.output.initialize(init.Xavier())
 
 ```{.python .input  n=12}
 def train(net, learning_rate, batch_size=128, epochs=5):
-    train_data = gluon.data.DataLoader(
+    train_data = gdata.DataLoader(
         train_imgs.transform_first(train_augs), batch_size, shuffle=True)
-    test_data = gluon.data.DataLoader(
+    test_data = gdata.DataLoader(
         test_imgs.transform_first(test_augs), batch_size)
 
     ctx = gb.try_all_gpus()
     net.collect_params().reset_ctx(ctx)
     net.hybridize()
-    loss = gluon.loss.SoftmaxCrossEntropyLoss()
+    loss = gloss.SoftmaxCrossEntropyLoss()
     trainer = gluon.Trainer(net.collect_params(), 'sgd', {
         'learning_rate': learning_rate, 'wd': 0.001})
     gb.train(train_data, test_data, net, loss, trainer, ctx, epochs)
@@ -136,7 +136,7 @@ train(finetune_net, 0.01)
 为了对比起见，我们训练同样的一个模型，但所有参数都初始成随机值。我们使用较大的学习率来加速收敛。
 
 ```{.python .input  n=14}
-scratch_net = gluon.model_zoo.vision.resnet18_v2(classes=2)
+scratch_net = model_zoo.vision.resnet18_v2(classes=2)
 scratch_net.initialize(init=init.Xavier())
 train(scratch_net, 0.1)
 ```

chapter_computer-vision/kaggle-gluon-cifar10.md

@@ -53,6 +53,7 @@ import datetime
 import gluonbook as gb
 from mxnet import autograd, gluon, init, nd
 from mxnet.gluon import data as gdata, nn, loss as gloss
+from mxnet.gluon.data.vision import transforms
 import numpy as np
 import os
 import pandas as pd
@@ -64,9 +65,9 @@ import shutil
 demo = True
 if demo:
     import zipfile
-    for fin in ['train_tiny.zip', 'test_tiny.zip', 'trainLabels.csv.zip']:
-        with zipfile.ZipFile('../data/kaggle_cifar10/' + fin, 'r') as zin:
-            zin.extractall('../data/kaggle_cifar10/')
+    for f in ['train_tiny.zip', 'test_tiny.zip', 'trainLabels.csv.zip']:
+        with zipfile.ZipFile('../data/kaggle_cifar10/' + f, 'r') as z:
+            z.extractall('../data/kaggle_cifar10/')
 ```
 
 ### 整理数据集
@@ -150,31 +151,31 @@ reorg_cifar10_data(data_dir, label_file, train_dir, test_dir, input_dir,
 为避免过拟合，我们在这里使用`transforms`来增广数据集。例如我们加入`transforms.RandomFlipLeftRight()`即可随机对每张图片做镜面反转。我们也通过`transforms.Normalize()`对彩色图像RGB三个通道分别做[标准化](../chapter_supervised-learning/kaggle-gluon-kfold.md)。以下我们列举了所有可能用到的操作，这些操作可以根据需求来决定是否调用，它们的参数也都是可调的。
 
 ```{.python .input  n=4}
-transform_train = gdata.vision.transforms.Compose([
-    # gdata.vision.transforms.CenterCrop(32),
-    # gdata.vision.transforms.RandomFlipTopBottom(),
-    # gdata.vision.transforms.RandomColorJitter(brightness=0.0, contrast=0.0,
-    #                                           saturation=0.0, hue=0.0),
-    # gdata.vision.transforms.RandomLighting(0.0),
-    # gdata.vision.transforms.Cast('float32'),
-    # gdata.vision.transforms.Resize(32),
+transform_train = transforms.Compose([
+     transforms.CenterCrop(32),
+     transforms.RandomFlipTopBottom(),
+     transforms.RandomColorJitter(brightness=0.0, contrast=0.0,
+                                               saturation=0.0, hue=0.0),
+     transforms.RandomLighting(0.0),
+     transforms.Cast('float32'),
+     transforms.Resize(32),
 
     # 随机按照 scale 和 ratio 裁剪，并放缩为 32 x 32 的正方形。
-    gdata.vision.transforms.RandomResizedCrop(32, scale=(0.08, 1.0),
+    transforms.RandomResizedCrop(32, scale=(0.08, 1.0),
                                               ratio=(3.0/4.0, 4.0/3.0)),
     # 随机左右翻转图片。
-    gdata.vision.transforms.RandomFlipLeftRight(),
+    transforms.RandomFlipLeftRight(),
     # 将图片像素值缩小到（0, 1）内，并将数据格式从“高*宽*通道”改为“通道*高*宽”。
-    gdata.vision.transforms.ToTensor(),
+    transforms.ToTensor(),
     # 对图片的每个通道做标准化。
-    gdata.vision.transforms.Normalize([0.4914, 0.4822, 0.4465],
+    transforms.Normalize([0.4914, 0.4822, 0.4465],
                                       [0.2023, 0.1994, 0.2010])
 ])
 
 # 测试时，无需对图像做标准化以外的增强数据处理。
-transform_test = gdata.vision.transforms.Compose([
-    gdata.vision.transforms.ToTensor(),
-    gdata.vision.transforms.Normalize([0.4914, 0.4822, 0.4465],
+transform_test = transforms.Compose([
+    transforms.ToTensor(),
+    transforms.Normalize([0.4914, 0.4822, 0.4465],
                                       [0.2023, 0.1994, 0.2010])
 ])
 ```