Add image classification models #52

wwhu · 2017-05-24T09:51:33Z

resolve #28
The classification accuracy has not been validated yet.

qingqing01

还需要使用文档，可以参考其他models里的文档，但模型介绍部分觉得不用详细讲了，都是比较经典的模型，book里都有介绍，可以指向链接指向book。

qingqing01 · 2017-05-25T08:15:48Z

image_classification/googlenet.py

+
+def inception(name, input, channels, filter1, filter3R, filter3, filter5R,
+              filter5, proj):
+    cov1 = paddle.layer.conv_projection(


这里采用这个配置里inception2 的写法吧，/~https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/paddle/image/googlenet.py#L19

conv_projection不适应CPU.

已添加inception2

直接去掉inception，只保留inception2吧~

qingqing01 · 2017-05-25T08:22:45Z

image_classification/resnet.py

+    res1 = layer_warp(block_func, pool1, 64, stages[0], 1)
+    res2 = layer_warp(block_func, res1, 128, stages[1], 2)
+    res3 = layer_warp(block_func, res2, 256, stages[2], 2)
+    res4 = layer_warp(block_func, res3, 512, stages[3], 2)


对于上面 " # TODO: bug fix for ch_in = input.num_filters "，一种办法是这里layer_warp这里可以显示的指定通道数。当然也可以等配置解析重写之后，看能不能获取layer的属性，再改也行~~

输入通道数的用处是判断是否需要在block之间进行projection，我这里直接通过b_projection参数显式指定是否需要projection

明白了，这样也可以，那上面的 TODO 可以删掉了吧~

qingqing01 · 2017-05-25T08:28:24Z

image_classification/train.py

        learning_rate_decay_a=0.1,
        learning_rate_decay_b=128000 * 35,
        learning_rate_schedule="discexp", )

    train_reader = paddle.batch(
-        paddle.reader.shuffle(reader.test_reader("train.list"), buf_size=1000),
+        paddle.reader.shuffle(
+            reader.test_reader(os.path.join(args.data_dir, 'train.list')),


等haoshuang的PR合入之后，需要换成flowers的数据，用加速的reader。同时文档里，可以告诉用户如果换成自己的数据集如何处理~

已在文档里说明如何表示数据

qingqing01 · 2017-05-25T08:32:19Z

image_classification/train.py

+        loss2 = paddle.layer.cross_entropy_cost(
+            input=out2, label=lbl, coeff=0.3)
+        paddle.evaluator.classification_error(input=out2, label=lbl)
+        extra_layers = [loss1, loss2]


要么把net, cost这些代码都放到各自的配置里去？这样train.py看着清爽一些~

把net放到各自配置里了，因为考虑到infer阶段不需要cost，所以cost还是放到了train.py

qingqing01 · 2017-06-01T11:27:14Z

image_classification/googlenet.py

+
+def inception(name, input, channels, filter1, filter3R, filter3, filter5R,
+              filter5, proj):
+    cov1 = paddle.layer.conv_projection(


直接去掉inception，只保留inception2吧~

qingqing01 · 2017-06-01T11:31:49Z

image_classification/README.md

+                                                     BATCH_SIZE),
+    learning_rate=0.001 / BATCH_SIZE,
+    learning_rate_decay_a=0.1,
+    learning_rate_decay_b=128000 * 35,


learning_rate_decay_a， learning_rate_decay_b解释下吧，参见book里的解释~

qingqing01 · 2017-06-01T11:38:53Z

image_classification/resnet.py

+    res1 = layer_warp(block_func, pool1, 64, stages[0], 1)
+    res2 = layer_warp(block_func, res1, 128, stages[1], 2)
+    res3 = layer_warp(block_func, res2, 256, stages[2], 2)
+    res4 = layer_warp(block_func, res3, 512, stages[3], 2)


明白了，这样也可以，那上面的 TODO 可以删掉了吧~

wwhu · 2017-06-02T06:08:14Z

@qingqing01 这几处已修改

qingqing01

再加个infer.py吧。

qingqing01 · 2017-06-05T06:38:28Z

image_classification/resnet.py

+
+
+def layer_warp(block_func, input, features, count, stride):
+    conv = block_func(input, features, stride, True)


第一个block_func的b_projection不是总为True的，还是的依赖于输入conv的channel数。所以我觉得还是shortcut里对判断输入和输出的channel数是否一致，决定是否用conv做升降维度好些~

qingqing01 · 2017-06-06T11:42:24Z

flowers dataset已经merge了，需要换下~

guoshengCS · 2017-06-08T10:05:51Z

能否稍微调整下basicblock和bottleneck中shortcut的顺序，将shortcut放在前面，保证拓扑顺序与原文尽可能一致，thx~

wwhu · 2017-06-13T11:12:15Z

添加了infer.py以及flowers数据，并调整了shortcut位置。
@qingqing01 @guoshengCS

qingqing01 · 2017-06-13T11:31:56Z

image_classification/infer.py

+        # image in RGB mode. It must swap the channel order.
+        im = im[(2, 1, 0), :, :]  # BGR
+        im = im.flatten()
+        im = im / 255.0


这里看下image.py里有没有函数可以直接用。而且这里 im/255.0和训练不对应吧，训练是减去均值吧。

qingqing01 · 2017-06-13T11:34:15Z

image_classification/googlenet.py

+__all__ = ['googlenet']
+
+
+def inception2(name, input, channels, filter1, filter3R, filter3, filter5R,


inception2 -> inception

qingqing01 · 2017-06-13T11:36:57Z

image_classification/vgg.py

@@ -17,7 +17,7 @@
 __all__ = ['vgg13', 'vgg16', 'vgg19']


-def vgg(input, nums):
+def vgg(input, nums, class_dim=100):


class_dim=100 -> class_dim，去掉100的默认值吧

qingqing01 · 2017-06-13T11:43:17Z

image_classification/resnet.py

+    return conv
+
+
+def resnet_imagenet(input, depth=50, class_dim=100):


class_dim=100 -> class_dim，去掉默认值100吧。

所有的默认值都去掉了

lcy-seso · 2017-06-14T23:53:38Z

If you are not going to finish this work. please tell me.

wwhu · 2017-06-15T00:38:28Z

I will finish it ASAP. Sorry for the delay. @lcy-seso

lcy-seso · 2017-06-15T00:44:36Z

@wwhu You are welcome. I think after some modifications, we can try to merge the already finished part and then based on a merged version, refactor, refine, and validate the learning performance. Also, thanks for your work.

qingqing01

LGTM. 精度需要进一步确保~

add v2 API for imagenet models

6dd3895

wwhu requested a review from qingqing01 May 24, 2017 09:51

qingqing01 reviewed May 25, 2017

View reviewed changes

add doc and reorginize net output

8848164

qingqing01 reviewed Jun 1, 2017

View reviewed changes

minor revision

d7d1ae5

qingqing01 reviewed Jun 5, 2017

View reviewed changes

add infer.py and flower dataset

0116bc8

wwhu added 2 commits June 13, 2017 19:34

fix bug for resnet_cifar10 and adjust learning rate

208ca38

fix bug

e9b94ca

qingqing01 reviewed Jun 13, 2017

View reviewed changes

wwhu added 2 commits June 15, 2017 10:21

add xmap for image list and modify the image reader of infer.py

bdffa40

resolve conflict

cf9cf4c

qingqing01 approved these changes Jun 15, 2017

View reviewed changes

lcy-seso merged commit 8e80e20 into PaddlePaddle:develop Jun 15, 2017

wwhu deleted the in-dev branch June 15, 2017 08:15

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add image classification models #52

Add image classification models #52

wwhu commented May 24, 2017

qingqing01 left a comment

qingqing01 May 25, 2017

wwhu Jun 1, 2017

qingqing01 Jun 1, 2017

qingqing01 May 25, 2017

wwhu Jun 1, 2017 •

edited

Loading

qingqing01 Jun 1, 2017

qingqing01 May 25, 2017

wwhu Jun 1, 2017

qingqing01 May 25, 2017

wwhu Jun 1, 2017

qingqing01 Jun 1, 2017

qingqing01 Jun 1, 2017

qingqing01 Jun 1, 2017

wwhu commented Jun 2, 2017

qingqing01 left a comment

qingqing01 Jun 5, 2017

wwhu Jun 13, 2017

qingqing01 commented Jun 6, 2017

guoshengCS commented Jun 8, 2017

wwhu commented Jun 13, 2017

qingqing01 Jun 13, 2017

wwhu Jun 15, 2017

qingqing01 Jun 13, 2017

qingqing01 Jun 13, 2017

qingqing01 Jun 13, 2017

wwhu Jun 15, 2017

lcy-seso commented Jun 14, 2017

wwhu commented Jun 15, 2017

lcy-seso commented Jun 15, 2017 •

edited

Loading

qingqing01 left a comment



		def layer_warp(block_func, input, features, count, stride):
		conv = block_func(input, features, stride, True)

		__all__ = ['googlenet']


		def inception2(name, input, channels, filter1, filter3R, filter3, filter5R,

		return conv


		def resnet_imagenet(input, depth=50, class_dim=100):

Add image classification models #52

Add image classification models #52

Conversation

wwhu commented May 24, 2017

qingqing01 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wwhu Jun 1, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wwhu commented Jun 2, 2017

qingqing01 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

qingqing01 commented Jun 6, 2017

guoshengCS commented Jun 8, 2017

wwhu commented Jun 13, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

lcy-seso commented Jun 14, 2017

wwhu commented Jun 15, 2017

lcy-seso commented Jun 15, 2017 • edited Loading

qingqing01 left a comment

Choose a reason for hiding this comment

wwhu Jun 1, 2017 •

edited

Loading

lcy-seso commented Jun 15, 2017 •

edited

Loading