Merge branch 'master' of https://github.com/dmlc/cxxnet

doc/README.md

@@ -21,5 +21,7 @@ In general, cxxnet configuration file contains four kinds of configurations in a
  - This page includes all the four tasks you could try by cxxnet.
 * [Other Setting](other.md)
  - Set other parameters for neural network, related to device selection, running control.
+* [Caffe Converter](caffe_converter.md)
+ - Convert the pretrained model in Caffe to cxxnet.
 * [Advanced Usages](advanced.md)
  - Some advanced usages of cxxnet can he found here.

doc/advanced.md

@@ -4,6 +4,7 @@ This page will introduce some advanced usages in cxxnet, including:
 
 #### Multi-label Training
 * To use multi-label training, you need the following three steps in additional to the case of single label training:
+ - For multi-label training, in ```imgrec```, you need to specify ```image_list``` field to indicate the list file that contains the labels.
  - First, you need to specify the number of labels in the network by setting ```label_width``` variable in global settings. The following setting denotes that we have 5 labels in the network.
  ```bash
  label_width = 5

doc/caffe_converter.md

@@ -0,0 +1,25 @@
+#### Introduction
+This page will introduce the usage of Caffe converter. It can convert a pretrained model in Caffe to cxxnet format.
+
+#### Preparation
+* To begin with the convert, a latest version of Caffe should be built. 
+* Currently, no automatic configuration file converter is provided. You need to convert the Caffe config file in prototxt to cxxnet configuration format by yourself. **Please make sure that all the layers in original Caffe model has corresponding layer in cxxnet!**
+* Converters are provided in both C++ and Python.
+ - To use the C++ converter, you need to specify the following paths in the config.mk. For example,
+
+ ```bash
+ # whether to build caffe converter
+ USE_CAFFE_CONVERTER = 1
+ CAFFE_ROOT = ~/caffe
+ CAFFE_INCLUDE = ~/caffe/include/
+ CAFFE_LIB = ~/caffe/build/lib/
+ ```
+
+ Then, run ```make all``` in the root of cxxnet. if everything is correct, you could find ```caffe_converter``` and ```caffe_mean_converter``` in the ```bin``` folder.
+
+ - To use the Python converter, you should first make sure the Python wrapper of Caffe is successfully built. Then you need to specify the paths of Caffe and cxxnet in ```tools/caffe_converter/convert.py```.
+
+#### Convert
+* Simply run '''bin/caffe_converter''' and '''tools/caffe_converter/convert.py''', and then follow the instructions.
+* To convert the mean file of Caffe, please use the C++ converter: ```caffe_mean_converter```. But we strongly recommend you to recompute the mean file in cxxnet due to the different data augmentation methods in Caffe and cxxnet.
+

doc/io.md

@@ -14,7 +14,7 @@ options 2 =
 ...
 iter = end
 ```
-* The basic iterator type is **mnist** , **image** , **imgbin**, **csv**
+* The basic iterator type is **mnist** , **image** , **imgrec**, **csv**
 * To use thread buffer, declare in this form
 ```bash
 iter = iterator_type
@@ -36,7 +36,7 @@ iter = end
 ```bash
 shuffle = 1
 ```
-* **shuffle** set 1 to shuffle the **training data**. Note that this option **does not** apply to  **imgbin** and **csv**.
+* **shuffle** set 1 to shuffle the **training data**. Note that this option **does not** apply to **csv**.
 
 =
 ### MNIST Iterator
@@ -83,12 +83,13 @@ A valid image list file is like the following (NO header):
 Image binary iterator aims to reduce to IO cost in random seek. It is especially useful when deal with large amount for data like in ImageNet.
 * Required field
 ```bash
+image_rec = path to the image binary file
+# The image list file can be removed, if only single label exists. It is a must in the multi-label case.
 image_list = path to the image list file
-image_bin = path to the image binary file
 ```
 * The **image_list** file is described [above](#image-list-file)
-* To generate **image_bin** file, you need to use the tool [im2bin](https://github.com/antinucleon/cxxnet/blob/master/tools/im2bin.cpp) in the tools folder.
-* You may check an example [here](https://github.com/antinucleon/cxxnet/blob/master/example/ImageNet/ImageNet.conf)
+* To generate **image_rec** file, you need to use the tool [im2rec](../tools/im2rec.cc) in the tools folder.
+* You may check examples [here](../example/ImageNet/)
 
 #### Realtime Preprocessing Option for Image/Image Binary
 ```bash

example/ImageNet/ImageNet.conf

@@ -4,21 +4,21 @@
 # The scheduling parameters is adapted from Caffe(http:https://caffe.berkeleyvision.org/)
 
 data = train
-iter = imgbin
- image_list = "../../NameList.train"
- image_bin = "../../TRAIN.BIN"
- image_root = "../../data/resize256/"
+iter = imgrec
+# image_list = "../../NameList.train"
+ image_rec = "../../TRAIN.BIN"
+# image_root = "../../data/resize256/"
  image_mean = "models/image_net_mean.bin"
  rand_crop=1
  rand_mirror=1
 iter = threadbuffer
 iter = end
 
 eval = test
-iter = imgbin
- image_list = "../../NameList.test"
- image_bin = "../../TEST.BIN"
- image_root = "../../data/resize256/"
+iter = imgrec
+# image_list = "../../NameList.test"
+ image_rec = "../../TEST.BIN"
+# image_root = "../../data/resize256/"
  image_mean = "models/image_net_mean.bin"
 # no random crop and mirror in test
 iter = end

example/ImageNet/Inception-BN.conf

@@ -21,9 +21,9 @@
 # 10.07% for top 5) from single crop.
 #
 data = train
-iter = imgbin
- image_list = "imagenet/train.lst"
- image_bin = "imagenet/train.bin"
+iter = imgrec
+# image_list = "imagenet/train.lst"
+ image_rec = "imagenet/train.bin"
  image_mean = "models/mean_224.bin"
  rand_crop=1
  rand_mirror=1
@@ -32,9 +32,9 @@ iter = threadbuffer
 iter = end
 
 eval = val
-iter = imgbin
- image_list = "imagenet/val.lst"
- image_bin = "imagenet/val.bin"
+iter = imgrec
+# image_list = "imagenet/val.lst"
+ image_rec = "imagenet/val.bin"
  image_mean = "models/mean_224.bin"
 #no random crop and mirror in test
 iter = end

example/ImageNet/README.md

@@ -10,7 +10,7 @@ This tutorial will guide you train your own super vision model. The default conf
 * If you want to still use a large batch with not enough RAM, You can set ```update_period=2``` and ```batch_size=128```, this means the parameter update is done every 2 batches, which is equivalent to ```batch_size=256```
 
 ### 0.Before you start
-Make sure you have downloaded the ImageNet training data. Resize the picture into size 256 * 256 *3 for later we will crop random 227 * 227 * 3 image while training.
+Make sure you have downloaded the ImageNet training data. You don't need to resize the images by yourself, currently ```im2rec``` could resize it automatically. You could check the promoting message of ```im2rec``` for details.
 
 ### 1.Make the image list
 After you get the data, you need to make a [image list file](../../doc/io.md#image-list-file) first. The format is
@@ -35,14 +35,16 @@ A sample file is provided here
 ```
 
 ### 2.Make the binary file
-Although you can use image iterator now. the disk random seek will make the training process extremely slow. So **you'd better generate binary file for training and use imgbin iterator** .
+Although you can use image iterator now. the disk random seek will make the training process extremely slow. So **you'd better generate binary file for training and use imgrec iterator** .
 
-To generate binary image, you need to use *im2bin* in the tool folder. The im2bin will take the path of _image list file_ you generated just now, _root path_ of the images and the _output file path_ as input. These processes usually take several hours, so be patient. :)
+To generate binary image, you need to use *im2rec* in the tool folder. The im2rec will take the path of _image list file_ you generated just now, _root path_ of the images and the _output file path_ as input. These processes usually take several hours, so be patient. :)
 
 A sample command:
 ```bash
-im2bin ./train.lst ./resized256_images/ TRAIN.BIN
+./bin/im2rec image.lst image_root_dir output.bin resize=256
 ```
+More details can be found by running ```./bin/im2rec```.
+
 ### 3.Set correct configuration file
 Change the iterator path in the [ImageNet.conf](ImageNet.conf) or [kaiming.conf](kaiming.conf) to point to your _image list file_ and _image binary file_ correctly, then just run as MNIST example. After about 20 round, you can see some reasonable result. We strongly recommend to use [kaiming.conf](kaiming.conf), since it could provide much better results than Alexnet, while keeping the time cost unchanged.
 By calling

example/ImageNet/kaiming.conf

@@ -4,9 +4,9 @@
 # J' model in the paper above
 
 data = train
-iter = imginst
- image_list = "train.lst"
- image_bin = "train.bin"
+iter = imgrec
+# image_list = "train.lst"
+ image_rec = "train.bin"
  image_mean = "mean_224.bin"
  rand_crop=1
  rand_mirror=1
@@ -17,9 +17,9 @@ iter = threadbuffer
 iter = end
 
 eval = val
-iter = imginst
- image_list = "val.lst"
- image_bin = "val.bin"
+iter = imgrec
+# image_list = "val.lst"
+ image_rec = "val.bin"
  image_mean = "mean_224.bin"
 # no random crop and mirror in test
 iter = end

src/cxxnet_main.cpp

@@ -132,7 +132,7 @@ class CXXNetLearnTask {
  if (!strcmp(name,"model_dir")) name_model_dir= val;
  if (!strcmp(name,"num_round" )) num_round = atoi(val);
  if (!strcmp(name,"max_round")) max_round = atoi(val);
- if (!strcmp(name, "silent")) silent = atoi(val);
+ if (!strcmp(name, "silent"))  silent = atoi(val);
  if (!strcmp(name, "task")) task = val;
  if (!strcmp(name, "dev")) {
  device = val;
@@ -431,8 +431,8 @@ class CXXNetLearnTask {
 #if MSHADOW_RABIT_PS
  is_root = rabit::GetRank() == 0;
  print_tracker = rabit::IsDistributed();
-#endif
  silent = !is_root;
+#endif
  time_t start = time(NULL);
  unsigned long elapsed = 0;
  if (continue_training == 0 && name_model_in == "NULL") {
@@ -470,22 +470,26 @@ class CXXNetLearnTask {
  if (++ sample_counter % print_step == 0) {
  elapsed = (long)(time(NULL) - start);
  if (!silent) {
-  std::ostringstream os;
-  os << "round " << std::setw(8) << start_counter - 1
-  << ":[" << std::setw(8) << sample_counter << "] " << elapsed << " sec elapsed";
-  if (print_tracker) {
- utils::TrackerPrint(os.str().c_str());
-  } else {
- printf("\r \r");
- printf("%s", os.str().c_str());
- fflush(stdout);
-  }
+  std::ostringstream os;
+  os << "round " << std::setw(8) << start_counter - 1
+  << ":[" << std::setw(8) << sample_counter << "] " << elapsed << " sec elapsed";
+  if (print_tracker) {
+ utils::TrackerPrint(os.str().c_str());
+  } else {
+ printf("\r \r");
+ printf("%s", os.str().c_str());
+ fflush(stdout);
+  }
  }
  }
  }
 
  if (test_io == 0) {
  std::ostringstream os;
+ if (silent) {
+ os << "Finish " << sample_counter << " epochs in "
+ << (long)(time(NULL) - start) << " sec. ";
+ }
  os << '[' << start_counter << ']';
  // handle only with eval_train = 1, but not val data
  if (itr_evals.size() == 0) {
@@ -498,9 +502,9 @@ class CXXNetLearnTask {
  utils::TrackerPrint(os.str());
  }
  elapsed = (unsigned long)(time(NULL) - start);
- if (is_root) {
-  this->SaveModel();
- }
+ if (is_root) {
+  this->SaveModel();
+ }
  }
 
  if (!silent) {

tools/network_maker/config_generator.py

@@ -246,7 +246,7 @@ def ConvFactory(nchannel, kernel_size=1, pad=0, stride = 1, bn = True, act = "re
 
 def DFS(layer, table, seq):
  out_layers = graph[layer]
- for l in out_layers:
+ for l in out_layers - table:
  DFS(l, table, seq)
  if layer not in table:
  table.add(layer)