From 1c6888f31270261367d89184c9fa91d1c83423cc Mon Sep 17 00:00:00 2001
From: Francesco Gatti <gattifrancesco@hotmail.it>
Date: Wed, 9 Aug 2017 14:13:17 +0000
Subject: [PATCH] auto download

---
 CMakeLists.txt | 15 ++++++++--
 README.md      | 74 ++++++++------------------------------------------
 2 files changed, 25 insertions(+), 64 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index a133f1c..802d2ad 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,7 +1,18 @@
 cmake_minimum_required(VERSION 2.8)
-
 project (tkDNN)
 
+set(BUILD_DEPS true CACHE BOOL "If true download deps")
+
+if( ${BUILD_DEPS} )
+	message("Launching pre-build dependency installer script...")
+
+    execute_process (COMMAND bash -c "bash build_models.sh download"
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/tests)
+
+	set(BUILD_DEPS false CACHE BOOL "If true download deps" FORCE)
+	message("Finished dowloading test weights")
+endif()
+
 find_package(CUDA QUIET REQUIRED)
 
 cuda_include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include ${CUDA_INCLUDE_DIRS})
@@ -31,4 +42,4 @@ add_executable(test_yolo tests/yolo/yolo.cpp)
 target_link_libraries(test_yolo tkDNN)
 
 add_executable(test_yolo_tiny tests/yolo-tiny/yolo-tiny.cpp)
-target_link_libraries(test_yolo_tiny tkDNN)
\ No newline at end of file
+target_link_libraries(test_yolo_tiny tkDNN)
diff --git a/README.md b/README.md
index 2ff77ef..0f08d93 100644
--- a/README.md
+++ b/README.md
@@ -1,15 +1,11 @@
 # tkDNN
 tkDNN is a Deep Neural Network library built with cuDNN primitives specifically thought to work on NVIDIA TK1 board.<br>
 The main scope is to do high performance inference on already trained models.
-Currently supports the following layers:
 
-* Dense, fully interconnected
-* Activation (RELU, ELU, SIGMOID, TANH)
-* Convolutional 2D
-* Convolutional 3D
-* Max and Average Pooling
-* Flatten 
-* Data preprocessing
+this branch is actually work on every NVIDIA GPU that support the dependencies:
+* CUDA 8
+* CUDNN 6
+* TENSORRT 2
 
 ## Workflow
 The recommended workflow follow these step:
@@ -26,60 +22,14 @@ cd build
 cmake ..
 make
 ```
+during the cmake configuration it will be dowloaded the weights needed for running
+the tests
 
 ## Test
-There is a ready to use example on *test* directory, to try it you must generate the weights with Keras
-```
-cd tests
-python test_model.py
-```
-And then execute the inference on build directory
-```
-cd build
-./tkDNNtest 
-```
-this should output the same prediction as Keras.
+Assumiung you have correctly builded the library these are the test ready to exec:
+* test_simple: a simple convolutional and dense network (CUDNN only)
+* test_mnist: the famous mnist netwok (CUDNN and TENSORRT)
+* test_mnistRT: the mnist network hardcoded in using tensorRT apis (TENSORRT only)
+* test_yolo: YOLO detection network (CUDNN and TENSORRT)
+* test_yolo_tiny: smaller version of YOLO (CUDNN and TENSRRT)
 
-## Simple example
-Here is a example of the entire workflow on a simple model.
-Using the following Keras model save it to a file
-```python
-model = Sequential()
-model.add(Reshape((20, 1), input_shape=(20)))
-model.add(Dense(256))
-model.compile()
-
-# save model
-model.save("path/to/model.h5")
-```
-
-After the model is created the weights can be exported for tkDNN inference
-```
-python weights_exporter model.h5 dense --output=weights/path
-```
-the exporter take as arguments, in order:
-* input model
-* layer type ["dense", "conv2d", conv3d"]
-* { layer type ["dense", "conv2d", conv3d"] for each layer to export }
-* optional argument --output define path where export weights   
-
-Then we can create a c++ program to do inference on tk1
-```c++
-#include<tkdnn.h>   //library include
-
-//Network object 
-tkDNN::Network net;
-//input dimension
-tkDNN::dataDim_t dim(1, 20, 1, 1, 1);
-//Dense layer
-tkDNN::Dense d0(&net, dim, 256, "weights/path", "bias/path");
-
-//here load the input data to CUDA
-//value_type is an alias of "float"
-value_type *data_d = [...] 
-
-//do inference
-value_type *output_d = d0.infer(dim, data_d);
-//dim will be updated with the output dimension
-``` 
-The result is finally stored on output_d in device memory.
\ No newline at end of file