The prefix for symbols not exported from the library and not
local to one translation unit is ff_ (or FF for types).
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
'void *' is too flexible, since we can derive info from
AVFilterContext*, so we just unify the interface with this data
structure.
Signed-off-by: Xie, Lin <lin.xie@intel.com>
Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
function fill_model_input_ov and infer_completion_callback are
extracted, it will help the async execution for reuse.
Signed-off-by: Xie, Lin <lin.xie@intel.com>
Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
TensorFlow C library accepts config for session options to
set different parameters for the inference. This patch exports
this interface.
The config is a serialized tensorflow.ConfigProto proto, so we need
two steps to use it:
1. generate the serialized proto with python (see script example below)
the output looks like: 0xab...cd
where 0xcd is the least significant byte and 0xab is the most significant byte.
2. pass the python script output into ffmpeg with
dnn_processing=options=sess_config=0xab...cd
The following script is an example to specify one GPU. If the system contains
3 GPU cards, the visible_device_list could be '0', '1', '2', '0,1' etc.
'0' does not mean physical GPU card 0, we need to try and see.
And we can also add more opitions here to generate more serialized proto.
script example to generate serialized proto which specifies one GPU:
import tensorflow as tf
gpu_options = tf.GPUOptions(visible_device_list='0')
config = tf.ConfigProto(gpu_options=gpu_options)
s = config.SerializeToString()
b = ''.join("%02x" % int(ord(b)) for b in s[::-1])
print('0x%s' % b)
for some cases (for example, super resolution), the DNN model changes
the frame size which impacts the filter behavior, so the filter needs
to know the out frame size at very beginning.
Currently, the filter reuses DNNModule.execute_model to query the
out frame size, it is not clear from interface perspective, so add
a new explict interface DNNModel.get_output for such query.
suppose we have a detect and classify filter in the future, the
detect filter generates some bounding boxes (BBox) as AVFrame sidedata,
and the classify filter executes DNN model for each BBox. For each
BBox, we need to crop the AVFrame, copy data to DNN model input and do
the model execution. So we have to save the in_frame at DNNModel.set_input
and use it at DNNModule.execute_model, such saving is not feasible
when we support async execute_model.
This patch sets the in_frame as execution_model parameter, and so
all the information are put together within the same function for
each inference. It also makes easy to support BBox async inference.
Currently, every filter needs to provide code to transfer data from
AVFrame* to model input (DNNData*), and also from model output
(DNNData*) to AVFrame*. Actually, such transfer can be implemented
within DNN module, and so filter can focus on its own business logic.
DNN module also exports the function pointer pre_proc and post_proc
in struct DNNModel, just in case that a filter has its special logic
to transfer data between AVFrame* and DNNData*. The default implementation
within DNN module is used if the filter does not set pre/post_proc.
currently, output is set both at DNNModel.set_input_output and
DNNModule.execute_model, it makes sense that the output name is
provided at model inference time so all the output info is set
at a single place.
and so DNNModel.set_input_output is renamed to DNNModel.set_input
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
different backend might need different options for a better performance,
so, add the parameter into dnn interface, as a preparation.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
We can try with the srcnn model from sr filter.
1) get srcnn.pb model file, see filter sr
2) convert srcnn.pb into openvino model with command:
python mo_tf.py --input_model srcnn.pb --data_type=FP32 --input_shape [1,960,1440,1] --keep_shape_ops
See the script at https://github.com/openvinotoolkit/openvino/tree/master/model-optimizer
We'll see srcnn.xml and srcnn.bin at current path, copy them to the
directory where ffmpeg is.
I have also uploaded the model files at https://github.com/guoyejun/dnn_processing/tree/master/models
3) run with openvino backend:
ffmpeg -i input.jpg -vf format=yuv420p,scale=w=iw*2:h=ih*2,dnn_processing=dnn_backend=openvino:model=srcnn.xml:input=x:output=srcnn/Maximum -y srcnn.ov.jpg
(The input.jpg resolution is 720*480)
Also copy the logs on my skylake machine (4 cpus) locally with openvino backend
and tensorflow backend. just for your information.
$ time ./ffmpeg -i 480p.mp4 -vf format=yuv420p,scale=w=iw*2:h=ih*2,dnn_processing=dnn_backend=tensorflow:model=srcnn.pb:input=x:output=y -y srcnn.tf.mp4
…
frame= 343 fps=2.1 q=31.0 Lsize= 2172kB time=00:00:11.76 bitrate=1511.9kbits/s speed=0.0706x
video:1973kB audio:187kB subtitle:0kB other streams:0kB global headers:0kB muxing overhead: 0.517637%
[aac @ 0x2f5db80] Qavg: 454.353
real 2m46.781s
user 9m48.590s
sys 0m55.290s
$ time ./ffmpeg -i 480p.mp4 -vf format=yuv420p,scale=w=iw*2:h=ih*2,dnn_processing=dnn_backend=openvino:model=srcnn.xml:input=x:output=srcnn/Maximum -y srcnn.ov.mp4
…
frame= 343 fps=4.0 q=31.0 Lsize= 2172kB time=00:00:11.76 bitrate=1511.9kbits/s speed=0.137x
video:1973kB audio:187kB subtitle:0kB other streams:0kB global headers:0kB muxing overhead: 0.517640%
[aac @ 0x31a9040] Qavg: 454.353
real 1m25.882s
user 5m27.004s
sys 0m0.640s
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
OpenVINO is a Deep Learning Deployment Toolkit at
https://github.com/openvinotoolkit/openvino, it supports CPU, GPU
and heterogeneous plugins to accelerate deep learning inferencing.
Please refer to https://github.com/openvinotoolkit/openvino/blob/master/build-instruction.md
to build openvino (c library is built at the same time). Please add
option -DENABLE_MKL_DNN=ON for cmake to enable CPU path. The header
files and libraries are installed to /usr/local/deployment_tools/inference_engine/
with default options on my system.
To build FFmpeg with openvion, take my system as an example, run with:
$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/deployment_tools/inference_engine/lib/intel64/:/usr/local/deployment_tools/inference_engine/external/tbb/lib/
$ ../ffmpeg/configure --enable-libopenvino --extra-cflags=-I/usr/local/deployment_tools/inference_engine/include/ --extra-ldflags=-L/usr/local/deployment_tools/inference_engine/lib/intel64
$ make
Here are the features provided by OpenVINO inference engine:
- support more DNN model formats
It supports TensorFlow, Caffe, ONNX, MXNet and Kaldi by converting them
into OpenVINO format with a python script. And torth model
can be first converted into ONNX and then to OpenVINO format.
see the script at https://github.com/openvinotoolkit/openvino/tree/master/model-optimizer/mo.py
which also does some optimization at model level.
- optimize at inference stage
It optimizes for X86 CPUs with SSE, AVX etc.
It also optimizes based on OpenCL for Intel GPUs.
(only Intel GPU supported becuase Intel OpenCL extension is used for optimization)
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
vf_dnn_processing.c recently changed to use swscale to trasfer data
between AVFrame and dnn model.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Linjie Fu <linjie.fu@intel.com>
The Y channel is handled by dnn, and also resized by dnn. The UV channels
are resized with swscale.
The command to use espcn.pb (see vf_sr) looks like:
./ffmpeg -i 480p.jpg -vf format=yuv420p,dnn_processing=dnn_backend=tensorflow:model=espcn.pb:input=x:output=y -y tmp.espcn.jpg
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Reviewed-by: Pedro Arthur <bygrandao@gmail.com>
Only the Y channel is handled by dnn, the UV channels are copied
without changes.
The command to use srcnn.pb (see vf_sr) looks like:
./ffmpeg -i 480p.jpg -vf format=yuv420p,scale=w=iw*2:h=ih*2,dnn_processing=dnn_backend=tensorflow:model=srcnn.pb:input=x:output=y -y srcnn.jpg
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Reviewed-by: Pedro Arthur <bygrandao@gmail.com>
The following is a python script to halve the value of the gray
image. It demos how to setup and execute dnn model with python+tensorflow.
It also generates .pb file which will be used by ffmpeg.
import tensorflow as tf
import numpy as np
from skimage import color
from skimage import io
in_img = io.imread('input.jpg')
in_img = color.rgb2gray(in_img)
io.imsave('ori_gray.jpg', np.squeeze(in_img))
in_data = np.expand_dims(in_img, axis=0)
in_data = np.expand_dims(in_data, axis=3)
filter_data = np.array([0.5]).reshape(1,1,1,1).astype(np.float32)
filter = tf.Variable(filter_data)
x = tf.placeholder(tf.float32, shape=[1, None, None, 1], name='dnn_in')
y = tf.nn.conv2d(x, filter, strides=[1, 1, 1, 1], padding='VALID', name='dnn_out')
sess=tf.Session()
sess.run(tf.global_variables_initializer())
graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['dnn_out'])
tf.train.write_graph(graph_def, '.', 'halve_gray_float.pb', as_text=False)
print("halve_gray_float.pb generated, please use \
path_to_ffmpeg/tools/python/convert.py to generate halve_gray_float.model\n")
output = sess.run(y, feed_dict={x: in_data})
output = output * 255.0
output = output.astype(np.uint8)
io.imsave("out.jpg", np.squeeze(output))
To do the same thing with ffmpeg:
- generate halve_gray_float.pb with the above script
- generate halve_gray_float.model with tools/python/convert.py
- try with following commands
./ffmpeg -i input.jpg -vf format=grayf32,dnn_processing=model=halve_gray_float.model:input=dnn_in:output=dnn_out:dnn_backend=native out.native.png
./ffmpeg -i input.jpg -vf format=grayf32,dnn_processing=model=halve_gray_float.pb:input=dnn_in:output=dnn_out:dnn_backend=tensorflow out.tf.png
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
do not request AVFrame's format in vf_ddn_processing with 'fmt',
but to add another filter for the format.
command examples:
./ffmpeg -i input.jpg -vf format=bgr24,dnn_processing=model=halve_first_channel.model:input=dnn_in:output=dnn_out:dnn_backend=native -y out.native.png
./ffmpeg -i input.jpg -vf format=rgb24,dnn_processing=model=halve_first_channel.model:input=dnn_in:output=dnn_out:dnn_backend=native -y out.native.png
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
This filter accepts all the dnn networks which do image processing.
Currently, frame with formats rgb24 and bgr24 are supported. Other
formats such as gray and YUV will be supported next. The dnn network
can accept data in float32 or uint8 format. And the dnn network can
change frame size.
The following is a python script to halve the value of the first
channel of the pixel. It demos how to setup and execute dnn model
with python+tensorflow. It also generates .pb file which will be
used by ffmpeg.
import tensorflow as tf
import numpy as np
import imageio
in_img = imageio.imread('in.bmp')
in_img = in_img.astype(np.float32)/255.0
in_data = in_img[np.newaxis, :]
filter_data = np.array([0.5, 0, 0, 0, 1., 0, 0, 0, 1.]).reshape(1,1,3,3).astype(np.float32)
filter = tf.Variable(filter_data)
x = tf.placeholder(tf.float32, shape=[1, None, None, 3], name='dnn_in')
y = tf.nn.conv2d(x, filter, strides=[1, 1, 1, 1], padding='VALID', name='dnn_out')
sess=tf.Session()
sess.run(tf.global_variables_initializer())
output = sess.run(y, feed_dict={x: in_data})
graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['dnn_out'])
tf.train.write_graph(graph_def, '.', 'halve_first_channel.pb', as_text=False)
output = output * 255.0
output = output.astype(np.uint8)
imageio.imsave("out.bmp", np.squeeze(output))
To do the same thing with ffmpeg:
- generate halve_first_channel.pb with the above script
- generate halve_first_channel.model with tools/python/convert.py
- try with following commands
./ffmpeg -i input.jpg -vf dnn_processing=model=halve_first_channel.model:input=dnn_in:output=dnn_out:fmt=rgb24:dnn_backend=native -y out.native.png
./ffmpeg -i input.jpg -vf dnn_processing=model=halve_first_channel.pb:input=dnn_in:output=dnn_out:fmt=rgb24:dnn_backend=tensorflow -y out.tf.png
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
If filter changes frame w/h, AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC
cannot be supported.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
to support dnn networks more general, we need to know the input info
of the dnn model.
background:
The data type of dnn model's input could be float32, uint8 or fp16, etc.
And the w/h of input image could be fixed or variable.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
so, we can make a filter more general to accept different network
models, by adding a data type convertion after getting data from network.
After we add dt field into struct DNNData, it becomes the same as
DNNInputData, so merge them with one struct: DNNData.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
Unlike other tf.*.conv2d layers, tf.nn.conv2d does not create many
nodes (within a scope) in the graph, it just acts like other layers.
tf.nn.conv2d only creates one node in the graph, and no internal
nodes such as 'kernel' are created.
The format of native model file is also changed, a flag named
has_bias is added, so change the version number.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
and also change CONV to DLT_CONV2D for better description
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
The reason to add this layer is that it is used by srcnn in vf_sr.
This layer is currently ignored in native mode. After this patch,
we can add multiple outputs support for native mode.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
the logic is that one layer in one separated source file to make
the source files simple for maintaining.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
the logic is that one layer in one separated source file to make
the source files simple for maintaining.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
currently, the layer number is at the beginning of the .model file,
so we have to scan twice in python script, the first scan to get the
layer number. Only one scan needed after put the layer number at the
end of .model file.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
the info can be saved in dnn operand object without regenerating again and again,
and it is also needed for layer split/merge, and for memory reuse.
to make things step by step, this patch just focuses on c code,
the change within python script will be added later.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
background:
DNN (deep neural network) is a sub module of libavfilter, and FATE/dnn
is unit test for the DNN module, one unit test for one dnn layer.
The unit tests are not based on the APIs exported by libavfilter,
they just directly call into the functions within DNN submodule.
There is an issue when run the following command:
build$ ../ffmpeg/configure --disable-static --enable-shared
make
make fate-dnn-layer-pad
And part of error message:
tests/dnn/dnn-layer-pad-test.o: In function `test_with_mode_symmetric':
/work/media/ffmpeg/build/src/tests/dnn/dnn-layer-pad-test.c:73: undefined reference to `dnn_execute_layer_pad'
The root cause is that function dnn_execute_layer_pad is a LOCAL symbol
in libavfilter.so, and so the linker could not find it when build dnn-layer-pad-test.
To check it, just run: readelf -s libavfilter/libavfilter.so | grep dnn
So, add dependency in fate/dnn Makefile with ffmpeg static libraries.
This is the same method used in fate/checkasm
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
if it is initialized randomly, the tensorflow lib will report
error message such as:
Attempt to add output -7920 of depth_to_space4 not in range [0, 1) to node with type Identity
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
conv2d with dilation > 1 generates tens of nodes in graph, it is not
easy to parse each node one by one, so we do special tricks to parse
the conv2d layer.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
since tf.pad is enabled, the conv2d(valid) changes back to its original behavior.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
the reason to add this layer first is that vf_sr uses it in its
tensorflow model, and the next plan is to update the python script
to convert tf.pad into native model.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
it is expected that there will be more files to support native mode,
so put all the dnn codes under libavfilter/dnn
The main change of this patch is to move the file location, see below:
modified: libavfilter/Makefile
new file: libavfilter/dnn/Makefile
renamed: libavfilter/dnn_backend_native.c -> libavfilter/dnn/dnn_backend_native.c
renamed: libavfilter/dnn_backend_native.h -> libavfilter/dnn/dnn_backend_native.h
renamed: libavfilter/dnn_backend_tf.c -> libavfilter/dnn/dnn_backend_tf.c
renamed: libavfilter/dnn_backend_tf.h -> libavfilter/dnn/dnn_backend_tf.h
renamed: libavfilter/dnn_interface.c -> libavfilter/dnn/dnn_interface.c
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
For example, given TensorFlow model file espcn.pb,
to generate native model file espcn.model, just run:
python convert.py espcn.pb
In current implementation, the native model file is generated for
specific dnn network with hard-code python scripts maintained out of ffmpeg.
For example, srcnn network used by vf_sr is generated with
https://github.com/HighVoltageRocknRoll/sr/blob/master/generate_header_and_model.py#L85
In this patch, the script is designed as a general solution which
converts general TensorFlow model .pb file into .model file. The script
now has some tricky to be compatible with current implemention, will
be refined step by step.
The script is also added into ffmpeg source tree. It is expected there
will be many more patches and community needs the ownership of it.
Another technical direction is to do the conversion in c/c++ code within
ffmpeg source tree. While .pb file is organized with protocol buffers,
it is not easy to do such work with tiny c/c++ code, see more discussion
at http://ffmpeg.org/pipermail/ffmpeg-devel/2019-May/244496.html. So,
choose the python script.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
currently, only float is supported as model input, actually, there
are other data types, this patch adds uint8.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
some models such as ssd, yolo have more than one output.
the clean up code in this patch is a little complex, it is because
that set_input_output_tf could be called for many times together
with ff_dnn_execute_model_tf, we have to clean resources for the
case that the two interfaces are called interleaved.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
Currently, within interface set_input_output, the dims/memory of the tensorflow
dnn model output is determined by executing the model with zero input,
actually, the output dims might vary with different input data for networks
such as object detection models faster-rcnn, ssd and yolo.
This patch moves the logic from set_input_output to execute_model which
is suitable for all the cases. Since interface changed, and so dnn_backend_native
also changes.
In vf_sr.c, it knows it's srcnn or espcn by executing the model with zero input,
so execute_model has to be called in function config_props
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
remove the requirment that the name of DNN model input/output
should be "x"/"y",
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
remove 'else' since there is always 'return' in 'if' scope,
so the code will be clean for later maintenance
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
otherwise, the following check will return error if layer_add_res
is randomly initialized.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>
This patch just enables the path from ffmpeg to libx264,
the more encoders can be added later.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
The encoders such as libx264 support different QPs offset for different MBs,
it makes possible for ROI-based encoding. It makes sense to add support
within ffmpeg to generate/accept ROI infos and pass into encoders.
Typical usage: After AVFrame is decoded, a ffmpeg filter or user's code
generates ROI info for that frame, and the encoder finally does the
ROI-based encoding.
The ROI info is maintained as side data of AVFrame.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
Guo, Yejun
