eval using rd-cost

Author: wolverinn

README.md

@@ -36,17 +36,13 @@ The **loss** of our trained model on test set is: 3.1049
 
 The **accuracy** of each label predicted on test set is: 66.12%
 
-The best way to evaluate the model is to integrate the model into the HEVC encoder. I've conceived a pipeline:
+The best way to evaluate the model is to integrate the model into the HEVC encoder. I've realized a pipeline, see it in:
 
-1. When HEVC encoder starts to process a new frame with frame number FrameNumber, it calls a command: ```python use_model.py -f FrameNumber```
-2. The script ```use_model.py``` takes FrameNumber as input. It also reads the ```bitstream.cfg``` to get the YUV filename the HEVC is currently processing. If FrameNumber is 0, it first use FFmpeg to extract frames from the YUV file. Then, it processes certain frame, and for all the CTUs in this frame, it generates a ```CtuNumber.txt``` with a 16x16 matrix in it and store all the txt in a folder ```ctu```.
-3. When HEVC encoder starts to process the CTU numbered CtuNumber, it goes to the ```ctu``` folder, find ```CtuNumber.txt``` and read the depths. In ```xCompressCU()```, if it's not at the predicted depth, then skip ```xCheckRDCostIntra()``` function.
+[HEVC-deep-learning-pipeline](https://github.com/wolverinn/HEVC-deep-learning-pipeline)
 
-I've already realized ```use_model.py```. Find it in ```./model test pipeline```.
+Using this evaluating pipeline, we can compare the change in encoding time and BDBR at the same time.
 
-Using this evaluating pipeline, we can compare the encoding time and BDBR at the same time.
-
-I use a simpler approach to evaluate the increase in RD-cost for each YUV file. As ```xCompressCU()``` in HEVC encoder calculates the RD-cost exhaustively at each depth, we can get the RD-cost for every possible depth decision. Thus, we can realize comparison of RD-cost between the original encoder and the CNN model. See the ```model test pipeline``` folder for codes.
+Another way to evaluate the model is to compare the difference in RD-cost for each YUV file encoded. As ```xCompressCU()``` in HEVC encoder calculates the RD-cost exhaustively at each depth, we can get the RD-cost for every possible depth decision. Thus, we can realize comparison of RD-cost between the original encoder and the CNN model. See the ```test RD-cost``` folder for codes.
 
 The increase in **RD cost** of our model is: 2.1% (tested only on one YUV sequence)
 

test RD-cost/TAppEncoder.exe

@@ -0,0 +1,9 @@
+InputFile : E:\HM\trunk\workspace\test_cost\flower_352x288_20.yuv
+InputBitDepth : 8
+InputChromaFormat : 420
+FrameRate : 20
+FrameSkip : 0
+SourceWidth : 352
+SourceHeight : 288
+FramesToBeEncoded : 10000
+Level : 3.1

test RD-cost/config/bitstream.cfg

@@ -0,0 +1,101 @@
+#======== File I/O =====================
+BitstreamFile                 : .\\config\\str.bin
+ReconFile                     : .\\config\\rec.yuv
+
+#======== Profile ================
+Profile                       : main
+
+#======== Unit definition ================
+MaxCUWidth                    : 64          # Maximum coding unit width in pixel
+MaxCUHeight                   : 64          # Maximum coding unit height in pixel
+MaxPartitionDepth             : 4           # Maximum coding unit depth
+QuadtreeTULog2MaxSize         : 5           # Log2 of maximum transform size for
+                                            # quadtree-based TU coding (2...6)
+QuadtreeTULog2MinSize         : 2           # Log2 of minimum transform size for
+                                            # quadtree-based TU coding (2...6)
+QuadtreeTUMaxDepthInter       : 3
+QuadtreeTUMaxDepthIntra       : 3
+
+#======== Coding Structure =============
+IntraPeriod                   : 1           # Period of I-Frame ( -1 = only first)
+DecodingRefreshType           : 1           # Random Accesss 0:none, 1:CRA, 2:IDR, 3:Recovery Point SEI
+GOPSize                       : 1           # GOP Size (number of B slice = GOPSize-1)
+ReWriteParamSetsFlag          : 1           # Write parameter sets with every IRAP
+
+#=========== Motion Search =============
+FastSearch                    : 1           # 0:Full search  1:TZ search
+SearchRange                   : 64          # (0: Search range is a Full frame)
+HadamardME                    : 1           # Use of hadamard measure for fractional ME
+FEN                           : 1           # Fast encoder decision
+FDM                           : 1           # Fast Decision for Merge RD cost
+
+#======== Quantization =============
+QP                            : 37          # Quantization parameter(0-51)
+MaxDeltaQP                    : 0           # CU-based multi-QP optimization
+MaxCuDQPDepth                 : 0           # Max depth of a minimum CuDQP for sub-LCU-level delta QP
+DeltaQpRD                     : 0           # Slice-based multi-QP optimization
+RDOQ                          : 1           # RDOQ
+RDOQTS                        : 1           # RDOQ for transform skip
+
+#=========== Deblock Filter ============
+LoopFilterOffsetInPPS         : 1           # Dbl params: 0=varying params in SliceHeader, param = base_param + GOP_offset_param; 1 (default) =constant params in PPS, param = base_param)
+LoopFilterDisable             : 0           # Disable deblocking filter (0=Filter, 1=No Filter)
+LoopFilterBetaOffset_div2     : 0           # base_param: -6 ~ 6
+LoopFilterTcOffset_div2       : 0           # base_param: -6 ~ 6
+DeblockingFilterMetric        : 0           # blockiness metric (automatically configures deblocking parameters in bitstream). Applies slice-level loop filter offsets (LoopFilterOffsetInPPS and LoopFilterDisable must be 0)
+
+#=========== Misc. ============
+InternalBitDepth              : 8           # codec operating bit-depth
+
+#=========== Coding Tools =================
+SAO                           : 1           # Sample adaptive offset  (0: OFF, 1: ON)
+AMP                           : 1           # Asymmetric motion partitions (0: OFF, 1: ON)
+TransformSkip                 : 1           # Transform skipping (0: OFF, 1: ON)
+TransformSkipFast             : 1           # Fast Transform skipping (0: OFF, 1: ON)
+SAOLcuBoundary                : 0           # SAOLcuBoundary using non-deblocked pixels (0: OFF, 1: ON)
+
+#============ Slices ================
+SliceMode                : 0                # 0: Disable all slice options.
+                                            # 1: Enforce maximum number of LCU in an slice,
+                                            # 2: Enforce maximum number of bytes in an 'slice'
+                                            # 3: Enforce maximum number of tiles in a slice
+SliceArgument            : 1500             # Argument for 'SliceMode'.
+                                            # If SliceMode==1 it represents max. SliceGranularity-sized blocks per slice.
+                                            # If SliceMode==2 it represents max. bytes per slice.
+                                            # If SliceMode==3 it represents max. tiles per slice.
+
+LFCrossSliceBoundaryFlag : 1                # In-loop filtering, including ALF and DB, is across or not across slice boundary.
+                                            # 0:not across, 1: across
+
+#============ PCM ================
+PCMEnabledFlag                      : 0                # 0: No PCM mode
+PCMLog2MaxSize                      : 5                # Log2 of maximum PCM block size.
+PCMLog2MinSize                      : 3                # Log2 of minimum PCM block size.
+PCMInputBitDepthFlag                : 1                # 0: PCM bit-depth is internal bit-depth. 1: PCM bit-depth is input bit-depth.
+PCMFilterDisableFlag                : 0                # 0: Enable loop filtering on I_PCM samples. 1: Disable loop filtering on I_PCM samples.
+
+#============ Tiles ================
+TileUniformSpacing                  : 0                # 0: the column boundaries are indicated by TileColumnWidth array, the row boundaries are indicated by TileRowHeight array
+                                                       # 1: the column and row boundaries are distributed uniformly
+NumTileColumnsMinus1                : 0                # Number of tile columns in a picture minus 1
+TileColumnWidthArray                : 2 3              # Array containing tile column width values in units of CTU (from left to right in picture)   
+NumTileRowsMinus1                   : 0                # Number of tile rows in a picture minus 1
+TileRowHeightArray                  : 2                # Array containing tile row height values in units of CTU (from top to bottom in picture)
+
+LFCrossTileBoundaryFlag             : 1                # In-loop filtering is across or not across tile boundary.
+                                                       # 0:not across, 1: across 
+
+#============ WaveFront ================
+WaveFrontSynchro                    : 0                # 0:  No WaveFront synchronisation (WaveFrontSubstreams must be 1 in this case).
+                                                       # >0: WaveFront synchronises with the LCU above and to the right by this many LCUs.
+
+#=========== Quantization Matrix =================
+ScalingList                   : 0                      # ScalingList 0 : off, 1 : default, 2 : file read
+ScalingListFile               : scaling_list.txt       # Scaling List file name. If file is not exist, use Default Matrix.
+
+#============ Lossless ================
+TransquantBypassEnable     : 0                         # Value of PPS flag.
+CUTransquantBypassFlagForce: 0                         # Force transquant bypass mode, when transquant_bypass_enable_flag is enabled
+
+### DO NOT ADD ANYTHING BELOW THIS LINE ###
+### DO NOT DELETE THE EMPTY LINE BELOW ###