update readme

Author: fracape

compressai/sadl_codec/readme.md

@@ -3,16 +3,16 @@
 ## Workflow
 
 1- compressAI:
- - train a model convertible to SADL (for example FactorizedPriorBasic), see below for details
+ - train a model convertible to SADL (for example FactorizedPriorReLU), see below for details
  - dump a pth containing the model (g\_a, g\_s, possibly the cdfs and quantizers if needed)
  - prepare a training set of raw patch 256x256x3 in npy format (uint8) to be used to recompute statistics of the latents
- 
+
 
 2- Conversion
  - initialize the SADL submodule
  - run the script build\_codec.sh in a working directory:
  ```shell
- compressai/sadl_codec/build_codec.sh --model model.pth --training_dataset trainingdataset.npy 
+ compressai/sadl_codec/build_codec.sh --model model.pth --training_dataset trainingdataset.npy
  ```
 Note: it can take a while to generate the first time as it has to perform an inference on the whole training set.
 
@@ -23,7 +23,7 @@ Note: it can take a while to generate the first time as it has to perform an inf
  - OPTIONAL STEP 3/4: extract quantizers information and create a sadl int16 decoder.
  - STEP 5: build the C++ decoder
  - STEP 6: extract the encoder, build the C++ encoder
- 
+
 
 3- Run on kodak dataset
 ```shell
@@ -45,10 +45,10 @@ Several version of the encoder/decoder are available (float version, non SIMD et
 4- Details
 
 File details:
-- model\_dec.onnx: contains just the network part with the deconvolution and activation compatible with SADL (e.g. ReLU). 
-- model\_info.pkl: contains information beside the network itself: 
+- model\_dec.onnx: contains just the network part with the deconvolution and activation compatible with SADL (e.g. ReLU).
+- model\_info.pkl: contains information beside the network itself:
 the cdfs, cdfs length and cdfs offset and the quantizers for each deconv layers parameters inside a dict { 'cdfs': nparray, 'cdflen': nparray, 'cdfoff': nparray, 'quantizers': '0.weight': 8,  '0.bias': 10, ...} }
-- model_enc.onnx: contains just the network part with the convolution and activation compatible with SADL (e.g. ReLU). 
+- model_enc.onnx: contains just the network part with the convolution and activation compatible with SADL (e.g. ReLU).
 
 
 5- Train a model