Updating instructions for XLA export

README.md

@@ -54,3 +54,22 @@ To export a trained model in ONNX format, run:
 python3 export_model.py --checkpoint_name test_run --export_format onnx
 ```
 Here `test_run` is the `CHECKPOINT_NAME` used in examples above. This creates a file `model_export/onnx/CHECKPOINT_NAME.onnx` containing the model graph and weights.
+
+## Model export (XLA format)
+
+To export a trained model in XLA format (compiled `.so`), run:
+```bash
+python3 export_model.py --checkpoint_name test_run --export_format pbtxt
+```
+Here `test_run` is the `CHECKPOINT_NAME` used in examples above. This creates protobuf files describing the model:
+```
+graph.pbtxt
+graph-1.config.pbtxt
+graph-10.config.pbtxt
+graph-100.config.pbtxt
+graph-1000.config.pbtxt
+graph-10000.config.pbtxt
+```
+and puts them into `model_export/model_v4/` directory (previously created such files would get overwritten). The `graph.pbtxt` contains the frozen computation graph of the model with all the weight values. Files `graph-X.pbtxt` specify additional configuration for model build target - input and output tensor names and shapes (the number `X` standing for the batch size).
+
+After exporting the model graph in protobuf format, please follow [these instructions](model_export/README.md) to build the model `.so` files.

model_export/README.md

@@ -10,7 +10,7 @@ The `setup_centos.sh` script (`model_export/setup_centos.sh`) contains all the e
 
 *Note: this was only tested in a clean CentOS environment (in the docker container above), there might be conflicts with the main environment used for model training and evaluation.*
 
-The `setup_centos.sh` script also creates a symlink to the `model_export/model_v4` folder and puts it into `tensorflow`. This folder contains bazel build instructions and needs to be inside the `tensorflow` directory to be built.
+The `setup_centos.sh` script also creates a symlink to the `model_export/model_v4` folder (which should contain the `*.pbtxt` files describing the exported model) and puts it into `tensorflow`. This folder contains bazel build instructions and needs to be inside the `tensorflow` directory to be built.
 ```bash
 ./setup_centos.sh
 ```
@@ -23,7 +23,7 @@ cd tensorflow/
 # rest options to default
 BAZEL_LINKLIBS=-l%:libstdc++.a bazel build --local_resources 30000,15,1.0 -c opt model_v4:all_models
 ```
-The `--local_resources` parameter from the last line allows to limit the resources to be used by bazel. 30000 is memory in megabytes, 15 is the number of CPU cores and 1.0 is the I/O capability. Not limiting the resources may result in a compiler crash.
+The `--local_resources` parameter from the last line allows to limit the resources to be used by bazel. 30000 is memory in megabytes, 15 is the number of CPU cores and 1.0 is the I/O capability. Not limiting the resources may result in a compiler crash. To further rebuild the model (e.g., after updating the `*.pbtxt` files) only the last command is needed.
 
 # Test:
 ```bash