Improve Docs (#147)

docs/2-getting-started.md

@@ -47,6 +47,7 @@ For users with AMD GPUs.
 - Install the [HIP SDK](https://rocm.docs.amd.com/projects/install-on-windows/en/latest/how-to/install.html)
 - The `HIP_PATH` environment variable must be set to the HIP install location (should contain the `bin`, `lib` and `include` directories)
 - The system `PATH` should contain `%HIP_PATH%\bin` (or equivalent for Linux)
+- On Linux, you will need to specify the `GCN_ARCH_NAME` environment variable, which you should be able to find using `rocminfo`.
 
 #### CPU
 If you need to train on CPU you can use the [legacy branch](https://github.com/jw1912/bullet/tree/legacy).

docs/4-saved-networks.md

@@ -1,11 +1,11 @@
 ## Saved Networks
 
 When a checkpoint is saved to a directory `<out_dir>/<checkpoint_name>`, it will contain
-- `unquantised.bin`, the raw floating point (`f32`) parameters of the network
+- `raw.bin`, the raw floating point (`f32`) parameters of the network
 - `quantised.bin`, the quantised network, padded to be a multiple of 64 bytes
-- `optimiser_state`, the internal state of the optimiser
+- `optimiser_state/`, the internal state of the optimiser
 
 If quantisation fails (due to integer overflow), then it will not save the quantised network, but training will be otherwise unaffected.
 
-For the two network files, the format of these is `(layer 1 weights)(layer 1 biases)(layer 2 weights)...` stored
+If you are using the `TrainerBuilder`, the format of the two network files is `(layer 1 weights)(layer 1 biases)(layer 2 weights)...` stored
 contiguously, with the weights matrices being stored column-major.