Skip to content

Commit

Permalink
docs: document more differences among different backends
Browse files Browse the repository at this point in the history
Signed-off-by: Jinzhe Zeng <jinzhe.zeng@rutgers.edu>
  • Loading branch information
njzjz committed Nov 20, 2024
1 parent ca6a00d commit 1f8d8cd
Show file tree
Hide file tree
Showing 5 changed files with 27 additions and 2 deletions.
8 changes: 7 additions & 1 deletion doc/model/train-fitting-tensor.md
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,8 @@ $deepmd_source_dir/examples/water_tensor/polar/polar_input_torch.json

The training and validation data are also provided our examples. But note that **the data provided along with the examples are of limited amount, and should not be used to train a production model.**

Similar to the `input.json` used in `ener` mode, training JSON is also divided into {ref}`model <model>`, {ref}`learning_rate <learning_rate>`, {ref}`loss <loss>` and {ref}`training <training>`. Most keywords remain the same as `ener` mode, and their meaning can be found [here](train-se-e2-a.md). To fit a tensor, one needs to modify {ref}`model[standard]/fitting_net <model[standard]/fitting_net>` and {ref}`loss <loss>`.
Similar to the `input.json` used in `ener` mode, training JSON is also divided into {ref}`model <model>`, {ref}`learning_rate <learning_rate>`, {ref}`loss <loss>` and {ref}`training <training>`. Most keywords remain the same as `ener` mode, and their meaning can be found [here](train-se-e2-a.md).
To fit a tensor, one needs to modify {ref}`fitting_net <model[standard]/fitting_net>` and {ref}`loss <loss>`.

## Theory

Expand Down Expand Up @@ -241,3 +242,8 @@ One may notice that in each step, some of the local loss and global loss will be
```
During training, at each step when the `lcurve.out` is printed, the system used for evaluating the training (validation) error may be either with only global or only atomic labels, thus the corresponding atomic or global errors are missing and are printed as zeros.
## Difference among different backends
To only fit against a subset of atomic types, in the TensorFlow backend, {ref}`fitting_net/sel_type <model[standard]/fitting_net[dipole]/sel_type>` should be set to selected types;
in other backends, {ref}`atom_exclude_types <model/atom_exclude_types>` should be set to excluded types.
7 changes: 6 additions & 1 deletion doc/model/train-se-atten.md
Original file line number Diff line number Diff line change
Expand Up @@ -152,11 +152,16 @@ In the TensorFlow backend, the {ref}`type_embedding <model/type_embedding>` sect

In other backends, type embedding is within this descriptor with the {ref}`tebd_dim <model[standard]/descriptor[se_atten_v2]/tebd_dim>` argument.

## Difference between TensorFlow and other backends
## Difference among different backends

TensorFlow and other backends have different implementations for {ref}`smooth_type_embedding <model[standard]/descriptor[se_atten_v2]/smooth_type_embedding>`.
The results are inconsistent when `smooth_type_embedding` is `true`.

In the TensorFlow backend, {ref}`scaling_factor <model[standard]/descriptor[se_atten]/scaling_factor>` cannot set to a value other than `1.0`;
{ref}`normalize <model[standard]/descriptor[se_atten]/normalize>` cannot be set to `false`;
{ref}`temperature <model[standard]/descriptor[se_atten]/temperature>` cannot be set;
{ref}`concat_output_tebd <model[standard]/descriptor[se_atten]/concat_output_tebd>` cannot be set to `false`.

## Type map

For training large systems, especially those with dozens of elements, the {ref}`type <model/type_map>` determines the element index of training data:
Expand Down
5 changes: 5 additions & 0 deletions doc/model/train-se-e2-a.md
Original file line number Diff line number Diff line change
Expand Up @@ -100,6 +100,11 @@ The construction of the descriptor is given by section {ref}`descriptor <model[s
Type embdding is only supported in the TensorFlow backends.
`se_e2_a` with type embedding and [`se_atten`](./train-se-atten.md) (or its updated version) without any attention layer are mathematically equivalent, so `se_atten` can be a substitute in other backends.

## Difference among different backends

In the TensorFlow backend, {ref}`env_protection <model[standard]/descriptor[se_e2_a]/env_protection>` cannot be set to a non-zero value.
In the JAX backend, {ref}`type_one_side <model[standard]/descriptor[se_e2_a]/type_one_side>` cannot be set to `false`.

## Model compression

Model compression is supported when type embedding is not used.
Expand Down
5 changes: 5 additions & 0 deletions doc/model/train-se-e2-r.md
Original file line number Diff line number Diff line change
Expand Up @@ -74,6 +74,11 @@ The type of the descriptor is set by the key {ref}`type <model[standard]/descrip

Type embdding is only supported in the TensorFlow backends.

## Difference among different backends

In the TensorFlow backend, {ref}`env_protection <model[standard]/descriptor[se_e2_r]/env_protection>` cannot be set to a non-zero value.
In the PyTorch, JAX, and DP backend, {ref}`type_one_side <model[standard]/descriptor[se_e2_r]/type_one_side>` cannot be set to `false`.

## Model compression

Model compression is supported when type embedding is not used.
4 changes: 4 additions & 0 deletions doc/model/train-se-e3.md
Original file line number Diff line number Diff line change
Expand Up @@ -69,6 +69,10 @@ The type of the descriptor is set by the key {ref}`type <model[standard]/descrip

Use [`se_e3_tebd`](./train-se-e3-tebd.md) for type embedding support.

## Difference among different backends

In the TensorFlow backend, {ref}`env_protection <model[standard]/descriptor[se_e3]/env_protection>` cannot be set to a non-zero value.

## Model compression

Model compression is supported.

0 comments on commit 1f8d8cd

Please sign in to comment.