Skip to content

fix bugs of single precision training and transfer #423

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
amcadmus merged 4 commits into from
Mar 19, 2021
Merged

fix bugs of single precision training and transfer #423

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
147732e
Merge pull request #17 from deepmodeling/api
denghuilu Mar 15, 2021
f5af444
Merge pull request #18 from deepmodeling/api
denghuilu Mar 18, 2021
ebdb404
fix bugs of single precision training and transfer
denghuilu Mar 18, 2021
3dc9768
fix bug of nbor sorting
denghuilu Mar 18, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 12 additions & 18 deletions deepmd/entrypoints/transfer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,7 +72,7 @@ def transfer(*, old_model: str, raw_model: str, output: str, **kwargs):
new_graph_def = transform_graph(raw_graph, old_graph)
with tf.gfile.GFile(output, mode="wb") as f:
f.write(new_graph_def.SerializeToString())
log.info("the output model is saved in {output:s}")
log.info("the output model is saved in " + output)


def load_graph(graph_name: str) -> tf.Graph:
Expand Down Expand Up @@ -136,37 +136,31 @@ def transform_graph(raw_graph: tf.Graph, old_graph: tf.Graph) -> tf.Graph:
if raw_graph_dtype == np.float16:
if old_graph_dtype == np.float64 or old_graph_dtype == np.float32:
if (len(tensor_shape) != 1) or (tensor_shape[0] != 1):
tensor = np.frombuffer(old_node.tensor_content, dtype=np.float16)
cp_attr.from_array(tensor, tf.float16, shape=tensor_shape)
tensor = np.frombuffer(old_node.tensor_content).astype(raw_graph_dtype)
cp_attr.from_array(tensor, tf.float16, shape = tensor_shape)
else:
tensor = load_tensor(old_node, old_graph_dtype, np.float16)
tensor = load_tensor(old_node, old_graph_dtype, raw_graph_dtype)
cp_attr.from_array(tensor, tf.float16, [1])

elif old_graph_dtype == np.float16:
tensor = convert_matrix(np.array(old_node.half_val), tensor_shape)
cp_attr.from_array(tensor, tf.float16)
elif old_graph_dtype[1] == "float16":
tensor = convertMatrix(np.array(old_node.half_val), tensor_shape)
cp_attr.from_array(tensor, raw_graph_dtype)

elif raw_graph_dtype == np.float64 or raw_graph_dtype == np.float32:
if old_graph_dtype == np.float64 or old_graph_dtype == np.float32:
if (len(tensor_shape) != 1) or (tensor_shape[0] != 1):
tensor = np.frombuffer(
old_node.tensor_content, dtype=raw_graph_dtype
)
tensor = np.frombuffer(old_node.tensor_content).astype(raw_graph_dtype)
cp_attr.from_str(tensor)
else:
tensor = load_tensor(old_node, old_graph_dtype, raw_graph_dtype)
cp_attr.from_array(tensor, raw_graph_dtype, shape=[1])

elif old_graph_dtype == np.float16:
if (len(tensor_shape) != 1) or (tensor_shape[0] != 1):
tensor = convert_matrix(
np.array(old_node.half_val), tensor_shape, dtype=raw_graph_dtype
)
tensor = convertMatrix(np.array(old_node.half_val), tensor_shape).astype(raw_graph_dtype)
cp_attr.from_str(tensor)
else:
tensor = convert_matrix(
np.array(old_node.half_val), tensor_shape, dtype=raw_graph_dtype
)
tensor = convertMatrix(np.array(old_node.half_val), tensor_shape).astype(raw_graph_dtype)
cp_attr.from_array(tensor, raw_graph_dtype)

return raw_graph_def
Expand All @@ -191,9 +185,9 @@ def from_str(self, tensor: np.ndarray):

def load_tensor(node: tf.Tensor, dtype_old: type, dtype_new: type) -> np.ndarray:
if dtype_old == np.float64:
tensor = np.array(node.double_val, dtype=dtype_new)
tensor = np.array(node.double_val).astype(dtype_new)
elif dtype_old == np.float32:
tensor = np.array(node.float_val, dtype=dtype_new)
tensor = np.array(node.float_val).astype(dtype_new)

return tensor

Expand Down
2 changes: 1 addition & 1 deletion deepmd/loss/ener.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -119,7 +119,7 @@ def build (self,

# only used when tensorboard was set as true
self.l2_loss_summary = tf.summary.scalar('l2_loss', tf.sqrt(l2_loss))
self.l2_loss_ener_summary = tf.summary.scalar('l2_ener_loss', tf.sqrt(l2_ener_loss) / global_cvt_2_tf_float(natoms[0]))
self.l2_loss_ener_summary = tf.summary.scalar('l2_ener_loss', global_cvt_2_tf_float(tf.sqrt(l2_ener_loss)) / global_cvt_2_tf_float(natoms[0]))
self.l2_loss_force_summary = tf.summary.scalar('l2_force_loss', tf.sqrt(l2_force_loss))
self.l2_loss_virial_summary = tf.summary.scalar('l2_virial_loss', tf.sqrt(l2_virial_loss) / global_cvt_2_tf_float(natoms[0]))

Expand Down
4 changes: 2 additions & 2 deletions source/lib/src/cuda/prod_env_mat.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -114,7 +114,7 @@ __global__ void format_nlist_fill_a(
}
FPTYPE rr = sqrt(dev_dot(diff, diff));
if (rr <= rcut) {
key_in[idy] = type[j_idx] * 1E15+ (int_64)(rr * 1.0E13) / 100000 * 100000 + j_idx;
key_in[idy] = type[j_idx] * 1E15+ (int_64)(rr * 1.0E13) / 10000000 * 10000000 + j_idx;
}
}

Expand Down Expand Up @@ -144,7 +144,7 @@ __global__ void format_nlist_fill_b(
for (unsigned int kk = 0; key_out[kk] != key_out[max_nbor_size - 1]; kk++) {
const int & nei_type = key_out[kk] / 1E15;
if (nei_iter[nei_type] < sec[nei_type + 1]) {
row_nlist[nei_iter[nei_type]++] = key_out[kk] % 100000;
row_nlist[nei_iter[nei_type]++] = key_out[kk] % 10000000;
}
}
}
Expand Down