def create_assets(basename,
sig_input,
sig_output,
sig_name,
lengths_key=None,
beam=None,
return_labels=False,
preproc='client'):
"""Save required variables for running an exported model from baseline's services.
:basename the base model name. e.g. /path/to/tagger-26075
:sig_input the input dictionary
:sig_output the output namedTuple
:lengths_key the lengths_key from the model.
used to translate batch output. Exported models will return a flat list,
and it needs to be reshaped into per-example lists. We use this key to tell
us which feature holds the sequence lengths.
"""
inputs = [k for k in sig_input]
outputs = sig_output._fields
model_name = basename.split("/")[-1]
directory = basename.split("/")[:-1]
metadata = create_metadata(inputs,
outputs,
sig_name,
model_name,
lengths_key,
beam=beam,
return_labels=return_labels,
preproc=preproc)
return metadata
def run(self, basename, output_dir, project=None, name=None, model_version=None, **kwargs):
logger.warning("Pytorch exporting is experimental and is not guaranteed to work for plugin models.")
client_output, server_output = get_output_paths(
output_dir,
project, name,
model_version,
kwargs.get('remote', True),
)
logger.info("Saving vectorizers and vocabs to %s", client_output)
logger.info("Saving serialized model to %s", server_output)
model, vectorizers, model_name = self.load_model(basename)
order = monkey_patch_embeddings(model)
data, lengths = create_fake_data(VECTORIZER_SHAPE_MAP, vectorizers, order)
meta = create_metadata(
order, ['output'],
self.sig_name,
model_name, model.lengths_key,
exporter_type=self.preproc_type()
)
exportable = self.wrapper(model)
logger.info("Tracing Model.")
traced = torch.jit.trace(exportable, (data, lengths))
traced.save(os.path.join(server_output, 'model.pt'))
logger.info("Saving metadata.")
save_to_bundle(client_output, basename, assets=meta)
logger.info('Successfully exported model to %s', output_dir)
def create_assets(basename, sig_input, sig_output, sig_name, lengths_key=None, beam=None, return_labels=False, preproc='client'):
"""Save required variables for running an exported model from baseline's services.
:basename the base model name. e.g. /path/to/tagger-26075
:sig_input the input dictionary
:sig_output the output namedTuple
:lengths_key the lengths_key from the model.
used to translate batch output. Exported models will return a flat list,
and it needs to be reshaped into per-example lists. We use this key to tell
us which feature holds the sequence lengths.
"""
inputs = [k for k in sig_input]
outputs = sig_output._fields
model_name = basename.split("/")[-1]
directory = basename.split("/")[:-1]
metadata = create_metadata(inputs, outputs, sig_name, model_name, lengths_key, beam=beam,
return_labels=return_labels, preproc=preproc)
return metadata
def run(self,
basename,
output_dir,
project=None,
name=None,
model_version=None,
**kwargs):
logger.warning(
"Pytorch exporting is experimental and is not guaranteed to work for plugin models."
)
client_output, server_output = get_output_paths(
output_dir,
project,
name,
model_version,
kwargs.get('remote', True),
)
logger.info("Saving vectorizers and vocabs to %s", client_output)
logger.info("Saving serialized model to %s", server_output)
model, vectorizers, model_name = self.load_model(basename)
order = monkey_patch_embeddings(model)
data, lengths = create_fake_data(VECTORIZER_SHAPE_MAP, vectorizers,
order)
meta = create_metadata(order, ['output'],
self.sig_name,
model_name,
model.lengths_key,
exporter_type=self.preproc_type())
exportable = self.wrapper(model)
logger.info("Tracing Model.")
traced = torch.jit.trace(exportable, (data, lengths))
traced.save(os.path.join(server_output, 'model.pt'))
logger.info("Saving metadata.")
save_to_bundle(client_output, basename, assets=meta)
logger.info('Successfully exported model to %s', output_dir)
def _run(self,
basename,
output_dir,
project=None,
name=None,
model_version=None,
use_version=False,
zip_results=True,
**kwargs):
logger.warning(
"Pytorch exporting is experimental and is not guaranteed to work for plugin models."
)
client_output, server_output = get_output_paths(
output_dir,
project,
name,
model_version,
kwargs.get('remote', False),
use_version=use_version)
logger.info("Saving vectorizers and vocabs to %s", client_output)
logger.info("Saving serialized model to %s", server_output)
model, vectorizers, vocabs, model_name = self.load_model(basename)
model = self.apply_model_patches(model)
data = self.create_example_input(vocabs, vectorizers)
example_output = self.create_example_output(model)
inputs = self.create_model_inputs(model)
outputs = self.create_model_outputs(model)
dynamics = self.create_dynamic_axes(model, vectorizers, inputs,
outputs)
meta = create_metadata(inputs, outputs, self.sig_name, model_name,
model.lengths_key)
if not self.tracing:
model = torch.jit.script(model)
logger.info("Exporting Model.")
logger.info("Model inputs: %s", inputs)
logger.info("Model outputs: %s", outputs)
torch.onnx.export(model,
data,
verbose=True,
dynamic_axes=dynamics,
f=f'{server_output}/{model_name}.onnx',
input_names=inputs,
output_names=outputs,
opset_version=self.onnx_opset,
example_outputs=example_output)
logger.info("Saving metadata.")
save_to_bundle(client_output,
basename,
assets=meta,
zip_results=zip_results)
logger.info('Successfully exported model to %s', output_dir)
return client_output, server_output
def _run(self,
basename,
output_dir,
project=None,
name=None,
model_version=None,
use_version=False,
zip_results=True,
remote=False,
**kwargs):
client_output, server_output = get_output_paths(
output_dir,
project,
name,
model_version,
remote,
use_version=use_version)
logger.info("Saving vectorizers and vocabs to %s", client_output)
logger.info("Saving serialized model to %s", server_output)
model, vectorizers, vocabs, model_name = self.load_model(basename)
# Triton server wants to see a specific name
model = self.apply_model_patches(model)
data = self.create_example_input(vocabs, vectorizers)
example_output = self.create_example_output(model)
inputs = self.create_model_inputs(model)
outputs = self.create_model_outputs(model)
dynamics = self.create_dynamic_axes(model, vectorizers, inputs,
outputs)
meta = create_metadata(inputs, outputs, self.sig_name, model_name,
model.lengths_key)
if not self.tracing:
model = torch.jit.script(model)
logger.info("Exporting Model.")
logger.info("Model inputs: %s", inputs)
logger.info("Model outputs: %s", outputs)
onnx_model_name = REMOTE_MODEL_NAME if remote else model_name
torch.onnx.export(
model,
data,
verbose=True,
dynamic_axes=dynamics,
f=f'{server_output}/{onnx_model_name}.onnx',
input_names=inputs,
output_names=outputs,
opset_version=self.onnx_opset,
#propagate=True,
example_outputs=example_output)
logger.info("Saving metadata.")
save_to_bundle(client_output,
basename,
assets=meta,
zip_results=zip_results)
logger.info('Successfully exported model to %s', output_dir)
return client_output, server_output