def test_inception(inception_model_pytorch):
"""Test the InceptionBlocks model that WebGME folks provided us."""
galaxy_images_output = torch.zeros((1, 5, 64, 64))
ebv_output = torch.zeros((1,))
# Run the model once to get some ground truth outpot (from PyTorch)
output = inception_model_pytorch(galaxy_images_output, ebv_output).detach().numpy()
# Convert to MDF
mdf_model, params_dict = pytorch_to_mdf(
model=inception_model_pytorch,
args=(galaxy_images_output, ebv_output),
example_outputs=output,
trace=True,
)
# Get the graph
mdf_graph = mdf_model.graphs[0]
# Add inputs to the parameters dict so we can feed this to the EvaluableGraph for initialization of all
# graph inputs.
params_dict["input1"] = galaxy_images_output.numpy()
params_dict["input2"] = ebv_output.numpy()
eg = EvaluableGraph(graph=mdf_graph, verbose=False)
eg.evaluate(initializer=params_dict)
assert np.allclose(
output,
eg.enodes["Add_381"].evaluable_outputs["_381"].curr_value,
)
def test_simple_function():
"""Test a simple function"""
def simple(x, y):
return x + y
mdf_model, param_dict = pytorch_to_mdf(
model=simple,
args=(torch.tensor(0.0), torch.tensor(0.0)),
example_outputs=(torch.tensor(0.0)),
use_onnx_ops=True,
)
_check_model(mdf_model)
def test_simple_module():
"""Test a simple torch.nn.Module"""
class Simple(torch.nn.Module):
def forward(self, x, y):
return x + y
mdf_model, param_dict = pytorch_to_mdf(
model=Simple(),
args=(torch.tensor(0.0), torch.tensor(0.0)),
example_outputs=(torch.tensor(0.0)),
use_onnx_ops=True,
)
_check_model(mdf_model)
def main():
ddm_params = dict(
starting_value=0.0,
drift_rate=0.3,
non_decision_time=0.15,
threshold=0.6,
noise=1.0,
time_step_size=0.001,
)
# Move params to device
for key, val in ddm_params.items():
ddm_params[key] = torch.tensor(val).to(dev)
# Run a single ddm
rt, decision = ddm(**ddm_params)
mdf_model, param_dict = pytorch_to_mdf(
model=ddm,
args=tuple(ddm_params.values()),
example_outputs=(rt, decision),
use_onnx_ops=True,
)
# Output the model to JSON
mdf_model.to_json_file("ddm.json")
import sys
if "-graph" in sys.argv:
mdf_model.to_graph_image(
engine="dot",
output_format="png",
view_on_render=False,
level=2,
filename_root="ddm",
only_warn_on_fail=
True, # Makes sure test of this doesn't fail on Windows on GitHub Actions
)
def main():
# changed import call
from modeci_mdf.execution_engine import EvaluableGraph
# Create some test inputs for the model
galaxy_images_output = torch.zeros((1, 5, 64, 64))
ebv_output = torch.zeros((1, ))
# Seed the random number generator to get deterministic behavior for weight initialization
torch.manual_seed(0)
model = InceptionBlocks()
# Turn on eval mode for model to get rid of any randomization due to things like BatchNorm or Dropout
model.eval()
# Run the model once to get some ground truth outpot (from PyTorch)
output = model(galaxy_images_output, ebv_output).detach().numpy()
# Convert to MDF
mdf_model, params_dict = pytorch_to_mdf(
model=model,
args=(galaxy_images_output, ebv_output),
example_outputs=output,
trace=True,
)
# Get the graph
mdf_graph = mdf_model.graphs[0]
# Add inputs to the parameters dict so we can feed this to the EvaluableGraph for initialization of all
# graph inputs.
params_dict["input1"] = galaxy_images_output.numpy()
params_dict["input2"] = ebv_output.numpy()
# Evaluate the model via the MDF scheduler
eg = EvaluableGraph(graph=mdf_graph, verbose=False)
eg.evaluate(initializer=params_dict)
# Make sure the results are the same betweeen PyTorch and MDF
assert np.allclose(
output,
eg.enodes["Add_381"].evaluable_outputs["_381"].curr_value,
)
print("Passed all comparison tests!")
# Output the model to JSON
mdf_model.to_json_file("inception.json")
import sys
if "-graph" in sys.argv:
mdf_model.to_graph_image(
engine="dot",
output_format="png",
view_on_render=False,
level=1,
filename_root="inception",
only_warn_on_fail=
True, # Makes sure test of this doesn't fail on Windows on GitHub Actions
)