def test_run():
g = graph_from_file("files/add.onnx")
example = {"x1": np.array([2]).astype(np.float32), "x2": np.array([5]).astype(np.float32)}
result = run(g,
inputs=example,
outputs=["sum"]
)
assert result[0] == 7, "Add output not correct."
result = run(g, inputs="", outputs="sum")
assert not result, "Model with this input should not run."
def test_graph_from_file():
g = graph_from_file("files/non-existing-file.onnx")
assert not g, "Graph from file failed to check emtpy file."
g = graph_from_file("files/example01.onnx")
assert type(g) is xpb2.GraphProto, "Graph from file failed to open file."
import sclblonnx as so
import numpy as np
from PIL import Image
"""
EXAMPLE 3: Using a previously exported pyTorch model
Here we open an existing and pre-trained Resnet model (trained on the cifar data).
For training details see:
https://github.com/scailable/sclbl-tutorials/tree/master/sclbl-pytorch-onnx
Here we simply evaluate one specific image.
"""
# Retrieve the graph from the stored .onnx model:
g = so.graph_from_file("onnx/cifar10-resnet20.onnx")
# Clean, check, and display (this model passes all the checks).
g = so.clean(g)
so.check(g)
so.display(g)
# To open an image we write a small utility function using Pillow to transform an image to a numpy array.
def process_image(image_path):
# Load Image
img = Image.open(image_path)
# Get the dimensions of the image
width, height = img.size
# Test the resize graph:
large_input = {"large_image": large_img.astype(np.int32)}
result = so.run(sg1, inputs=large_input, outputs=['small_image'])
# Round values in array and cast as 8-bit integer to store back as JPG:
img_arr = np.array(np.round(result[0]), dtype=np.uint8)
out = Image.fromarray(img_arr, mode="RGB")
out.save("images/1-Resized.JPG") # Yes, this works.
# Store the resize onnx:
so.graph_to_file(sg1, "onnx/resize-image-450x600-300x400.onnx")
# So, now we have a working (sub)graph that resizes an image (which obviously we can just load next time)
# Now, we open up the original image processing graph
sg2 = so.graph_from_file("onnx/check-container.onnx")
# The outputs of sg1 and the inputs of sg2 need to match; lets examine them
so.list_outputs(sg1)
so.list_inputs(sg2)
# Merge the two graphs, the outputs will be merged with the inputs in order of appearance:
g = so.merge(sg1, sg2, outputs=["small_image"], inputs=["in"])
so.check(g)
so.display(g)
# And now it works with the large image:
result = so.run(g, inputs=large_input, outputs=['result'])
# Print the result
if result[0]:
print("The container in the large image is empty.")
dnn_model.compile(loss='mean_absolute_error',
optimizer=tf.keras.optimizers.SGD())
# train the model (use .predict for local predictions)
history = dnn_model.fit(X, y, validation_split=0.2, verbose=0, epochs=300)
# Save model (note, the convert_keras() function is undergoing change in different
# versions of tf / onnx).
# You might need: tf.compat.v1.disable_eager_execution()
# or use the tf2onnx tool at https://github.com/onnx/tensorflow-onnx
onnx_model = keras2onnx.convert_keras(dnn_model, dnn_model.name)
keras2onnx.save_model(onnx_model, "onnx/tf-keras-dynamic.onnx")
# load the model using sclblonnx
g = so.graph_from_file("onnx/tf-keras-dynamic.onnx")
# so.display(g)
# check() and clean()
so.check(g)
g = so.clean(g) # Fails due to dynamic size
# Note, while this model passes check(), clean() provides a warning message due to the dynamic input (Nx10).
# This occurs because the training data is N long. However, for inference we would like it to be 1x10
# Let's fix this by changing the input to static.
so.list_inputs(g)
g = so.replace_input(g, "input_1", "FLOAT", [1, 10])
# And do the same for the output
output = so.replace_output(g, "output_1", "FLOAT", [1, 1]) # Check this one...