| resizer_node = onnx.helper.make_node("Resize", inputs=[model_input_name, "roi", "scales", "output_size"], outputs=[resizer_node_output_name]) | |
| transpose_node = onnx.helper.make_node("Transpose", inputs=[resizer_node_output_name], outputs=[model_output_name], perm=[0,3,1,2]) |
| # Permute dimension, after resize they are in this order | |
| # N W H C | |
| # 0 1 2 3 | |
| # After transpose we want them in shape | |
| # N C W H | |
| # 0 3 1 2 | |
| perm = onnx.helper.make_node("Constant", inputs=[], outputs=["perm"], name="perm-constant", | |
| value=onnx.helper.make_tensor(name="perm-values", | |
| data_type=onnx.TensorProto.INT64, | |
| dims=np.array([0,3,1,2]).shape, |
| # pip install wget | |
| # pip install -i https://test.pypi.org/simple/ onnx-weekly | |
| # pip install onnxruntime | |
| import onnx | |
| import numpy as np | |
| import onnxruntime as rt | |
| import matplotlib.pyplot as plt | |
| from PIL import Image |
| # !wget https://github.com/EliSchwartz/imagenet-sample-images/raw/master/n02085936_Maltese_dog.JPEG | |
| large_img = np.array(Image.open("n02085936_Maltese_dog.JPEG")) | |
| print(f"Image shape: {large_img.shape}") | |
| plt.imshow(large_img) | |
| # Output: | |
| # Image shape: (500, 375, 3) | |
| sess = rt.InferenceSession("combined_dynamic_input.onnx") # Start the inference session and open the model |
| import onnx | |
| # !wget https://github.com/onnx/models/raw/main/vision/classification/squeezenet/model/squeezenet1.0-12.onnx | |
| model1 = onnx.load('resizer.onnx') | |
| model2 = onnx.load('squeezenet1.0-12.onnx') | |
| combined_model = onnx.compose.merge_models( | |
| model1, model2, |
| graph = onnx.helper.make_graph([roi, scales, output_size, resizer_node, transpose_node], "resizer", [X], [Y]) | |
| print(graph) | |
| # Create the model (ModelProto instance) | |
| model_def = onnx.helper.make_model(g, producer_name="onnx-zaitra") | |
| # these are optional, just to show what you can set here | |
| model_def.opset_import[0].version = 12 | |
| model_def.ir_version = 7 |
| resizer_node_output_name = "Resize_Y" | |
| Y_resize = onnx.helper.make_tensor_value_info(resizer_node_output_name, | |
| onnx.TensorProto.FLOAT, | |
| [None, 224, 224, 3]) | |
| roi = onnx.helper.make_node("Constant", inputs=[], outputs=["roi"], name="roi-constant", | |
| value=onnx.helper.make_tensor(name="roi-values", | |
| data_type=onnx.TensorProto.FLOAT, |
| model_input_name = "X" | |
| X = onnx.helper.make_tensor_value_info( | |
| model_input_name, | |
| onnx.TensorProto.FLOAT, | |
| [None, None, None, 3]) # notice that we are leaving just channels dimension fixed, Batch size, width and heigh is not known in advance | |
| model_output_name = "Y" | |
| Y = onnx.helper.make_tensor_value_info( | |
| model_output_name, |
| # Adapted from this StackOverflow answer: https://gis.stackexchange.com/a/14717 | |
| """ | |
| Usage: | |
| file_path = "/tmp/file.tiff" | |
| split_tiff("/tmp/file.tiff", 256, 256) | |
| The resulting output will be chunks /tmp/file_0_0.tiff ... |
Summer is not over yet, but coding part of Google summer of code yes. This is my final report.
I was working with fedora organization on conu - python API for your containers. Before GSOC, you could use library just to work with standalone containers. Goal for GSOC was to implement support also for orchestration tools like Kubernetes and OpenShift. For more information see my proposal.