Pavel Macenauer | 59e057f | 2020-04-15 14:17:26 +0000 | [diff] [blame^] | 1 | #!/usr/bin/env python3 |
Pavel Macenauer | d0fedae | 2020-04-15 14:52:57 +0000 | [diff] [blame] | 2 | # Copyright 2020 NXP |
| 3 | # SPDX-License-Identifier: MIT |
| 4 | |
| 5 | from zipfile import ZipFile |
| 6 | import numpy as np |
| 7 | import pyarmnn as ann |
| 8 | import example_utils as eu |
| 9 | import os |
| 10 | |
| 11 | |
| 12 | def unzip_file(filename): |
| 13 | """Unzips a file to its current location. |
| 14 | |
| 15 | Args: |
| 16 | filename (str): Name of the archive. |
| 17 | |
| 18 | Returns: |
| 19 | str: Directory path of the extracted files. |
| 20 | """ |
| 21 | with ZipFile(filename, 'r') as zip_obj: |
| 22 | zip_obj.extractall(os.path.dirname(filename)) |
| 23 | return os.path.dirname(filename) |
| 24 | |
| 25 | |
| 26 | if __name__ == "__main__": |
| 27 | # Download resources |
| 28 | archive_filename = eu.download_file( |
| 29 | 'https://storage.googleapis.com/download.tensorflow.org/models/tflite/mobilenet_v1_1.0_224_quant_and_labels.zip') |
| 30 | dir_path = unzip_file(archive_filename) |
| 31 | # names of the files in the archive |
| 32 | labels_filename = os.path.join(dir_path, 'labels_mobilenet_quant_v1_224.txt') |
| 33 | model_filename = os.path.join(dir_path, 'mobilenet_v1_1.0_224_quant.tflite') |
| 34 | kitten_filename = eu.download_file('https://s3.amazonaws.com/model-server/inputs/kitten.jpg') |
| 35 | |
| 36 | # Create a network from the model file |
| 37 | net_id, graph_id, parser, runtime = eu.create_tflite_network(model_filename) |
| 38 | |
| 39 | # Load input information from the model |
| 40 | # tflite has all the need information in the model unlike other formats |
| 41 | input_names = parser.GetSubgraphInputTensorNames(graph_id) |
| 42 | assert len(input_names) == 1 # there should be 1 input tensor in mobilenet |
| 43 | |
| 44 | input_binding_info = parser.GetNetworkInputBindingInfo(graph_id, input_names[0]) |
| 45 | input_width = input_binding_info[1].GetShape()[1] |
| 46 | input_height = input_binding_info[1].GetShape()[2] |
| 47 | |
| 48 | # Load output information from the model and create output tensors |
| 49 | output_names = parser.GetSubgraphOutputTensorNames(graph_id) |
| 50 | assert len(output_names) == 1 # and only one output tensor |
| 51 | output_binding_info = parser.GetNetworkOutputBindingInfo(graph_id, output_names[0]) |
| 52 | output_tensors = ann.make_output_tensors([output_binding_info]) |
| 53 | |
| 54 | # Load labels file |
| 55 | labels = eu.load_labels(labels_filename) |
| 56 | |
| 57 | # Load images and resize to expected size |
| 58 | image_names = [kitten_filename] |
| 59 | images = eu.load_images(image_names, input_width, input_height) |
| 60 | |
| 61 | for idx, im in enumerate(images): |
| 62 | # Create input tensors |
| 63 | input_tensors = ann.make_input_tensors([input_binding_info], [im]) |
| 64 | |
| 65 | # Run inference |
| 66 | print("Running inference on '{0}' ...".format(image_names[idx])) |
| 67 | runtime.EnqueueWorkload(net_id, input_tensors, output_tensors) |
| 68 | |
| 69 | # Process output |
| 70 | out_tensor = ann.workload_tensors_to_ndarray(output_tensors)[0][0] |
| 71 | results = np.argsort(out_tensor)[::-1] |
| 72 | eu.print_top_n(5, results, labels, out_tensor) |