Implements TensorFlow inference of a saved model, following the same API as the TensorRTInfer class.
| 27 | |
| 28 | |
| 29 | class TensorFlowInfer: |
| 30 | """ |
| 31 | Implements TensorFlow inference of a saved model, following the same API as the TensorRTInfer class. |
| 32 | """ |
| 33 | |
| 34 | def __init__(self, saved_model_path): |
| 35 | gpus = tf.config.experimental.list_physical_devices("GPU") |
| 36 | for gpu in gpus: |
| 37 | tf.config.experimental.set_memory_growth(gpu, True) |
| 38 | |
| 39 | self.model = tf.saved_model.load(saved_model_path) |
| 40 | self.pred_fn = self.model.signatures["serving_default"] |
| 41 | |
| 42 | # Setup I/O bindings |
| 43 | self.inputs = [] |
| 44 | fn_inputs = self.pred_fn.structured_input_signature[1] |
| 45 | for i, input in enumerate(list(fn_inputs.values())): |
| 46 | self.inputs.append( |
| 47 | { |
| 48 | "index": i, |
| 49 | "name": input.name, |
| 50 | "dtype": np.dtype(input.dtype.as_numpy_dtype()), |
| 51 | "shape": input.shape.as_list(), |
| 52 | } |
| 53 | ) |
| 54 | self.outputs = [] |
| 55 | fn_outputs = self.pred_fn.structured_outputs |
| 56 | for i, output in enumerate(list(fn_outputs.values())): |
| 57 | self.outputs.append( |
| 58 | { |
| 59 | "index": i, |
| 60 | "name": output.name, |
| 61 | "dtype": np.dtype(output.dtype.as_numpy_dtype()), |
| 62 | "shape": output.shape.as_list(), |
| 63 | } |
| 64 | ) |
| 65 | |
| 66 | def input_spec(self): |
| 67 | return self.inputs[0]["shape"], self.inputs[0]["dtype"] |
| 68 | |
| 69 | def output_spec(self): |
| 70 | return self.outputs[0]["shape"], self.outputs[0]["dtype"] |
| 71 | |
| 72 | def infer(self, batch, top=1): |
| 73 | # Process I/O and execute the network |
| 74 | input = {self.inputs[0]["name"]: tf.convert_to_tensor(batch)} |
| 75 | output = self.pred_fn(**input) |
| 76 | output = output[self.outputs[0]["name"]].numpy() |
| 77 | |
| 78 | # Read and process the results |
| 79 | classes = np.argmax(output, axis=1) |
| 80 | scores = np.max(output, axis=1) |
| 81 | top = max(top, output.shape[1]) |
| 82 | top_classes = np.flip(np.argsort(output, axis=1), axis=1)[:, 0:top] |
| 83 | top_scores = np.flip(np.sort(output, axis=1), axis=1)[:, 0:top] |
| 84 | |
| 85 | return classes, scores, [top_classes, top_scores] |
| 86 |