| import sys, numpy as np, torch, onnxruntime as ort | |
| p = torch.load("probe.pt", weights_only=True) | |
| a, b, ref = p["a"].numpy(), p["b"].numpy(), p["ref"].numpy() | |
| if len(sys.argv) > 1: | |
| import tensorflow as tf | |
| it = tf.lite.Interpreter(model_path=sys.argv[1]); it.allocate_tensors() | |
| ins = it.get_input_details(); outs = it.get_output_details() | |
| print("tflite inputs", [(d["name"], tuple(d["shape"]), d["dtype"].__name__) for d in ins], "out", tuple(outs[0]["shape"])) | |
| for d in ins: | |
| x = a if "0" in d["name"] else b | |
| it.set_tensor(d["index"], np.transpose(x, (0, 2, 3, 1)).astype(d["dtype"])) | |
| it.invoke() | |
| y = np.transpose(it.get_tensor(outs[0]["index"]), (0, 3, 1, 2)) | |
| d = np.abs(y - ref) | |
| print("tflite vs torch: max", float(d.max()), "mean", float(d.mean()), "share above 1/255:", float((d > 1 / 255).mean())) | |