xref: /aosp_15_r20/external/pytorch/test/test_bundled_images.py (revision da0073e96a02ea20f0ac840b70461e3646d07c45) 
1#!/usr/bin/env python3
2# Owner(s): ["oncall: mobile"]
3# mypy: allow-untyped-defs
4
5import io
6
7import cv2
8
9import torch
10import torch.utils.bundled_inputs
11from torch.testing._internal.common_utils import TestCase
12
13
14torch.ops.load_library("//caffe2/torch/fb/operators:decode_bundled_image")
15
16
17def model_size(sm):
18    buffer = io.BytesIO()
19    torch.jit.save(sm, buffer)
20    return len(buffer.getvalue())
21
22
23def save_and_load(sm):
24    buffer = io.BytesIO()
25    torch.jit.save(sm, buffer)
26    buffer.seek(0)
27    return torch.jit.load(buffer)
28
29
30"""Return an InflatableArg that contains a tensor of the compressed image and the way to decode it
31
32    keyword arguments:
33    img_tensor -- the raw image tensor in HWC or NCHW with pixel value of type unsigned int
34                  if in NCHW format, N should be 1
35    quality -- the quality needed to compress the image
36"""
37
38
39def bundle_jpeg_image(img_tensor, quality):
40    # turn NCHW to HWC
41    if img_tensor.dim() == 4:
42        assert img_tensor.size(0) == 1
43        img_tensor = img_tensor[0].permute(1, 2, 0)
44    pixels = img_tensor.numpy()
45    encode_param = [int(cv2.IMWRITE_JPEG_QUALITY), quality]
46    _, enc_img = cv2.imencode(".JPEG", pixels, encode_param)
47    enc_img_tensor = torch.from_numpy(enc_img)
48    enc_img_tensor = torch.flatten(enc_img_tensor).byte()
49    obj = torch.utils.bundled_inputs.InflatableArg(
50        enc_img_tensor, "torch.ops.fb.decode_bundled_image({})"
51    )
52    return obj
53
54
55def get_tensor_from_raw_BGR(im) -> torch.Tensor:
56    raw_data = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
57    raw_data = torch.from_numpy(raw_data).float()
58    raw_data = raw_data.permute(2, 0, 1)
59    raw_data = torch.div(raw_data, 255).unsqueeze(0)
60    return raw_data
61
62
63class TestBundledImages(TestCase):
64    def test_single_tensors(self):
65        class SingleTensorModel(torch.nn.Module):
66            def forward(self, arg):
67                return arg
68
69        im = cv2.imread("caffe2/test/test_img/p1.jpg")
70        tensor = torch.from_numpy(im)
71        inflatable_arg = bundle_jpeg_image(tensor, 90)
72        input = [(inflatable_arg,)]
73        sm = torch.jit.script(SingleTensorModel())
74        torch.utils.bundled_inputs.augment_model_with_bundled_inputs(sm, input)
75        loaded = save_and_load(sm)
76        inflated = loaded.get_all_bundled_inputs()
77        decoded_data = inflated[0][0]
78
79        # raw image
80        raw_data = get_tensor_from_raw_BGR(im)
81
82        self.assertEqual(len(inflated), 1)
83        self.assertEqual(len(inflated[0]), 1)
84        self.assertEqual(raw_data.shape, decoded_data.shape)
85        self.assertEqual(raw_data, decoded_data, atol=0.1, rtol=1e-01)
86
87        # Check if fb::image_decode_to_NCHW works as expected
88        with open("caffe2/test/test_img/p1.jpg", "rb") as fp:
89            weight = torch.full((3,), 1.0 / 255.0).diag()
90            bias = torch.zeros(3)
91            byte_tensor = torch.tensor(list(fp.read())).byte()
92            im2_tensor = torch.ops.fb.image_decode_to_NCHW(byte_tensor, weight, bias)
93            self.assertEqual(raw_data.shape, im2_tensor.shape)
94            self.assertEqual(raw_data, im2_tensor, atol=0.1, rtol=1e-01)
95