| 50 | } |
| 51 | |
| 52 | void Compute(OpKernelContext* c) override { |
| 53 | const Tensor& tags = c->input(0); |
| 54 | const Tensor& tensor = c->input(1); |
| 55 | OP_REQUIRES(c, IsLegacyScalar(tags.shape()), |
| 56 | errors::InvalidArgument("Tags must be a scalar")); |
| 57 | OP_REQUIRES(c, |
| 58 | tensor.dims() == 4 && |
| 59 | (tensor.dim_size(3) == 1 || tensor.dim_size(3) == 3 || |
| 60 | tensor.dim_size(3) == 4), |
| 61 | errors::InvalidArgument( |
| 62 | "Tensor must be 4-D with last dim 1, 3, or 4, not ", |
| 63 | tensor.shape().DebugString())); |
| 64 | const string& base_tag = tags.scalar<tstring>()(); |
| 65 | |
| 66 | OP_REQUIRES(c, |
| 67 | tensor.dim_size(0) < (1LL << 31) && |
| 68 | tensor.dim_size(1) < (1LL << 31) && |
| 69 | tensor.dim_size(2) < (1LL << 31) && |
| 70 | (tensor.dim_size(1) * tensor.dim_size(2)) < (1LL << 29), |
| 71 | errors::InvalidArgument("Tensor too large for summary ", |
| 72 | tensor.shape().DebugString())); |
| 73 | |
| 74 | // The casts and h * w cannot overflow because of the limits above. |
| 75 | const int batch_size = static_cast<int>(tensor.dim_size(0)); |
| 76 | const int h = static_cast<int>(tensor.dim_size(1)); |
| 77 | const int w = static_cast<int>(tensor.dim_size(2)); |
| 78 | const int hw = h * w; // Compact these two dims for simplicity |
| 79 | const int depth = static_cast<int>(tensor.dim_size(3)); |
| 80 | |
| 81 | OP_REQUIRES(c, hw > 0 && depth > 0, |
| 82 | errors::InvalidArgument( |
| 83 | "input tensor must have non-zero dims. Found: [", |
| 84 | batch_size, ", ", h, ", ", w, ", ", depth, "].")); |
| 85 | |
| 86 | Summary s; |
| 87 | if (tensor.dtype() == DT_UINT8) { |
| 88 | // For uint8 input, no normalization is necessary |
| 89 | auto ith_image = [&tensor, batch_size, hw, depth](int i) { |
| 90 | auto values = tensor.shaped<uint8, 3>({batch_size, hw, depth}); |
| 91 | return typename TTypes<uint8>::ConstMatrix( |
| 92 | &values(i, 0, 0), Eigen::DSizes<Eigen::DenseIndex, 2>(hw, depth)); |
| 93 | }; |
| 94 | OP_REQUIRES_OK( |
| 95 | c, AddImages(base_tag, batch_size, w, h, depth, ith_image, &s)); |
| 96 | } else if (tensor.dtype() == DT_HALF) { |
| 97 | NormalizeAndAddImages<Eigen::half>(c, tensor, h, w, hw, depth, batch_size, |
| 98 | base_tag, &s); |
| 99 | } else if (tensor.dtype() == DT_FLOAT) { |
| 100 | NormalizeAndAddImages<float>(c, tensor, h, w, hw, depth, batch_size, |
| 101 | base_tag, &s); |
| 102 | } else { // tensor.dtype() = DT_DOUBLE |
| 103 | NormalizeAndAddImages<double>(c, tensor, h, w, hw, depth, batch_size, |
| 104 | base_tag, &s); |
| 105 | } |
| 106 | |
| 107 | Tensor* summary_tensor = nullptr; |
| 108 | OP_REQUIRES_OK(c, c->allocate_output(0, TensorShape({}), &summary_tensor)); |
| 109 | CHECK(SerializeToTString(s, &summary_tensor->scalar<tstring>()())); |
nothing calls this directly
no test coverage detected