add image with raw pixel data

[Mirocos]

Hi, I read all codes in the test file that I actually didn't find an example to save images with its raw pixel data.
Does it already supported or not?

[RustingSword]

Not yet, it has to be encoded into one of the supported formats first.

[Mirocos]

Hi, I was trying to encode raw pixels data to png format and send it to add_image.
My code looks like below:

            int out_len = 0;
            int width = t.size(0), height = t.size(1), channel = t.size(2);
            auto encoded_bytes = stbi_write_png_to_mem((const unsigned char*)v.data(), 0, height, width, channel, &out_len);
            std::string png(encoded_bytes, encoded_bytes+out_len);
            logger.add_image("Tensor raw data plugin", 1, png, height, width, channel, "imrod");
            std::ofstream file("./demo/imord.png", std::ios::binary);
            file << png;

where stbi_write_png_to_mem is a function from stbi_image project,.

I found that the imord.png is correct which ouput by the binary stream. But the file output by the logger, say imrod.pb seems to be wrong. Because tensorboard tell that no dashboard active data in it.

[RustingSword]

Thanks for the feedback, I will try to look into this as soon as possible. Maybe you can do a quick inspection to do a round trip by stbi_load(some_valid_png_file) then stbi_write_png_to_mem to see what's wrong. If this works, then the diff comes from the raw pixel data; if this doesn't work, maybe stbi encodes png file differently?

[RustingSword]

I just tried this, and it worked:

diff --git a/CMakeLists.txt b/CMakeLists.txt
index b99ba57..7b885d0 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -23,3 +23,7 @@ target_link_libraries(tensorboard_logger PUBLIC ${Protobuf_LIBRARIES})

 add_executable(tensorboard_logger_test tests/test_tensorboard_logger.cc)
 target_link_libraries(tensorboard_logger_test tensorboard_logger)
+target_include_directories(tensorboard_logger_test PUBLIC
+    ../stb/deprecated/
+    ../stb/
+)
diff --git a/tests/test_tensorboard_logger.cc b/tests/test_tensorboard_logger.cc
index d9f1f3e..bde9d07 100644
--- a/tests/test_tensorboard_logger.cc
+++ b/tests/test_tensorboard_logger.cc
@@ -5,6 +5,14 @@
 #include <vector>

 #include "tensorboard_logger.h"
+#define STB_IMAGE_WRITE_IMPLEMENTATION
+#include "stb_image_write.h"
+
+#define STB_IMAGE_IMPLEMENTATION
+#include "stb_image.h"
+
+#define STB_DEFINE
+#include "stb.h"

 using namespace std;

@@ -57,6 +65,21 @@ int test_log_image(TensorBoardLogger& logger) {
     logger.add_image("TensorBoard Audo Plugin", 1, image2, 1766, 814, 3,
                      "TensorBoard", "Audio");

+    // add image from memory
+    int x, y, c, size;
+    const char* test_file = "tux.png";
+    auto image3 = stbi_load(test_file, &x, &y, &c, 0);
+    auto encoded = stbi_write_png_to_mem(image3, 0, x, y, c, &size);
+    cout << test_file << " x=" << x << " y=" << y << " c=" << c <<
+        " re-encoded size=" << size << endl;
+    string png(encoded, encoded + size);
+    logger.add_image("image from memory", 1, png, x, y, c, "Tux");
+    ofstream file("reencoded_tux.png", ios::binary);
+    file << png;
+    file.close();
+    stbi_image_free(image3);
+    STBIW_FREE(encoded);
+
     // add multiple images
     // FIXME This seems doesn't work anymore.
     // logger.add_images(

Test file is this one https://cdn.freebiesupply.com/logos/large/2x/tux-logo-png-transparent.png

[Mirocos]

Hi, thanks for that check. I use the reference picture tux.png and load it first than write it to encoded memory. I don't what happened that I open the target imrod.pb in hex format and all I saw is like below snap.

[Mirocos]

I know what happened why tensorboard didn't detect the event file. From this issue #13, I found that my target ouput filename did not have a substring like tfevents that is what tensorboard seach for. Sorry to waste your time, my apologies

[RustingSword]

🤦 I'm planning to add a warning message if the file name is not valid.

[Mirocos]

What about assert if the filename does not have substring "tfevents"?

[RustingSword]

Yes, on second thought I think throwing a runtime_error is more appropriate than a warning message. Just pushed the code, thanks for the suggestion.