Crash when loading a 4 channel image when channel option is forced to 3 in StbImage.read_file

[ChristopheBelpaire]

Hello,
I'm reading the awesome book Machine learning with elixir and
on my M1 macbook with erlang 26.2.2 and elixir 1.16.1,
the following script crashes (from chapter 7 of the book, the dataset comes from kaggle : https://www.kaggle.com/datasets/shaunthesheep/microsoft-catsvsdogs-dataset) :

Mix.install([{:axon, "~> 0.6"}, {:nx, path: "./nx/nx", override: true}, {:exla, path: "./nx/exla", override: true}, {:stb_image, "0.5.2"}, {:kino, "~> 0.8"}], system_env: %{"DEBUG" => 1})
Process.sleep(1000)
Nx.tensor(1, backend: EXLA.Backend)


defmodule CatsAndDogs do
  def pipeline(paths, batch_size, target_height, target_width) do
    paths
    |> Enum.shuffle()
    |> Task.async_stream(&parse_image/1)
    |> Stream.filter(fn
       {:ok, {%StbImage{shape: {_, _, 3}}, _}} -> true
       _ -> false end)
    |> Stream.map(&to_tensors(&1, target_height, target_width))
    |> Stream.chunk_every(batch_size, batch_size, :discard)
    |> Stream.map(fn chunks ->
      {img_chunk, label_chunk} = Enum.unzip(chunks)
      {Nx.stack(img_chunk), Nx.stack(label_chunk)}
    end)
  end

  defp parse_image(path) do
   label = if String.contains?(path, "Cat"), do: 0, else: 1
   case StbImage.read_file(path, channels: 3) do
      {:ok, img} -> {img, label}
      _error -> :error
    end
  end

  defp to_tensors({:ok, {img, label}}, target_height, target_width) do img_tensor =
    img
    |> StbImage.resize(target_height, target_width)
    |> StbImage.to_nx()
    |> Nx.divide(255)
    label_tensor = Nx.tensor([label])
    {img_tensor, label_tensor}
  end

end


{test_paths, train_paths} =(Path.wildcard("/Users/christophebelpaire/perso/machine-learning-in-elixir/PetImages/Cat/*.jpg")
++ Path.wildcard("/Users/christophebelpaire/perso/machine-learning-in-elixir/PetImages/Dog/*.jpg"))
|> Enum.shuffle()
|> Enum.split(1000)


target_height = 96
target_width = 96
batch_size = 32

train_pipeline = CatsAndDogs.pipeline(
  train_paths, batch_size, target_height, target_width
)
test_pipeline = CatsAndDogs.pipeline(
  test_paths, batch_size, target_height, target_width
)

mlp_model =
  Axon.input("images", shape: {nil, target_height, target_width, 3}) |> Axon.flatten()
  |> Axon.dense(256, activation: :relu)
  |> Axon.dense(128, activation: :relu)
  |> Axon.dense(1, activation: :sigmoid)


  IO.gets("Press enter to continue - #{System.pid()}")

  _mlp_trained_model_state =
  mlp_model
  |> Axon.Loop.trainer(:binary_cross_entropy, :adam)
  |> Axon.Loop.metric(:accuracy)
  |> Axon.Loop.run(train_pipeline, %{}, epochs: 5, compiler: EXLA)

Here is the debug message from lldb :

Architecture set to: arm64-apple-macosx-.
(lldb) continue
Process 13117 resuming
Process 13117 stopped
* thread #25, name = 'erts_dios_1', stop reason = EXC_BAD_ACCESS (code=1, address=0x14a394000)
    frame #0: 0x0000000189b496b0 libsystem_platform.dylib`_platform_memmove + 96
libsystem_platform.dylib`:
->  0x189b496b0 <+96>:  ldnp   q0, q1, [x1]
    0x189b496b4 <+100>: add    x1, x1, #0x20
    0x189b496b8 <+104>: subs   x2, x2, #0x20
    0x189b496bc <+108>: b.hi   0x189b496a8               ; <+88>
Target 0: (beam.smp) stopped.
(lldb) bt
* thread #25, name = 'erts_dios_1', stop reason = EXC_BAD_ACCESS (code=1, address=0x14a394000)
  * frame #0: 0x0000000189b496b0 libsystem_platform.dylib`_platform_memmove + 96
    frame #1: 0x0000000148127134 stb_image_nif.so`pack_data + 96
    frame #2: 0x0000000148125c08 stb_image_nif.so`read_file + 812
    frame #3: 0x0000000102d746c8 beam.smp`erts_call_dirty_nif(esdp=0x00000001439c1c80, c_p=0x0000000145eb12b8, I=0x00000001502b7278, reg=0x000000010308c5c0) at erl_nif.c:466:18 [opt]
    frame #4: 0x0000000102c4e0e0 beam.smp`erts_dirty_process_main(esdp=0x00000001439c1c80) at beam_common.c:280:23 [opt]
    frame #5: 0x0000000102bc68e8 beam.smp`sched_dirty_io_thread_func(vesdp=0x00000001439c1c80) at erl_process.c:8768:5 [opt]
    frame #6: 0x0000000102e13cbc beam.smp`thr_wrapper(vtwd=0x000000016d242560) at ethread.c:116:25 [opt]
    frame #7: 0x0000000189b1a034 libsystem_pthread.dylib`_pthread_start + 136

[ChristopheBelpaire]

Ok, the problem comes from the way I'm loading the images StbImage.read_file(path, channels: 3)
Some of my images seems to have 4 channels, the crash seems to be caused by the fact the load is forced to 3 channel for a 4 channel image.

[cocoa-xu]

Hi @ChristopheBelpaire, thanks for reporting this issue. I'll try to fix this issue sometime this week!

[cocoa-xu]

$ file /Users/cocoa/Downloads/archive/PetImages/Cat/10404.jpg
/Users/cocoa/Downloads/archive/PetImages/Cat/10404.jpg: Adobe Photoshop Image, 432 x 363, RGB, 3x 8-bit channels

I found one in the dataset that can cause this issue, but it's definitely not a JPEG image and seems to have wrong metadata perhaps. I'll look into this further

[ChristopheBelpaire]

Awesome, thanks!

I removed the "bad" images with :

    |> Stream.filter(fn
       {:ok, {%StbImage{shape: {_, _, 3}}, _}} -> true
       _ -> false end)

To filter the channel, everything can be loaded with this filter.

[cocoa-xu]

Hi @ChristopheBelpaire, v0.6.6 is out and should fix this issue!

[ChristopheBelpaire]

Awesome!
I tried my scenario again and it seems to load all the images.
Thanks for the fix!