johndpope
commented
3 weeks ago it seems the onnx has problem with fusedleakyrelu -
related - https://github.com/rosinality/stylegan2-pytorch/issues/322
am using this same function in own project - https://github.com/bes-dev/MobileStyleGAN.pytorch/blob/a9776ff8f05a868b2d3b637bda14eca4c074d2a3/core/models/modules/ops/fused_act.py#L22
i have this class that spits out onnx - the opsets are very important as lower versions dont have the support https://github.com/johndpope/IMF/blob/main/onnxconv.py#L153
but this didn't resolve things for me yet.
still digging....
wait a second maybe this recent upstream fix on pytorch solves things - https://github.com/pytorch/pytorch/issues/125753 need to test
pip install onnxscript -U
pip install onnxruntime -U
pip install onnxconverter_common -U
pip install onnx -UN.B. - this library has a diagnosis for onnx https://github.com/webonnx/wonnx
cargo install --git https://github.com/webonnx/wonnx.git wonnx-cli
nnx info ./data/models/opt-squeeze.onnx
[2024-10-10T21:07:06Z ERROR nnx::info] Node '/latent_token_encoder/linear_layers.2_1/Gemm' input '/latent_token_encoder/activation_7/LeakyRelu_output_0' has unknown shape
[2024-10-10T21:07:06Z ERROR nnx::info] Node '/latent_token_encoder/activation_8/LeakyRelu' input '/latent_token_encoder/linear_layers.2_1/Gemm_output_0' has unknown shape
[2024-10-10T21:07:06Z ERROR nnx::info] Node '/latent_token_encoder/linear_layers.3_1/Gemm' input '/latent_token_encoder/activation_8/LeakyRelu_output_0' has unknown shape
[2024-10-10T21:07:06Z ERROR nnx::info] Node '/latent_token_encoder/activation_9/LeakyRelu' input '/latent_token_encoder/linear_layers.3_1/Gemm_output_0' has unknown shape
[2024-10-10T21:07:06Z ERROR nnx::info] Node '/latent_token_encoder/final_linear_1/Gemm' input '/latent_token_encoder/activation_9/LeakyRelu_output_0' has unknown shape
+------------------+-------------------------------------------------------------------+
| Model version | 0 |
+------------------+-------------------------------------------------------------------+
| IR version | 8 |
+------------------+-------------------------------------------------------------------+
| Producer name | pytorch |
+------------------+-------------------------------------------------------------------+
| Producer version | 2.4.0 |
+------------------+-------------------------------------------------------------------+
| Opsets | 15 |
+------------------+-------------------------------------------------------------------+
| Inputs | +-------------+-------------+----------------------------+------+ |
| | | Name | Description | Shape | Type | |
| | +-------------+-------------+----------------------------+------+ |
| | | x_current | | batch_size x 3 x 256 x 256 | f32 | |
| | +-------------+-------------+----------------------------+------+ |
| | | x_reference | | batch_size x 3 x 256 x 256 | f32 | |
| | +-------------+-------------+----------------------------+------+ |
+------------------+-------------------------------------------------------------------+
| Outputs | +------+-------------+----------------------------+------+ |
| | | Name | Description | Shape | Type | |
| | +------+-------------+----------------------------+------+ |
| | | f_r | | batch_size x 128 x 64 x 64 | f32 | |
| | +------+-------------+----------------------------+------+ |
| | | t_r | | batch_size x 256 x 32 x 32 | f32 | |
| | +------+-------------+----------------------------+------+ |
| | | t_c | | batch_size x 512 x 16 x 16 | f32 | |
| | +------+-------------+----------------------------+------+ |
| | | x | | batch_size x 512 x 8 x 8 | f32 | |
| | +------+-------------+----------------------------+------+ |
| | | 3032 | | Gemm3032_dim_0 x 32 | f32 | |
| | +------+-------------+----------------------------+------+ |
| | | 4799 | | Gemm4799_dim_0 x 32 | f32 | |
| | +------+-------------+----------------------------+------+ |
+------------------+-------------------------------------------------------------------+
| Ops used | +-----------------+---------------------+ |
| | | Op | Attributes | |
| | +-----------------+---------------------+ |
| | | Add | | |
| | +-----------------+---------------------+ |
| | | Cast | to=7 | |
| | +-----------------+---------------------+ |
| | | Concat | axis=0 | |
| | +-----------------+---------------------+ |
| | | Constant | value=<TENSOR> | |
| | +-----------------+---------------------+ |
| | | ConstantOfShape | value=<TENSOR> | |
| | +-----------------+---------------------+ |
| | | Conv | dilations=<INTS> | |
| | | | group=1 | |
| | | | kernel_shape=<INTS> | |
| | | | pads=<INTS> | |
| | | | strides=<INTS> | |
| | +-----------------+---------------------+ |
| | | Div | | |
| | +-----------------+---------------------+ |
| | | Gather | axis=0 | |
| | +-----------------+---------------------+ |
| | | Gemm | alpha=1 | |
| | | | beta=1 | |
| | | | transB=1 | |
| | +-----------------+---------------------+ |
| | | Identity | | |
| | +-----------------+---------------------+ |
| | | LeakyRelu | alpha=0.2 | |
| | +-----------------+---------------------+ |
| | | Mul | | |
| | +-----------------+---------------------+ |
| | | Pad | mode=constant | |
| | +-----------------+---------------------+ |
| | | ReduceMean | axes=<INTS> | |
| | | | keepdims=0 | |
| | +-----------------+---------------------+ |
| | | Relu | | |
| | +-----------------+---------------------+ |
| | | Reshape | allowzero=0 | |
| | +-----------------+---------------------+ |
| | | Shape | | |
| | +-----------------+---------------------+ |
| | | Slice | | |
| | +-----------------+---------------------+ |
| | | Sub | | |
| | +-----------------+---------------------+ |
| | | Transpose | perm=<INTS> | |
| | +-----------------+---------------------+ |
| | | Unsqueeze | | |
| | +-----------------+---------------------+ |
+------------------+-------------------------------------------------------------------+
| Memory usage | +--------------+-----------+ |
| | | Inputs | 0 B | |
| | +--------------+-----------+ |
| | | Outputs | 0 B | |
| | +--------------+-----------+ |
| | | Intermediate | 0 B | |
| | +--------------+-----------+ |
| | | Weights | 204.4 MiB | |
| | +--------------+-----------+ |
| | | Total | 204.4 MiB | |
| | +--------------+-----------+ |
+------------------+-------------------------------------------------------------------+
Hi!
Thank you for an amazing repository. I successfully converted my StyleGAN2-ada rosinality checkpoint, by running the following line:
python convert_rosinality_ckpt.py --ckpt {path_to_rosinality_stylegan2_ckpt} --ckpt-mnet output/mnet.ckpt --ckpt-snet output/snet.ckpt --cfg-path output/config.jsonI tested the checkpoint with
demo.pyand it produces images as expected.I then converted it to ONNX by running
python train.py --cfg output/config.json --export-model onnx --export-dir onnx-2and tried to use the converted checkpoint in MobileStyleGAN web demo (https://github.com/cyrildiagne/mobilestylegan-web-demo). It produces uniform grey images for all seeds. The web demo works fine with the authors' ffhq checkpoint so it seems to be an issue with the converted model.Do you have any thoughts on what might be causing this?