Converts a UniDiffuser autoencoder_kl.pth checkpoint to a diffusers AutoencoderKL.
(ckpt, diffusers_model, num_head_channels=1)
| 346 | |
| 347 | # Based on diffusers.pipelines.stable_diffusion.convert_from_ckpt.convert_ldm_vae_checkpoint |
| 348 | def convert_vae_to_diffusers(ckpt, diffusers_model, num_head_channels=1): |
| 349 | """ |
| 350 | Converts a UniDiffuser autoencoder_kl.pth checkpoint to a diffusers AutoencoderKL. |
| 351 | """ |
| 352 | # autoencoder_kl.pth ckpt is a torch state dict |
| 353 | vae_state_dict = torch.load(ckpt, map_location="cpu") |
| 354 | |
| 355 | new_checkpoint = {} |
| 356 | |
| 357 | new_checkpoint["encoder.conv_in.weight"] = vae_state_dict["encoder.conv_in.weight"] |
| 358 | new_checkpoint["encoder.conv_in.bias"] = vae_state_dict["encoder.conv_in.bias"] |
| 359 | new_checkpoint["encoder.conv_out.weight"] = vae_state_dict["encoder.conv_out.weight"] |
| 360 | new_checkpoint["encoder.conv_out.bias"] = vae_state_dict["encoder.conv_out.bias"] |
| 361 | new_checkpoint["encoder.conv_norm_out.weight"] = vae_state_dict["encoder.norm_out.weight"] |
| 362 | new_checkpoint["encoder.conv_norm_out.bias"] = vae_state_dict["encoder.norm_out.bias"] |
| 363 | |
| 364 | new_checkpoint["decoder.conv_in.weight"] = vae_state_dict["decoder.conv_in.weight"] |
| 365 | new_checkpoint["decoder.conv_in.bias"] = vae_state_dict["decoder.conv_in.bias"] |
| 366 | new_checkpoint["decoder.conv_out.weight"] = vae_state_dict["decoder.conv_out.weight"] |
| 367 | new_checkpoint["decoder.conv_out.bias"] = vae_state_dict["decoder.conv_out.bias"] |
| 368 | new_checkpoint["decoder.conv_norm_out.weight"] = vae_state_dict["decoder.norm_out.weight"] |
| 369 | new_checkpoint["decoder.conv_norm_out.bias"] = vae_state_dict["decoder.norm_out.bias"] |
| 370 | |
| 371 | new_checkpoint["quant_conv.weight"] = vae_state_dict["quant_conv.weight"] |
| 372 | new_checkpoint["quant_conv.bias"] = vae_state_dict["quant_conv.bias"] |
| 373 | new_checkpoint["post_quant_conv.weight"] = vae_state_dict["post_quant_conv.weight"] |
| 374 | new_checkpoint["post_quant_conv.bias"] = vae_state_dict["post_quant_conv.bias"] |
| 375 | |
| 376 | # Retrieves the keys for the encoder down blocks only |
| 377 | num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "encoder.down" in layer}) |
| 378 | down_blocks = { |
| 379 | layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks) |
| 380 | } |
| 381 | |
| 382 | # Retrieves the keys for the decoder up blocks only |
| 383 | num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "decoder.up" in layer}) |
| 384 | up_blocks = { |
| 385 | layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks) |
| 386 | } |
| 387 | |
| 388 | for i in range(num_down_blocks): |
| 389 | resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key] |
| 390 | |
| 391 | if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict: |
| 392 | new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop( |
| 393 | f"encoder.down.{i}.downsample.conv.weight" |
| 394 | ) |
| 395 | new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop( |
| 396 | f"encoder.down.{i}.downsample.conv.bias" |
| 397 | ) |
| 398 | |
| 399 | paths = renew_vae_resnet_paths(resnets) |
| 400 | meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"} |
| 401 | assign_to_checkpoint( |
| 402 | paths, |
| 403 | new_checkpoint, |
| 404 | vae_state_dict, |
| 405 | additional_replacements=[meta_path], |
no test coverage detected