Can not load pretrained model properly #3090
Description
Is your feature request related to a problem? Please describe.
I use pipeline.save_pretrained(config.output_dir, variant=str(epoch)) to save model every n epochs, this leads to there are multiple model in the unet folder:
Then when I want to load pretrained model using pipeline = DDPMPipeline.from_pretrained('./eye-144192_100_4atten', local_files_only=True), it throw error:
╭─────────────────────────────── Traceback (most recent call last) ────────────────────────────────╮
│ in <module> │
│ │
│ /opt/conda/lib/python3.8/site-packages/diffusers/pipelines/pipeline_utils.py:944 in │
│ from_pretrained │
│ │
│ 941 │ │ │ │ │
│ 942 │ │ │ │ # check if the module is in a subdirectory │
│ 943 │ │ │ │ if os.path.isdir(os.path.join(cached_folder, name)): │
│ ❱ 944 │ │ │ │ │ loaded_sub_model = load_method(os.path.join(cached_folder, name), ** │
│ 945 │ │ │ │ else: │
│ 946 │ │ │ │ │ # else load from the root directory │
│ 947 │ │ │ │ │ loaded_sub_model = load_method(cached_folder, **loading_kwargs) │
│ │
│ /opt/conda/lib/python3.8/site-packages/diffusers/models/modeling_utils.py:527 in from_pretrained │
│ │
│ 524 │ │ │ │ except: # noqa: E722 │
│ 525 │ │ │ │ │ pass │
│ 526 │ │ │ if model_file is None: │
│ ❱ 527 │ │ │ │ model_file = _get_model_file( │
│ 528 │ │ │ │ │ pretrained_model_name_or_path, │
│ 529 │ │ │ │ │ weights_name=_add_variant(WEIGHTS_NAME, variant), │
│ 530 │ │ │ │ │ cache_dir=cache_dir, │
│ │
│ /opt/conda/lib/python3.8/site-packages/diffusers/models/modeling_utils.py:821 in _get_model_file │
│ │
│ 818 │ │ │ model_file = os.path.join(pretrained_model_name_or_path, subfolder, weights_ │
│ 819 │ │ │ return model_file │
│ 820 │ │ else: │
│ ❱ 821 │ │ │ raise EnvironmentError( │
│ 822 │ │ │ │ f"Error no file named {weights_name} found in directory {pretrained_mode │
│ 823 │ │ │ ) │
│ 824 │ else: │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
OSError: Error no file named diffusion_pytorch_model.bin found in directory ./eye-144192_100_4atten/unet.
Describe the solution you'd like
diffusers can automatically choose the last model, e.g. here is eye-144192_100/unet/diffusion_pytorch_model.99.bin
Describe alternatives you've considered
User specify the model , e.g. pipeline = DDPMPipeline.from_pretrained('./eye-144192_100_4atten/unet/diffusion_pytorch_model.99.bin', local_files_only=True)