Python3.11运行pix2pix-zero遇TypeError：deprecate()不支持message参数

错误类型 #

TypeError: deprecate() got an unexpected keyword argument 'message'

错误详情 #

Fetching 16 files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████| 16/16 [00:00<00:00, 66117.11it/s]
  0%|                                                                                                                                          | 0/50 [00:01<?, ?it/s]
Traceback (most recent call last):
  File "/data1/wz/research/pix2pix-zero/src/inversion.py", line 58, in <module>
    x_inv, x_inv_image, x_dec_img = pipe(
                                    ^^^^^
  File "/data1/wz/research/pix2pix-zero/src/utils/ddim_inv.py", line 106, in __call__
    noise_pred = self.unet(latent_model_input,t,encoder_hidden_states=prompt_embeds,cross_attention_kwargs=cross_attention_kwargs,).sample
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/data1/wz/anaconda3/envs/pix2pix-zero/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/data1/wz/anaconda3/envs/pix2pix-zero/lib/python3.11/site-packages/diffusers/models/unet_2d_condition.py", line 582, in forward
    sample, res_samples = downsample_block(
                          ^^^^^^^^^^^^^^^^^
  File "/data1/wz/anaconda3/envs/pix2pix-zero/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/data1/wz/anaconda3/envs/pix2pix-zero/lib/python3.11/site-packages/diffusers/models/unet_2d_blocks.py", line 837, in forward
    hidden_states = attn(
                    ^^^^^
  File "/data1/wz/anaconda3/envs/pix2pix-zero/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/data1/wz/anaconda3/envs/pix2pix-zero/lib/python3.11/site-packages/diffusers/models/transformer_2d.py", line 265, in forward
    hidden_states = block(
                    ^^^^^^
  File "/data1/wz/anaconda3/envs/pix2pix-zero/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/data1/wz/anaconda3/envs/pix2pix-zero/lib/python3.11/site-packages/diffusers/models/attention.py", line 291, in forward
    attn_output = self.attn1(
                  ^^^^^^^^^^^
  File "/data1/wz/anaconda3/envs/pix2pix-zero/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/data1/wz/anaconda3/envs/pix2pix-zero/lib/python3.11/site-packages/diffusers/models/cross_attention.py", line 205, in forward
    return self.processor(
           ^^^^^^^^^^^^^^^
  File "/data1/wz/research/pix2pix-zero/src/utils/cross_attention.py", line 7, in __call__
    attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/data1/wz/anaconda3/envs/pix2pix-zero/lib/python3.11/site-packages/diffusers/models/cross_attention.py", line 260, in prepare_attention_mask
    deprecate(
TypeError: deprecate() got an unexpected keyword argument 'message'

涉及的cross_attention.py函数 #

def prepare_attention_mask(self, attention_mask, target_length, batch_size=None):
        if batch_size is None:
            deprecate(
                "batch_size=None",
                "0.0.15",
                message=(
                    "Not passing the `batch_size` parameter to `prepare_attention_mask` can lead to incorrect"
                    " attention mask preparation and is deprecated behavior. Please make sure to pass `batch_size` to"
                    " `prepare_attention_mask` when preparing the attention_mask."
                ),
            )
            batch_size = 1

        head_size = self.heads
        if attention_mask is None:
            return attention_mask

        if attention_mask.shape[-1] != target_length:
            if attention_mask.device.type == "mps":
                # HACK: MPS: Does not support padding by greater than dimension of input tensor.
                # Instead, we can manually construct the padding tensor.
                padding_shape = (attention_mask.shape[0], attention_mask.shape[1], target_length)
                padding = torch.zeros(padding_shape, dtype=attention_mask.dtype, device=attention_mask.device)
                attention_mask = torch.cat([attention_mask, padding], dim=2)
            else:
                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)

        if attention_mask.shape[0] < batch_size * head_size:
            attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
        return attention_mask

问题原因与修复 #

这个错误不是Python 3.11版本不兼容导致的，核心原因是你安装的diffusers库版本和代码中调用的deprecate函数参数不匹配：代码里的prepare_attention_mask函数调用deprecate时传入了message参数，但当前环境的diffusers版本中，deprecate函数并不支持该参数。

修复方案： #

1. 升级diffusers到兼容版本
   执行命令更新到最新稳定版：

   pip install --upgrade diffusers
   

   或者指定一个已知兼容的版本（比如0.19.0及以上，可根据pix2pix-zero项目要求调整）：

   pip install diffusers==0.19.0
   

2. 手动修改调用逻辑
   如果不想升级库，可以修改cross_attention.py中的代码，去掉message参数：

   # 修改前
   deprecate(
       "batch_size=None",
       "0.0.15",
       message="..."
   )
   # 修改后
   deprecate("batch_size=None", "0.0.15")
   

3. 传递batch_size参数
   既然错误是因为没传batch_size触发废弃警告导致的，也可以直接在调用prepare_attention_mask时传入batch_size，避免触发该调用：
   在/data1/wz/research/pix2pix-zero/src/utils/cross_attention.py第7行修改为：

   # 从attention_mask推导batch_size
   batch_size = attention_mask.shape[0] // attn.heads if attention_mask is not None else 1
   attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size=batch_size)
   

内容的提问来源于stack exchange，提问作者Our bank