正如我在这里概述的那样,由于硬件(例如),我一直在使用旧版本的 pytorch 和 torchvision。使用 ppc64le IBM 架构。
因此,我在不同计算机、集群和我的个人 Mac 之间发送和接收检查点时遇到问题。我想知道是否有什么方法可以加载模型来避免这个问题?例如使用 1.6.x 时,可能会以新旧格式保存模型。当然,从 1.3.1 到 1.6.x 是不可能的,但至少我希望有些东西能起作用。
有什么建议吗?当然,我理想的解决方案是我不必担心它,我总是可以加载和保存我的检查点以及我通常在所有硬件上统一腌制的所有内容。
我遇到的第一个错误是 zip jit 错误:
RuntimeError: /home/miranda9/data/f.pt is a zip archive (did you mean to use torch.jit.load()?)
所以我使用了它(和其他pickle库):
# %%
import torch
from pathlib import Path
def load(path):
import torch
import pickle
import dill
path = str(path)
try:
db = torch.load(path)
f = db['f']
except Exception as e:
db = torch.jit.load(path)
f = db['f']
#with open():
# db = pickle.load(open(path, "r+"))
# db = dill.load(open(path, "r+"))
#raise ValueError(f'FAILED: {e}')
return db, f
p = "~/data/f.pt"
path = Path(p).expanduser()
db, f = load(path)
Din, nb_examples = 1, 5
x = torch.distributions.Normal(loc=0.0, scale=1.0).sample(sample_shape=(nb_examples, Din))
y = f(x)
print(y)
print('Success!\a')
但是我收到了关于我被迫使用的不同 pytorch 版本的抱怨:
Traceback (most recent call last):
File "hal_pg.py", line 27, in <module>
db, f = load(path)
File "hal_pg.py", line 16, in load
db = torch.jit.load(path)
File "/home/miranda9/.conda/envs/wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/jit/__init__.py", line 239, in load
cpp_module = torch._C.import_ir_module(cu, f, map_location, _extra_files)
RuntimeError: version_number <= kMaxSupportedFileFormatVersion INTERNAL ASSERT FAILED at /opt/anaconda/conda-bld/pytorch-base_1581395437985/work/caffe2/serialize/inline_container.cc:131, please report a bug to PyTorch. Attempted to read a PyTorch file with version 3, but the maximum supported version for reading is 1. Your PyTorch installation may be too old. (init at /opt/anaconda/conda-bld/pytorch-base_1581395437985/work/caffe2/serialize/inline_container.cc:131)
frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&) + 0xbc (0x7fff7b527b9c in /home/miranda9/.conda/envs/wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libc10.so)
frame #1: caffe2::serialize::PyTorchStreamReader::init() + 0x1d98 (0x7fff1d293c78 in /home/miranda9/.conda/envs/wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libtorch.so)
frame #2: caffe2::serialize::PyTorchStreamReader::PyTorchStreamReader(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&) + 0x88 (0x7fff1d2950d8 in /home/miranda9/.conda/envs/wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libtorch.so)
frame #3: torch::jit::import_ir_module(std::shared_ptr<torch::jit::script::CompilationUnit>, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, c10::optional<c10::Device>, std::unordered_map<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::hash<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > >, std::equal_to<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > >, std::allocator<std::pair<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > > > >&) + 0x64 (0x7fff1e624664 in /home/miranda9/.conda/envs/wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libtorch.so)
frame #4: <unknown function> + 0x70e210 (0x7fff7c0ae210 in /home/miranda9/.conda/envs/wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libtorch_python.so)
frame #5: <unknown function> + 0x28efc4 (0x7fff7bc2efc4 in /home/miranda9/.conda/envs/wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libtorch_python.so)
<omitting python frames>
frame #26: <unknown function> + 0x25280 (0x7fff84b35280 in /lib64/libc.so.6)
frame #27: __libc_start_main + 0xc4 (0x7fff84b35474 in /lib64/libc.so.6)
有什么想法可以使集群中的一切保持一致吗?我什至无法打开 pickle 文件。
也许对于我被迫使用的当前 pytorch 版本来说这是不可能的:(
RuntimeError: version_number <= kMaxSupportedFileFormatVersion INTERNAL ASSERT FAILED at /opt/anaconda/conda-bld/pytorch-base_1581395437985/work/caffe2/serialize/inline_container.cc:131, please report a bug to PyTorch. Attempted to read a PyTorch file with version 3, but the maximum supported version for reading is 1. Your PyTorch installation may be too old. (init at /opt/anaconda/conda-bld/pytorch-base_1581395437985/work/caffe2/serialize/inline_container.cc:131)
frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&) + 0xbc (0x7fff83ba7b9c in /home/miranda9/.conda/envs/automl-meta-learning_wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libc10.so)
frame #1: caffe2::serialize::PyTorchStreamReader::init() + 0x1d98 (0x7fff25993c78 in /home/miranda9/.conda/envs/automl-meta-learning_wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libtorch.so)
frame #2: caffe2::serialize::PyTorchStreamReader::PyTorchStreamReader(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&) + 0x88 (0x7fff259950d8 in /home/miranda9/.conda/envs/automl-meta-learning_wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libtorch.so)
frame #3: torch::jit::import_ir_module(std::shared_ptr<torch::jit::script::CompilationUnit>, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, c10::optional<c10::Device>, std::unordered_map<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::hash<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > >, std::equal_to<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > >, std::allocator<std::pair<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > > > >&) + 0x64 (0x7fff26d24664 in /home/miranda9/.conda/envs/automl-meta-learning_wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libtorch.so)
frame #4: <unknown function> + 0x70e210 (0x7fff8472e210 in /home/miranda9/.conda/envs/automl-meta-learning_wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libtorch_python.so)
frame #5: <unknown function> + 0x28efc4 (0x7fff842aefc4 in /home/miranda9/.conda/envs/automl-meta-learning_wmlce-v1.7.0-py3.7/lib/python3.7/site-packages/torch/lib/libtorch_python.so)
<omitting python frames>
frame #23: <unknown function> + 0x25280 (0x7fff8d335280 in /lib64/libc.so.6)
frame #24: __libc_start_main + 0xc4 (0x7fff8d335474 in /lib64/libc.so.6)
使用代码:
from pathlib import Path
import torch
path = '/home/miranda9/data/dataset/'
path = Path(path).expanduser() / 'fi_db.pt'
path = str(path)
# db = torch.load(path)
# torch.jit.load(path)
db = torch.jit.load(str(path))
print(db)
相关链接:
我相信开发人员的目的是传递一个保存为泡菜的标志。只是默认行为的改变。
对于之前设置检查点的文件,在新环境中重新加载 zip 文件保存的权重(pytorch>=1.6),然后再次检查点作为 pickle(无需重新训练);
更新您的代码并从下次开始添加标志
我们已将 torch.save 切换为默认使用基于 zip 文件的格式 而不是旧的基于 Pickle 的格式。 torch.load 保留了 能够加载旧格式,但使用新格式是 受到推崇的。新格式是:
对于检查和构建用于操作的工具更加友好 保存文件修复了一个长期存在的问题,其中序列化 (getstate, setstate) 依赖于模块的函数 序列化的张量值得到了错误的数据,与 TorchScript序列化格式,使序列化更加一致 跨 PyTorch
使用方法如下:
m = MyMod()
torch.save(m.state_dict(), 'mymod.pt') # Saves a zipfile to mymod.pt
要使用旧格式,请传递标志
_use_new_zipfile_serialization=False
m = MyMod()
torch.save(m.state_dict(), 'mymod.pt', _use_new_zipfile_serialization=False) # Saves pickle
这不是一个理想的解决方案,但它适用于将检查点从新版本转移到旧版本。
我也使用ppc64le并面临同样的问题。可以将模型保存为任何 PyTorch 版本都可以读取的文本格式。我在 ppc64le 机器上安装了 PyTorch v1.3.0,在我的笔记本上安装了 v1.7.0(不需要显卡)。
第1步.通过较新的PyTorch版本保存模型
def save_model_txt(model, path):
fout = open(path, 'w')
for k, v in model.state_dict().items():
fout.write(str(k) + '\n')
fout.write(str(v.tolist()) + '\n')
fout.close()
在保存之前,我像这样加载模型
checkpoint = torch.load(path, map_location=torch.device('cpu'))
model.load_state_dict(checkpoint, strict=False)
第2步.传输文本文件
步骤 3. 在旧 PyTorch 中加载文本文件
def load_model_txt(model, path):
data_dict = {}
fin = open(path, 'r')
i = 0
odd = 1
prev_key = None
while True:
s = fin.readline().strip()
if not s:
break
if odd:
prev_key = s
else:
print('Iter', i)
val = eval(s)
if type(val) != type([]):
data_dict[prev_key] = torch.FloatTensor([eval(s)])[0]
else:
data_dict[prev_key] = torch.FloatTensor(eval(s))
i += 1
odd = (odd + 1) % 2
# Replace existing values with loaded
print('Loading...')
own_state = model.state_dict()
print('Items:', len(own_state.items()))
for k, v in data_dict.items():
if not k in own_state:
print('Parameter', k, 'not found in own_state!!!')
else:
try:
own_state[k].copy_(v)
except:
print('Key:', k)
print('Old:', own_state[k])
print('New:', v)
sys.exit(0)
print('Model loaded')
加载前必须初始化模型。空模型被传递到函数中。
限制
如果您的模型 state_dict 包含 (str: torch.Tensor) 值以外的其他内容,则此方法将不起作用。您可以使用
检查您的 state_dict 内容for k, v in model.state_dict().items():
...
阅读以下内容以了解:
https://pytorch.org/tutorials/recipes/recipes/ saving_and_loading_models_for_inference.html
https://discuss.pytorch.org/t/how-to-load-part-of-pre-trained-model/1113
基于 @maxim velikanov 的答案,我创建了一个单独的 OrderedDict,其中键与模型的原始状态字典相同,但每个张量值都转换为列表。
这个 OrderedDict 是它们转储到 JSON 文件中的。
def save_model_json(model, path):
actual_dict = OrderedDict()
for k, v in model.state_dict().items():
actual_dict[k] = v.tolist()
with open(path, 'w') as f:
json.dump(actual_dict, f)
然后,加载器可以将文件作为 JSON 加载,并且每个列表/整数将在将其值复制到原始状态字典之前转换回张量。
def load_model_json(model, path):
data_dict = OrderedDict()
with open(path, 'r') as f:
data_dict = json.load(f)
own_state = model.state_dict()
for k, v in data_dict.items():
print('Loading parameter:', k)
if not k in own_state:
print('Parameter', k, 'not found in own_state!!!')
if type(v) == list or type(v) == int:
v = torch.tensor(v)
own_state[k].copy_(v)
model.load_state_dict(own_state)
print('Model loaded')
我在加载处理后的数据时遇到了类似的问题。我之前在 torch 1.8 中将数据保存为“xxx.pt”,但在 torch 1.2 中将其加载。即使通过 torch.jit.load() 我也无法成功加载它。我唯一的解决办法是在旧版本中再次保存数据。