代码如下：
loss_fn = nn.MSELoss()
optimizer = nn.Adam(model.trainable_params(), 1e-8)
B = np.linspace(0.0001, 0.02, 300)
B = Tensor(B, dtype=mindspore.float32)
a = 1.0 - B
a_bar = ops.cumprod(a,0)
sqrt_a_bar = ops.sqrt(a_bar)
sqrt_one_minus_a_bar = ops.sqrt(1 - a_bar)
model.to_float(mindspore.float32)
loss_fn.to_float(mindspore.float32)
# 1\定义向前函数
def forward_fn(img, t, c, c_mask):
img_with_noise_t, targets_noise = q(img, t, sqrt_a_bar, sqrt_one_minus_a_bar)
logits_noise = model(img_with_noise_t, t, c, c_mask)
loss = loss_fn(logits_noise, targets_noise)
return loss, logits_noise
# 2\得到梯度函数
grad_fn = mindspore.value_and_grad(forward_fn, None, optimizer.parameters, has_aux=True)
# 3\定义一步训练函数
def train_step(img, t, c, c_mask):
output, inputs_gradient = grad_fn(img, t, c, c_mask)
# (loss, _), grads = grad_fn(img, t, c, c_mask)
optimizer(inputs_gradient)
return loss
def train(model, dataset):
size = dataset.get_dataset_size()
model.set_train()
for batch, (img, label) in enumerate(dataset.create_tuple_iterator()):
img = Tensor(img, dtype=mindspore.float32)
img = ops.cast(img, ms.float32)
t_np = np.random.randint(0, 300, size=(img.shape[0],), dtype=np.int32)
t = Tensor(t_np, dtype=mindspore.float32)
c = ops.gather(text_features, label, 0) # 匹配文本向量
c, c_mask = get_context_mask(c, 0.1)
loss = train_step(img, t, c, c_mask)
if batch % 100 == 0:
loss, current = loss.asnumpy(), batch
print(f"loss: {loss:>7f} [{current:>3d}/{size:>3d}]")
在执行训练代码的时候报错
epochs = 1
for t in range(epochs):
print(f"Epoch {t+1}\n-------------------------------")
ms.jit(train(model, train_dataset1))
print("Done!")
报错如下：
[ERROR] KERNEL(22770,ffffa80f80b0,python):2025-09-29-17:13:16.939.851 [mindspore/ccsrc/kernel/kernel_info.cc:72] GetMutableOutputAddr] Index [0] out of range
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
Cell In[54], line 6
4 for t in range(epochs):
5 print(f"Epoch {t+1}\n-------------------------------")
----> 6 ms.jit(train(model, train_dataset1))
7 print("Done!")
Cell In[52], line 45, in train(model, dataset)
42 c = ops.gather(text_features, label, 0) # 匹配文本向量
43 c, c_mask = get_context_mask(c, 0.1)
---> 45 loss = train_step(img, t, c, c_mask)
47 if batch % 100 == 0:
48 loss, current = loss.asnumpy(), batch
Cell In[52], line 30, in train_step(img, t, c, c_mask)
28 output, inputs_gradient = grad_fn(img, t, c, c_mask)
29 # (loss, _), grads = grad_fn(img, t, c, c_mask)
---> 30 optimizer(inputs_gradient)
31 return loss
File ~/anaconda3/envs/MindSpore/lib/python3.10/site-packages/mindspore/nn/cell.py:1270, in Cell.__call__(self, *args, **kwargs)
1267 if not (self.requires_grad or self._dynamic_shape_inputs or self.mixed_precision_type):
1268 if not (self._forward_pre_hook or self._forward_hook or self._backward_pre_hook or self._backward_hook or
1269 self._shard_fn or self._recompute_cell or (self.has_bprop and _pynative_executor.requires_grad())):
-> 1270 return self.construct(*args, **kwargs)
1272 return self._run_construct(*args, **kwargs)
1274 return self._complex_call(*args, **kwargs)
File ~/anaconda3/envs/MindSpore/lib/python3.10/site-packages/mindspore/common/api.py:1180, in jit.<locals>.wrap_func.<locals>.staging_specialize(*args, **kwargs)
1177 setattr(func, "amp_strategy", get_curr_amp_strategy())
1179 ms_function_executor = _JitExecutor(func, hash_obj, None, process_obj, jit_config, dynamic)
-> 1180 out = ms_function_executor(*args, **kwargs)
1181 return out
File ~/anaconda3/envs/MindSpore/lib/python3.10/site-packages/mindspore/common/api.py:180, in _wrap_func.<locals>.wrapper(*arg, **kwargs)
178 @wraps(fn)
179 def wrapper(*arg, **kwargs):
--> 180 results = fn(*arg, **kwargs)
181 return _convert_python_data(results)
File ~/anaconda3/envs/MindSpore/lib/python3.10/site-packages/mindspore/common/api.py:627, in _JitExecutor.__call__(self, *args, **kwargs)
625 except Exception as err:
626 _pynative_executor.clear_res()
--> 627 raise err
629 if context.get_context("precompile_only") or os.getenv('MS_DEV_PRECOMPILE_ONLY') == '1':
630 return None
File ~/anaconda3/envs/MindSpore/lib/python3.10/site-packages/mindspore/common/api.py:621, in _JitExecutor.__call__(self, *args, **kwargs)
619 if context.get_context("mode") == context.PYNATIVE_MODE:
620 _pynative_executor.set_jit_compile_status(True, phase)
--> 621 phase = self.compile(self.fn.__name__, *args_list, **kwargs)
622 _pynative_executor.set_jit_compile_status(False, phase)
623 else:
File ~/anaconda3/envs/MindSpore/lib/python3.10/site-packages/mindspore/common/api.py:738, in _JitExecutor.compile(self, method_name, *args, **kwargs)
736 if isinstance(self.obj, ms.nn.Cell):
737 self._graph_executor.set_weights_values(self.obj.parameters_dict())
--> 738 is_compile = self._graph_executor.compile(self.obj, compile_args, kwargs, phase)
740 if not is_compile:
741 raise RuntimeError("Executor compile failed.")
RuntimeError: Output_idx 0 of node @kernel_graph5:param_global_step node:0xaaab447b3220 output addr is not exist.
----------------------------------------------------
- C++ Call Stack: (For framework developers)
----------------------------------------------------
mindspore/ccsrc/backend/common/session/anf_runtime_algorithm.cc:909 GetMutableOutputAddr
感觉是在构建计算图的时候找不到输出接点，但是没有定位到具体哪里出了问题。
使用的modelarts notebook实例规格为：
mindspore_2.6.0rc1-cann_8.1.rc1-py_3.10-euler_2.10.11-aarch64-snt9b