jittor.nn.ConstantPad2d报错local variable 'i' referenced before assignment

[PhyllisJi]

Describe the bug

可能是上一层的shape为[1,128,]导致，但是该API没有对输入形状进行检查并合理地抛出异常，另外jittor不支持ConstantPad1d

Full Log

---------------------------------------------------------------------------
UnboundLocalError                         Traceback (most recent call last)
Cell In[6], line 55
     51     y = m(x)
     52     return list(y.shape)
---> 55 go()

Cell In[6], line 51, in go()
     49 x = jittor.randn([1, 3, 224, 224])
     50 m = alexnet()
---> 51 y = m(x)
     52 return list(y.shape)

File ~/miniconda3/envs/myconda/lib/python3.9/site-packages/jittor/__init__.py:1168, in Module.__call__(self, *args, **kw)
   1167 def __call__(self, *args, **kw):
-> 1168     return self.execute(*args, **kw)

Cell In[6], line 41, in alexnet.execute(self, x)
     39 x = self.relu5_mutated(x)
     40 print(x.shape)
---> 41 x = self.pool3_mutated(x)
     42 return x

File ~/miniconda3/envs/myconda/lib/python3.9/site-packages/jittor/__init__.py:1168, in Module.__call__(self, *args, **kw)
   1167 def __call__(self, *args, **kw):
-> 1168     return self.execute(*args, **kw)

File ~/miniconda3/envs/myconda/lib/python3.9/site-packages/jittor/nn.py:1694, in ConstantPad2d.execute(self, x)
   1692 for i in range(len(shape)-2):
   1693     tar_dims.append(f"i{i}")
-> 1694 tar_dims.append(f"i{i+1}-{self.pt}")
   1695 tar_dims.append(f"i{i+2}-{self.pl}")
   1696 return x.reindex(tar_shape, tar_dims, overflow_value=self.value)

UnboundLocalError: local variable 'i' referenced before assignment

Minimal Reproduce

import os
os.environ["disable_lock"] = "1"
import jittor
import jittor.nn as nn
import jittor.optim as optim
import numpy as np
import copy


class alexnet(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1_mutated = jittor.nn.Conv2d(in_channels=3, out_channels=64, kernel_size=6, stride=4, padding=2)
        self.relu1_mutated = jittor.nn.ReLU()
        self.pool1_mutated = jittor.nn.ZeroPad2d(padding=(1, 5, 7, 1))
        self.conv2_mutated = jittor.nn.UpsamplingNearest2d(scale_factor=1.0)
        self.relu2_mutated = jittor.nn.GELU()
        self.pool2_mutated = jittor.nn.MaxPool2d(kernel_size=(3, 6), stride=8, ceil_mode=False, return_indices=False, padding=(6, 3))
        self.conv3_mutated = jittor.nn.Sigmoid()
        self.relu3_mutated = jittor.nn.ELU()
        self.conv4_mutated = jittor.nn.MaxPool2d(kernel_size=(3, 7), return_indices=False, stride=(6, 2))
        self.relu4_mutated = jittor.nn.ReLU()
        self.conv5_mutated = jittor.nn.Flatten()
        self.relu5_mutated = jittor.nn.Softmax()
        self.pool3_mutated = jittor.nn.ConstantPad2d(padding=5, value=0.0)
    
    def execute(self, x):
        x = self.conv1_mutated(x)
        x = self.relu1_mutated(x)
        x = self.pool1_mutated(x)
        x = self.conv2_mutated(x)
        x = self.relu2_mutated(x)
        x = self.pool2_mutated(x)
        x = self.conv3_mutated(x)
        x = self.relu3_mutated(x)
        x = self.conv4_mutated(x)
        x = self.relu4_mutated(x)
        x = self.conv5_mutated(x)
        x = self.relu5_mutated(x)
        print(x.shape)
        x = self.pool3_mutated(x)
        return x

def go():
    jittor.flags.use_cuda = 1
    x = jittor.randn([1, 3, 224, 224])
    m = alexnet()
    y = m(x)
    return list(y.shape)

go()