The problem about using RECORDER in dev1.x (RecorderManager)

[cape-zck]

Thanks to the MMrazor team contribution, Now I am learning the use of MMrazor dev1.x about the RECORDER moudle, but I ran into a couple of problems.

1.I find no mmrazor.core in dev1.x. however,I find the module such as MethodOutputsRecorder, RecorderManager and so on in mmrazor.models.task_modules, this problem seems to have been solved.

2.But when I test the demo about RecorderManager, I found that if I tried to use ModuleOutputs and MethodOutputs in RecorderManager at the same time, the forward process would seem to run twice, and output MethodOutputs would report an out-of-range error. But if the Method(toy_func) isn’t in the same python file, this error doesn't seem to happen.

For example:

import torch
import random
from mmengine import ConfigDict
from torch import nn
from mmrazor.models.task_modules import RecorderManager

class Toy():
    def toy_func(self):
        return random.randint(0, 1000000)

class ToyModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(1, 1, 1)
        self.conv2 = nn.Conv2d(1, 1, 1)
        self.toy = Toy()

    def forward(self, x):
        return self.conv2(self.conv1(x)) + self.toy.toy_func()

# configure multi-recorders
conv1_rec = ConfigDict(type='ModuleOutputs', source='conv1')
conv2_rec = ConfigDict(type='ModuleOutputs', source='conv2')
func_rec = ConfigDict(type='MethodOutputs', source='demo.Toy.toy_func')
# instantiate RecorderManager with a dict that contains recorders' configs,
# you can customize their keys.
manager = RecorderManager(
    {'conv1_rec': conv1_rec,
     'conv2_rec': conv2_rec,
     'func_rec': func_rec})

model = ToyModel()
# initialize is to make specified module can be recorded by
# registering customized forward hook.
manager.initialize(model)

x = torch.rand(1, 1, 1, 1)
with manager:
    out = model(x)

conv2_out = manager.get_recorder('conv2_rec').get_record_data()
'''
 - mmengine - WARNING - The "task util" registry in mmrazor did not set import location. Fallback to call `mmrazor.utils.register_all_modules` instead.
tensor([[[[-0.2194]]]], grad_fn=<ThnnConv2DBackward>)
tensor([[[[-1.8282]]]], grad_fn=<ThnnConv2DBackward>)
'''

func_rec = manager.get_recorder('func_rec').get_record_data()
'''
569682
Traceback (most recent call last):
  File "demo.py", line 50, in <module>
    func_rec = manager.get_recorder('func_rec').get_record_data()
  File "mmrazor\mmrazor\models\task_modules\recorder\base_recorder.py", line 78, in get_record_data
    assert record_idx < len(self._data_buffer), \
AssertionError: record_idx is illegal. The length of data_buffer is 0, but record_idx is 0.
'''

But when I remove the MethodOutputs, everything seemed normal again.

import torch
import random
from mmengine import ConfigDict
from torch import nn
from mmrazor.models.task_modules import RecorderManager

class Toy():
    def toy_func(self):
        return random.randint(0, 1000000)

class ToyModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(1, 1, 1)
        self.conv2 = nn.Conv2d(1, 1, 1)
        self.toy = Toy()

    def forward(self, x):
        return self.conv2(self.conv1(x)) + self.toy.toy_func()

# configure multi-recorders
conv1_rec = ConfigDict(type='ModuleOutputs', source='conv1')
conv2_rec = ConfigDict(type='ModuleOutputs', source='conv2')
#func_rec = ConfigDict(type='MethodOutputs', source='demo.Toy.toy_func')
# instantiate RecorderManager with a dict that contains recorders' configs,
# you can customize their keys.
manager = RecorderManager(
    {'conv1_rec': conv1_rec,
     'conv2_rec': conv2_rec})

model = ToyModel()
# initialize is to make specified module can be recorded by
# registering customized forward hook.
manager.initialize(model)

x = torch.rand(1, 1, 1, 1)
with manager:
    out = model(x)

conv2_out = manager.get_recorder('conv2_rec').get_record_data()

'''
- mmengine - WARNING - The "task util" registry in mmrazor did not set import location. Fallback to call `mmrazor.utils.register_all_modules` instead.
tensor([[[[-0.7945]]]], grad_fn=<ThnnConv2DBackward>)

'''

[HIT-cwh]

Hi @cape-zck ! Thank you for your issue.
If we set the source module name to demo.Toy.toy_func, the method output recorder will first import this target module demo.Toy by import_modules_from_strings('demo.Toy') and then wrap its toy_func method by the method_record_wrapper method in MethodOutputsRecorder. In this way, we can record the output data of a specific class method.

While the Toy.toy_func method you used during the forward pass is not the same as demo.Toy.toy_func. So the method demo.Toy.toy_func is not executed actually and the output data of demo.Toy.toy_func is not recorded by the MethodOutputsRecorder. Hence it would report an out-of-range error.

[pkyzh2006]

Hi @cape-zck ! Thank you for your issue. If we set the source module name to demo.Toy.toy_func, the method output recorder will first import this target module demo.Toy by import_modules_from_strings('demo.Toy') and then wrap its toy_func method by the method_record_wrapper method in MethodOutputsRecorder. In this way, we can record the output data of a specific class method.

While the Toy.toy_func method you used during the forward pass is not the same as demo.Toy.toy_func. So the method demo.Toy.toy_func is not executed actually and the output data of demo.Toy.toy_func is not recorded by the MethodOutputsRecorder. Hence it would report an out-of-range error.

Could you help me check #636