How to define example_input_array for computational graph

[minwang-ai]

Could you please give me an example for defining self.example_input_array according to the document?

When and where should I log computational graph, train/Val/test step?

log_graph (bool) – Adds the computational graph to tensorboard. This requires that the user has defined the self.example_input_array attribute in their model.

[rohitgr7]

you can simply assign it to a variable:

class LitModel(LightningModule):
    def __init__(self, ...):
        ...
        self.example_input_array = ...

[minwang-ai]

Should I specify log in test_step or train_step?

def test_step(self, batch, batch_idx):
...
self.logger.log_graph(self)
...

[minwang-ai]

Do you mean define a random tensor with same size as input?
What if I pass a dictionary to my model as input and use different values for CNN input and pointNet input and fusing two network’s feature map together?

[rohitgr7]

Do you mean define a random tensor with same size as input?

yes

Apart from that, I only need to set the log_graph flag to True, right?

yes

Should I specify log in test_step or train_step?

no need, it will happen automatically for you.

What if I pass a dictionary to my model as input and use different values for CNN input and pointNet input and fusing two network’s feature map together?

I think it depends on your forward method. If it can handle dict input and it should work I guess.

[minwang-ai]

Thank you for your reply!

Do you mean define a random tensor with same size as input?

yes

Apart from that, I only need to set the log_graph flag to True, right?

yes

Should I specify log in test_step or train_step?

no need, it will happen automatically for you.

What if I pass a dictionary to my model as input and use different values for CNN input and pointNet input and fusing two network’s feature map together?

I think it depends on your forward method. If it can handle dict input and it should work I guess.

My forward method can handle dict input but I don’t know how to define self.example_input_array in this case, since my input is a nested dictionary, each value is also a dictionary and each value is a tensor.

[rohitgr7]

class LitModel(LightningModule):
    def __init__(self, ...):
        ...
        self.example_input_array = {'key1': {'nested_key1': torch.Tensor(...)}, 'key2': ...}

this doesn't work?

[sascharo]

```python
class LitModel(LightningModule):
    def __init__(self, ...):
        ...
        self.example_input_array = {'key1': {'nested_key1': torch.Tensor(...)}, 'key2': ...}

this doesn't work?

Not for me.