[Bug]Got "ValueError: not enough values to unpack (expected 2, got 1)" when BatchMatMul using two same tensors

[NoobCheese]

I got a tf graph , imported it into TVM succesfully using api "relay.frontend.from_tensorflow"
Error occurs when :

Error log like this :

The two input tensors of "batchMatMul" node are the same .both of their shapes are [1, 41, 16].

And i traces the soure code here:

i am deeply confused

[vinx13]

Seems that only one placeholder is created. Would you mind sharing the Relay IR?

[NoobCheese]

Seems that only one placeholder is created. Would you mind sharing the Relay IR?

i am sorry that the Relay IR which is generated from a online business model. But it seems u r right. i modified the sample batchmatmul in tvm and i got the same error hints.

The original smaple code is:
x = placeholder(xxxx),
y = placeholder(xxxx),
And what i did is :
x = placeholder(xxxx),
y = x,

However, my model works in online TF env, and what i should do if this case happened when i want to use TVM?

thanks bro!

[vinx13]

You can print the prim func here and check if the placeholders are correctly created. https://github.com/apache/tvm/blob/main/src/relay/backend/te_compiler_cache.cc#L125

…

On Sep 4, 2021, at 9:48 AM, Noob ***@***.***> wrote:  Seems that only one placeholder is created. Would you mind sharing the Relay IR? i am sorry that the Relay IR which is generated from a online business model. But it seems u r right. i modified the sample batchmatmul in tvm and i got the same error hints. The original smaple code is: x = placeholder(xxxx), y = placeholder(xxxx), And what i did is : x = placeholder(xxxx), y = x, However, my model works in online TF env, and what i should do if this case happened when i want to use TVM? thanks bro! — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[comaniac]

I don't think we can reuse the placeholder in this way. If your model results in this IR, then it is probably due to the incorrect Relay TF frontend, or unsupported behavior.

[NoobCheese]

You can print the prim func here and check if the placeholders are correctly created. https://github.com/apache/tvm/blob/main/src/relay/backend/te_compiler_cache.cc#L125
…
On Sep 4, 2021, at 9:48 AM, Noob @.***> wrote:  Seems that only one placeholder is created. Would you mind sharing the Relay IR? i am sorry that the Relay IR which is generated from a online business model. But it seems u r right. i modified the sample batchmatmul in tvm and i got the same error hints. The original smaple code is: x = placeholder(xxxx), y = placeholder(xxxx), And what i did is : x = placeholder(xxxx), y = x, However, my model works in online TF env, and what i should do if this case happened when i want to use TVM? thanks bro! — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

My co-worker hacked the tf subgraph, and added a placeholder into the bmm node this morning, Tvm built the Relay IR well :)
Anyway. one thing i am gonna comfirm, if only one placeholder in bmm, i.e a matrix multiply itself, does tvm support this?

Thanks bro!

[vinx13]

I checked the below example, seems Relay currently doesn't support this.

import tvm
from tvm import relay
a = relay.Var('a', relay.TensorType((1,8, 16)))
b = relay.nn.batch_matmul(a, a)
f = relay.Function([a], b)
with tvm.target.Target('llvm'):
    relay.build(f)

[NoobCheese]

I checked the below example, seems Relay currently doesn't support this.

import tvm
from tvm import relay
a = relay.Var('a', relay.TensorType((1,8, 16)))
b = relay.nn.batch_matmul(a, a)
f = relay.Function([a], b)
with tvm.target.Target('llvm'):
    relay.build(f)

thx, it really helps