C3_W4 UNQ_C5 : problem with loading the weights

[Florent_Tho]

Hello,

When I try to compute the accuracy with the Siamese model, I get an error that I don’t manage to solve (see trace below)

I suppose there is a discrepancy between the model I build and the shape of the weights in the pretrained model.
I tried to recharge the workbook several time, to no avail.

How can I investigate this error and find the problem ?
I didn’t find much usefull on the web

LayerError: Exception passing through layer Parallel (in pure_fn):
  layer created in file [...]/<ipython-input-20-54e6716dd7ce>, line 29
  layer input shapes: ShapeDtype{shape:(512, 64), dtype:int64}

  File [...]/trax/layers/base.py, line 707, in __setattr__
    super().__setattr__(attr, value)

  File [...]/trax/layers/base.py, line 454, in weights
    f'Number of weight elements ({len(weights)}) does not equal the

ValueError: Number of weight elements (512) does not equal the number of sublayers (2) in: Parallel_in2_out2[
  Serial[


    Embedding_41699_128
    LSTM_128

    Mean
    Normalize

  ]
  Serial[

    Embedding_41699_128
    LSTM_128

    Mean
    Normalize

  ]
]

[Florent_Tho]

I’m really puzzled.
it seems that defining the q1, q2 creates difficulties.

for example :

[reinoudbosch]

Hi Florent_Tho,

The comment above the invocation of the data generator states

“# use batch size chuncks of questions as Q1 & Q2 arguments of the data generator. e.g x[i:i + batch_size]”

Are you doing this in your code? The example you give [0:512] with a batch size of 10 does not conform to this comment.

[zzhu24]

I have run into the exact same issue. I think I have correctly used the “# use batch size chuncks of questions as Q1 & Q2 arguments of the data generator. e.g x[i:i + batch_size]”

Here is the data generator from my code:
q1, q2 = next(data_generator(test_Q1[i:i + batch_size], test_Q2[i:i + batch_size],pad =vocab[‘’], batch_size=batch_size,shuffle=False))

then when I call model(q1,q2), the same issue happens again.

my lab id is nhbjjhab

[ARNAV_GUPTA1]

Hint: use vocab['<PAD>'] for the pad argument of the data generator . This was the hint given to solve the question. You are using pad = vocab[‘’]. I guess using this hint you will be able to solve the question.

[zzhu24]

Thanks for the reply, but I was using vocab < PAD >. Sorry when I copy and paste it did not come over correctly.

This screenshot below is a test to replicate the issue I have and ran into the exact same issue.
I have this issue where I call next() on the generator to get q1 and q2 inputs, but when I call the model with q1 and q2 it fails.

[anchyzas]

Hey, I just solved the exact same problem. Try to pass them to the model as a tuple, with two pairs of brackets, like so:

model((temp1, temp2))

[glicerico]

It’s a shame there’s no single example how to run the trax’s Parallel model for inference. But @anchyzas gave the right answer, add both question lists as a tuple (or list), as Parallel takes both inputs as the first argument.

[joaopedrovtp]

yea, that solved my problem too! thanks!

[leggard]

@anchyzas

the model as a tuple,

Thanks, this helped me too!

[O1ena]

Thank you!!