Hi Tsvi,
I understand that you want to retrieve layer activations from a trained “bertDocumentClassifier” model.
There are two reasons why the “activations” function is not working as expected.
First, the “InputNames” property of the underlying network for the model shows that the model accepts three inputs instead of two. Namely, it expects the input IDs, an attention mask, and the segment IDs.
In your code, you are calling the function with only two inputs, the input IDs and the segment IDs.
Second, the “activations” function only works with networks represented as “DAGNetwork” objects or “SeriesNetwork” objects, as specified in the documentation: https://www.mathworks.com/help/releases/R2023b/deeplearning/ref/seriesnetwork.activations.html#d126e5157.
The underlying network for “bertDocumentClassifier” is a “dlnetwork” object: https://www.mathworks.com/help/releases/R2023b/textanalytics/ref/bertdocumentclassifier.html#mw_2480ef12-2a75-480d-aec3-eefb236d8afe. For such objects, you need to use the “predict” or the “forward” function, based on whether you want the model to output for inference or for training.
Please try the following code. I am assuming you want the model to output for inference and hence, using the “predict” function. If you want the model to output for training, change the “predict” call to a “forward” call. No other changes will be required:
% Extract the network
net = mdl.Network;
% Extract an example and encode it
example = textDataTrain(1, :);
[tokens, segments] = encode(mdl.Tokenizer, example);
% Since tokens and segments is a cell arrays with single vectors
% Extract the vectors
tokens = tokens{1};
segments = segments{1};
% Extract number of tokens
dims = size(tokens, 2);
% Convert the tokens and segments to dlarray
% BERT expects input in CTB format
tokens = dlarray(tokens, "CTB");
segments = dlarray(segments, "CTB");
% Create an attention mask of all zeros in CTB format
attentionMask = dlarray(zeros(1, dims), "CTB");
% Use predict function to get the output of layer 'out_fc2'
output = predict(net, tokens, attentionMask, segments, 'Outputs', 'out_fc2');
The code:
- Extracts the underlying network for the model.
- Extracts a single example from the training data and encodes it into input IDs and segment IDs.
- Since the outputs of the “encode” function are cell arrays (https://www.mathworks.com/help/releases/R2023b/textanalytics/ref/berttokenizer.encode.html#mw_3665f38f-7b0f-4514-a774-9578f8f519b4), extracts the vectors in those cell arrays.
- Since the ”forward” function expect inputs as “dlarray” objects (https://www.mathworks.com/help/releases/R2023b/deeplearning/ref/dlnetwork.forward.html#function_forward_sep_mw_bfae26e4-e62e-4b46-80e2-3d3f8305b520) and the “predict” function can work with them as well, converts the input IDs and the segment IDs into “dlarray” objects. Note that the “bertDocumentClassifier” expects input in the “CTB” data format.
- Creates an attention mask of all zeros as a “dlarray” object, in the “CTB” data format.
- Uses the “predict” function with the “Outputs” name-value argument to get the activations from the layer ‘out_fc2’.
This has the following output:
Please refer to the following resources for more information:
- “activations” function – https://www.mathworks.com/help/releases/R2023b/deeplearning/ref/seriesnetwork.activations.html.
- “encode” function - https://www.mathworks.com/help/releases/R2023b/textanalytics/ref/berttokenizer.encode.html.
- “predict” function – https://www.mathworks.com/help/releases/R2023b/deeplearning/ref/seriesnetwork.predict.html.
- “forward” function - https://www.mathworks.com/help/releases/R2023b/deeplearning/ref/dlnetwork.forward.html.
- “dlarray” objects - https://www.mathworks.com/help/releases/R2023b/deeplearning/ref/dlarray.html.
Hope this helps!
