Add attention export functionality to experiments (#134)

* Add attention export functionality

* Update experiments/attention_export.py

Co-authored-by: Anush <[email protected]>

---------

Co-authored-by: Anush <[email protected]>

experiments/attention_export.py

@@ -0,0 +1,18 @@
+import numpy as np
+import onnx
+import onnxruntime
+from optimum.exporters.onnx import main_export
+from transformers import AutoTokenizer
+
+model_id = "sentence-transformers/paraphrase-MiniLM-L6-v2"
+output_dir = f"models/{model_id.replace('/', '_')}"
+model_kwargs = {"output_attentions": True, "return_dict": True}
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+# export if the output model does not exist
+# try:
+#     sess = onnxruntime.InferenceSession(f"{output_dir}/model.onnx")
+#     print("Model already exported")
+# except FileNotFoundError:
+print(f"Exporting model to {output_dir}")
+main_export(model_id, output=output_dir, no_post_process=True, model_kwargs=model_kwargs)

experiments/try_attention_export.py

@@ -0,0 +1,30 @@
+import numpy as np
+import onnx
+import onnxruntime
+from optimum.exporters.onnx import main_export
+from transformers import AutoTokenizer
+
+model_id = "sentence-transformers/paraphrase-MiniLM-L6-v2"
+output_dir = f"models/{model_id.replace('/', '_')}"
+model_kwargs = {"output_attentions": True, "return_dict": True}
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+model_path = f"{output_dir}/model.onnx"
+onnx_model = onnx.load(model_path)
+ort_session = onnxruntime.InferenceSession(model_path)
+text = "This is a test sentence"
+tokenizer_output = tokenizer(text, return_tensors="np")
+input_ids = tokenizer_output["input_ids"]
+attention_mask = tokenizer_output["attention_mask"]
+print(attention_mask)
+# Prepare the input
+input_ids = np.array(input_ids).astype(np.int64)  # Replace your_input_ids with actual input data
+
+# Run the ONNX model
+outputs = ort_session.run(None, {"input_ids": input_ids, "attention_mask": attention_mask})
+
+# Get the attention weights
+attentions = outputs[-1]
+
+# Print the attention weights for the first layer and first head
+print(attentions[0][0])