q-ycong-p

Keras LSTM not converted correctly when precedent Embedding layer specifies mask_zero=True

17

**Describe the bug** When an `tf.keras.layers.Embedding` with attribute `mask_zero=True` attribute precede LSTM layer, the LSTM is **converted into loops instead of LSTM op**. **System information** - OS Platform and Distribution...

Add masked LSTM support

3

Masking is a intra-layer behavior in TF LSTM [1] but is not a intra-op behavior in ONNX LSTM [2]. When converted to ONNX, masked TF LSTM layer is converted to...

DEBT: Enforce thread-safety for ONNX opset schema API

5

### Description Enforce thread-safe access to the [static schema map](https://github.com/onnx/onnx/blob/main/onnx/defs/schema.cc#L1052) in the implementations that register and de-register schema. ### Motivation and Context This is a discussed feature following up to...

Missing type STRINGS when parsing attributes

1

The [current implementation](https://github.com/ToriML/onnx2pytorch/blob/master/onnx2pytorch/convert/attribute.py#L30) doesn't support STRINGS as attribute type. Per [ONNX sepc for LSTM](https://github.com/onnx/onnx/blob/main/docs/Operators.md#LSTM), one of the attributes can be `activations` which is of type STRINGS. This leads to conversion...