LSTM

Description

Computes a one-layer LSTM.

Input

• x: input tensor of type float16.
• w: input tensor of type float16.
• r: input tensor of type float16.
• b (optional): input tensor of type float16.
• sequence_lens (optional): input tensor of type int32.
• initial_h (optional): input tensor of type float16.
• initial_c (optional): input tensor of type float16.
• p (optional): input tensor of type float16.

Attribute

• direction: string. The default value is forward.
• hidden_size: int.
• activation_alpha (optional): float, indicating the scaling value used by some activation functions.
• activation_beta (optional): float, indicating the scaling value used by some activation functions.
• clip: float, indicating that elements of the tensor are clipped to the range of [-clip, +clip] and applied to the activated input. By default, no clipping is performed.
• input_forget: int32. The default value is 0. If the value is 1, the input and forget gate are coupled.

Output

• y: output tensor of type float16.
• y_h: output tensor of type float16.
• y_c: output tensor of type float16.

Constraints

None.

ONNX Opset Support

Opset v8/v9/v10/v11/v12/v13/v14/v15/v16