ctc_loss.py

import hipdnn
import torch


def build_ctc_loss_graph(hipdnn_handle, torch_tensor_probs, hipdnn_data_type):
    graph = hipdnn.pygraph(
        handle=hipdnn_handle,
        io_data_type=hipdnn_data_type,
        intermediate_data_type=hipdnn.data_type.FLOAT,
        compute_data_type=hipdnn.data_type.FLOAT,
        name="ctc_loss_inference",
    )
    hipdnn_tensor_probs = graph.tensor_like(torch_tensor_probs)
    losses, gradients = graph.ctc_loss(
        probs=hipdnn_tensor_probs,
        blank_label_id=0,
        apply_softmax=False,
        algo=0,
        labels=[1, 2, 3, 4, 2, 3, 2],
        label_lengths=[1, 2, 1, 3],
        input_lengths=[4, 100, 100, 200],
        name="ctc_loss",
    )
    losses.set_output(True)
    gradients.set_output(True)
    graph.build(hipdnn_handle)
    return (graph, hipdnn_tensor_probs, losses, gradients)


if __name__ == "__main__":
    batch, max_time, num_classes = 4, 500, 5
    hipdnn_data_type = hipdnn.data_type.FLOAT
    torch_data_type = torch.float32
    torch_tensor_probs = torch.rand(
        max_time, batch, num_classes, dtype=torch_data_type, device="cuda"
    )
    hipdnn_handle = hipdnn.create_handle()
    graph, hipdnn_tensor_probs, hipdnn_tensor_losses, hipdnn_tensor_gradients = (
        build_ctc_loss_graph(hipdnn_handle, torch_tensor_probs, hipdnn_data_type)
    )
    torch_tensor_losses = torch.empty(batch, dtype=torch_data_type, device="cuda")
    torch_tensor_gradients = torch.empty(
        batch, max_time, num_classes, dtype=torch_data_type, device="cuda"
    )
    variant_pack = {
        hipdnn_tensor_probs: torch_tensor_probs.data_ptr(),
        hipdnn_tensor_losses: torch_tensor_losses.data_ptr(),
        hipdnn_tensor_gradients: torch_tensor_gradients.data_ptr(),
    }
    workspace = torch.empty(graph.get_workspace_size(), dtype=torch.uint8, device="cuda")
    graph.exec(variant_pack=variant_pack, workspace=workspace.data_ptr())
    print("ctc_loss graph execution complete.")