MCPcopy Create free account
hub / github.com/NVIDIA/TensorRT / squad_output

Function squad_output

demo/BERT/builder_varseqlen.py:333–353  ·  view source on GitHub ↗

Create the squad output

(prefix, config, init_dict, network, input_tensor)

Source from the content-addressed store, hash-verified

331
332
333def squad_output(prefix, config, init_dict, network, input_tensor):
334 """
335 Create the squad output
336 """
337 hidden_size = config.hidden_size
338
339 W_out = init_dict[prefix + SQD_W]
340 B_out = init_dict[prefix + SQD_B]
341
342 if config.use_int8:
343 dense = network.add_convolution_nd(input_tensor, 2, (1, 1), W_out, B_out)
344 else:
345 dense = network.add_fully_connected(input_tensor, 2, W_out, B_out)
346
347 OUT = network.add_shuffle(dense.get_output(0))
348 if config.use_int8 and config.interleaved:
349 OUT.second_transpose = (1, 2, 0, 3)
350 else:
351 OUT.second_transpose = (1, 0, 2, 3)
352 set_output_name(OUT, prefix, "squad_logits")
353 return OUT
354
355def emb_layernorm(builder, network, config, weights_dict, builder_config, max_sequence_length, batch_sizes):
356 input_ids = network.add_input(name="input_ids", dtype=trt.int32, shape=(-1,))

Callers 1

bert_modelFunction · 0.70

Calls 1

set_output_nameFunction · 0.70

Tested by

no test coverage detected