Skip to content

constant batch in TensorRT engine build #483

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
1 commit merged into from May 6, 2020
+30 −45
Merged

constant batch in TensorRT engine build #483

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
39 changes: 12 additions & 27 deletions PyTorch/SpeechSynthesis/Tacotron2/exports/export_tacotron2_onnx.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -316,11 +316,9 @@ def main():
do_constant_folding=True,
input_names=["sequences", "sequence_lengths"],
output_names=["memory", "processed_memory", "lens"],
dynamic_axes={"sequences": {0: "batch_size", 1: "text_seq"},
"sequence_lengths": {0: "batch_size"},
"memory": {0: "batch_size", 1: "mem_seq"},
"processed_memory": {0: "batch_size", 1: "mem_seq"},
"lens": {0: "batch_size"},
dynamic_axes={"sequences": {1: "text_seq"},
"memory": {1: "mem_seq"},
"processed_memory": {1: "mem_seq"}
})

decoder_iter = DecoderIter(tacotron2)
Expand Down Expand Up @@ -379,26 +377,13 @@ def main():
"out_attention_weights",
"out_attention_weights_cum",
"out_attention_context"],
dynamic_axes={"decoder_input" : {0: "batch_size"},
"attention_hidden" : {0: "batch_size"},
"attention_cell" : {0: "batch_size"},
"decoder_hidden" : {0: "batch_size"},
"decoder_cell" : {0: "batch_size"},
"attention_weights" : {0: "batch_size", 1: "seq_len"},
"attention_weights_cum" : {0: "batch_size", 1: "seq_len"},
"attention_context" : {0: "batch_size"},
"memory" : {0: "batch_size", 1: "seq_len"},
"processed_memory" : {0: "batch_size", 1: "seq_len"},
"mask" : {0: "batch_size", 1: "seq_len"},
"decoder_output" : {0: "batch_size"},
"gate_prediction" : {0: "batch_size"},
"out_attention_hidden" : {0: "batch_size"},
"out_attention_cell" : {0: "batch_size"},
"out_decoder_hidden" : {0: "batch_size"},
"out_decoder_cell" : {0: "batch_size"},
"out_attention_weights" : {0: "batch_size", 1: "seq_len"},
"out_attention_weights_cum" : {0: "batch_size", 1: "seq_len"},
"out_attention_context" : {0: "batch_size"}
dynamic_axes={"attention_weights" : {1: "seq_len"},
"attention_weights_cum" : {1: "seq_len"},
"memory" : {1: "seq_len"},
"processed_memory" : {1: "seq_len"},
"mask" : {1: "seq_len"},
"out_attention_weights" : {1: "seq_len"},
"out_attention_weights_cum" : {1: "seq_len"}
})

postnet = Postnet(tacotron2)
Expand All @@ -410,8 +395,8 @@ def main():
do_constant_folding=True,
input_names=["mel_outputs"],
output_names=["mel_outputs_postnet"],
dynamic_axes={"mel_outputs": {0: "batch_size", 2: "mel_seq"},
"mel_outputs_postnet": {0: "batch_size", 2: "mel_seq"}})
dynamic_axes={"mel_outputs": {2: "mel_seq"},
"mel_outputs_postnet": {2: "mel_seq"}})

mel = test_inference(encoder, decoder_iter, postnet)
torch.save(mel, "mel.pt")
Expand Down
6 changes: 3 additions & 3 deletions PyTorch/SpeechSynthesis/Tacotron2/exports/export_waveglow_onnx.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,9 +198,9 @@ def export_onnx(parser, args):
do_constant_folding=True,
input_names=["mel", "z"],
output_names=["audio"],
dynamic_axes={"mel": {0: "batch_size", 2: "mel_seq"},
"z": {0: "batch_size", 2: "z_seq"},
"audio": {0: "batch_size", 1: "audio_seq"}})
dynamic_axes={"mel": {2: "mel_seq"},
"z": {2: "z_seq"},
"audio": {1: "audio_seq"}})


def main():
Expand Down
2 changes: 1 addition & 1 deletion PyTorch/SpeechSynthesis/Tacotron2/trt/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,5 +85,5 @@ Our results were obtained by running the `./trt/run_latency_tests_trt.sh` script

|Framework|Batch size|Input length|Precision|Avg latency (s)|Latency std (s)|Latency confidence interval 90% (s)|Latency confidence interval 95% (s)|Latency confidence interval 99% (s)|Throughput (samples/sec)|Speed-up PyT+TRT/TRT|Avg mels generated (81 mels=1 sec of speech)|Avg audio length (s)|Avg RTF|
|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
|PyT+TRT|1| 128| FP16| 1.13| 0.01| 1.15| 1.15| 1.16| 137,738| 1.45| 608| 7.06| 6.24|
|PyT+TRT|1| 128| FP16| 1.02| 0.05| 1.09| 1.10| 1.14| 150,439| 1.59| 602| 6.99| 6.86|
|PyT |1| 128| FP16| 1.63| 0.07| 1.71| 1.73| 1.81| 94,758| 1.00| 601| 6.98| 4.30|
28 changes: 14 additions & 14 deletions PyTorch/SpeechSynthesis/Tacotron2/trt/export_onnx2trt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,8 +66,8 @@ def main():
engine_prec = "_fp16" if args.fp16 else "_fp32"

# Encoder
shapes=[{"name": "sequences", "min": (1,4), "opt": (1,128), "max": (4,256)},
{"name": "sequence_lengths", "min": (1,), "opt": (1,), "max": (4,)}]
shapes=[{"name": "sequences", "min": (1,4), "opt": (1,128), "max": (1,256)},
{"name": "sequence_lengths", "min": (1,), "opt": (1,), "max": (1,)}]
if args.encoder != "":
print("Building Encoder ...")
encoder_engine = build_engine(args.encoder, shapes=shapes, fp16=args.fp16)
Expand All @@ -79,17 +79,17 @@ def main():
sys.exit()

# DecoderIter
shapes=[{"name": "decoder_input", "min": (1,80), "opt": (1,80), "max": (4,80)},
{"name": "attention_hidden", "min": (1,1024), "opt": (1,1024), "max": (4,1024)},
{"name": "attention_cell", "min": (1,1024), "opt": (1,1024), "max": (4,1024)},
{"name": "decoder_hidden", "min": (1,1024), "opt": (1,1024), "max": (4,1024)},
{"name": "decoder_cell", "min": (1,1024), "opt": (1,1024), "max": (4,1024)},
{"name": "attention_weights", "min": (1,4), "opt": (1,128), "max": (4,256)},
{"name": "attention_weights_cum", "min": (1,4), "opt": (1,128), "max": (4,256)},
{"name": "attention_context", "min": (1,512), "opt": (1,512), "max": (4,512)},
{"name": "memory", "min": (1,4,512), "opt": (1,128,512), "max": (4,256,512)},
{"name": "processed_memory", "min": (1,4,128), "opt": (1,128,128), "max": (4,256,128)},
{"name": "mask", "min": (1,4), "opt": (1,128), "max": (4,256)}]
shapes=[{"name": "decoder_input", "min": (1,80), "opt": (1,80), "max": (1,80)},
{"name": "attention_hidden", "min": (1,1024), "opt": (1,1024), "max": (1,1024)},
{"name": "attention_cell", "min": (1,1024), "opt": (1,1024), "max": (1,1024)},
{"name": "decoder_hidden", "min": (1,1024), "opt": (1,1024), "max": (1,1024)},
{"name": "decoder_cell", "min": (1,1024), "opt": (1,1024), "max": (1,1024)},
{"name": "attention_weights", "min": (1,4), "opt": (1,128), "max": (1,256)},
{"name": "attention_weights_cum", "min": (1,4), "opt": (1,128), "max": (1,256)},
{"name": "attention_context", "min": (1,512), "opt": (1,512), "max": (1,512)},
{"name": "memory", "min": (1,4,512), "opt": (1,128,512), "max": (1,256,512)},
{"name": "processed_memory", "min": (1,4,128), "opt": (1,128,128), "max": (1,256,128)},
{"name": "mask", "min": (1,4), "opt": (1,128), "max": (1,256)}]
if args.decoder != "":
print("Building Decoder ...")
decoder_iter_engine = build_engine(args.decoder, shapes=shapes, fp16=args.fp16)
Expand All @@ -101,7 +101,7 @@ def main():
sys.exit()

# Postnet
shapes=[{"name": "mel_outputs", "min": (1,80,32), "opt": (1,80,768), "max": (4,80,1664)}]
shapes=[{"name": "mel_outputs", "min": (1,80,32), "opt": (1,80,768), "max": (1,80,1664)}]
if args.postnet != "":
print("Building Postnet ...")
postnet_engine = build_engine(args.postnet, shapes=shapes, fp16=args.fp16)
Expand Down