maxmustermann/DeepLearningExamples
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

Merge pull request #91 from GrzegorzKarchNV/readme-fix

Browse source 
fixed batches for mixed precision and fp32, added info when no input …
This commit is contained in:
nvpstr 2019-07-01 11:43:29 +02:00 committed by GitHub
parent 40e074257f d7c20e5bfc
commit eabbedd1d4
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 10 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

17
PyTorch/SpeechSynthesis/Tacotron2/README.md
View file

@ -155,7 +155,9 @@ To run inference issue:
```bash
python inference.py --tacotron2 <Tacotron2_checkpoint> --waveglow <WaveGlow_checkpoint> -o output/ -i phrase.txt --fp16-run
```
The speech is generated from text file passed with `-i` argument. To run
The speech is generated from text file passed with `-i` argument.
If no file is provided or if the provided file cannot be opened, speech will be
generated from a default text located in the `inference.py` file. To run
inference in mixed precision, use `--fp16-run` flag. The output audio will
be stored in the path specified by `-o` argument.
@ -406,7 +408,7 @@ training script in the PyTorch-19.05-py3 NGC container on NVIDIA DGX-1 with
Tacotron 2 and output samples per second for WaveGlow) were averaged over
an entire training epoch.
This table shows the results for Tacotron 2, with batch size equal 48 and 80
This table shows the results for Tacotron 2, with batch size equal 80 and 48
for mixed precision and FP32 training, respectively.
|Number of GPUs|Mixed precision tokens/sec|FP32 tokens/sec|Speed-up with mixed precision|Multi-gpu weak scaling with mixed precision|Multi-gpu weak scaling with FP32|
@ -415,7 +417,8 @@ for mixed precision and FP32 training, respectively.
|**4**|7,768|5,683|1.37|3.04|3.27|
|**8**|12,524|10,484|1.19|4.90|6.03|
The following table shows the results for WaveGlow, with batch size equal 4 and 8 for mixed precision and FP32 training, respectively.
The following table shows the results for WaveGlow, with batch size equal 8 and
4 for mixed precision and FP32 training, respectively.
|Number of GPUs|Mixed precision samples/sec|FP32 samples/sec|Speed-up with mixed precision|Multi-gpu weak scaling with mixed precision|Multi-gpu weak scaling with FP32|
|---:|---:|---:|---:|---:|---:|
@ -427,17 +430,17 @@ To achieve these same results, follow the [Quick Start Guide](#quick-start-guide
### Expected training time
This table shows the expected training time for convergence for Tacotron 2 (1500 epochs).
This table shows the expected training time for convergence for Tacotron 2 (1500 epochs, time in hours).
|Number of GPUs|Expected training time with mixed precision|Expected training time with FP32|Speed-up with mixed precision|
|Number of GPUs|Expected training time in hours with mixed precision|Expected training time in hours with FP32|Speed-up with mixed precision|
|---:|---:|---:|---:|
|**1**| 197.39 | 302.32 | 1.38 |
|**4**| 63.29 | 88.07 | 1.25 |
|**8**| 33.72 | 45.51 | 1.33 |
This table shows the expected training time for convergence for WaveGlow (1000 epochs).
This table shows the expected training time for convergence for WaveGlow (1000 epochs, time in hours).
|Number of GPUs|Expected training time with mixed precision|Expected training time with FP32|Speed-up with mixed precision|
|Number of GPUs|Expected training time in hours with mixed precision|Expected training time in hours with FP32|Speed-up with mixed precision|
|---:|---:|---:|---:|
|**1**| 400.99 | 782.67 | 1.95 |
|**4**| 89.40 | 213.09 | 2.38 |