profile
viewpoint

jeremydvoss/Mizrach 0

Shows ya where to face while praying, with impeccable scientific accuracy

issue commentmozilla/DeepSpeech

Does batch size matter for inference?

Thanks for the help! Looks like I need to completely retrain since my checkpoints are of a different shape now. Anyways to still use my old checkpoints?

jeremydvoss

comment created time in a month

issue openedmozilla/DeepSpeech

Does batch size matter for inference?

I've trained my own network on a subset of the english common voice with these parameters (via a bash script): language=enLimited10000 epochs=100 batch_size=128 lr=0.0001 dropout_rate=0.30 n_hidden=1000 ./DeepSpeech.py --train_files ../CommonVoice/$language/clips/train.csv --dev_files ../CommonVoice/$language/clips/dev.csv --test_files ../CommonVoice/$language/clips/test.csv \ --epochs $epochs \ --test_batch_size $batch_size \ --train_batch_size $batch_size \ --dev_batch_size $batch_size \ --export_batch_size $batch_size \ --learning_rate $lr \ --export_dir $HOME/EXPORT_DIR/$language \ --dropout_rate $dropout_rate \ --n_hidden $n_hidden \ --train_cudnn \ --checkpoint_dir ~/.local/share/deepspeech/checkpoints$language/ \ --scorer data/lm/kenlm.scorer \

However when I do the inference, I get this error: `Loading model from file ../EXPORT_DIR/enLimited10000/output_graph.pb TensorFlow: v1.15.0-24-gceb46aa DeepSpeech: v0.7.0-alpha.2-0-g9dd63d5 Warning: reading entire model file into memory. Transform model file into an mmapped graph to reduce heap usage. 2020-03-04 00:35:56.526314: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA Loaded model in 0.0634s. Loading scorer from files ../DeepSpeech3/data/lm/kenlm_BACKUP.scorer Loaded scorer in 0.000286s. Running inference. Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [128] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [128] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [128] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [128] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [128] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [128] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [128] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]]

Inference took 0.486s for 1.975s audio file.`

When I tried the inference for a smaller network where I used a batch size of 32, I got this: `TensorFlow: v1.15.0-24-gceb46aa DeepSpeech: v0.7.0-alpha.2-0-g9dd63d5 Warning: reading entire model file into memory. Transform model file into an mmapped graph to reduce heap usage. 2020-03-04 01:07:25.709064: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA Loaded model in 0.0576s. Loading scorer from files ../DeepSpeech3/data/lm/kenlm_BACKUP.scorer Loaded scorer in 0.000378s. Running inference. Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [32] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [32] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [32] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [32] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [32] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [32] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]] Error running session: Invalid argument: Shapes of all inputs must match: values[0].shape = [1] != values[1].shape = [32] [[{{node cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/stack}}]]

Inference took 0.441s for 1.975s audio file.`

So it seems to me like it's saying that a network trained on a batch size greater than 1 cannot infer for a single audio file. Is that true? That would mean I could only train for a batch size of 1, limiting the utilization of my gpus. Perhaps I am misunderstanding the difference between train- dev- test- and export- bash sizes.

created time in a month

issue commentmozilla/DeepSpeech

error: unrecognized arguments: --scorer deepspeech-0.6.1-models/kenlm.scorer

Thanks! It works if specify

$ pip3 install deepspeech==0.7.0a2

... or rather I'm onto some other bug. Thanks for your help!

crowlogic

comment created time in a month

issue commentmozilla/DeepSpeech

error: unrecognized arguments: --scorer deepspeech-0.6.1-models/kenlm.scorer

Oh I see. Is there a way to install the newest version of deepspeech then? "pip3 install deepspeech" seems to install 0.6.1

crowlogic

comment created time in a month

issue commentmozilla/DeepSpeech

error: unrecognized arguments: --scorer deepspeech-0.6.1-models/kenlm.scorer

I am having this same issue. I have trained a model. I now what to test it on such an audio file, but get this error. The DeepSpeech I installed is up to date: TensorFlow: v1.14.0-21-ge77504a DeepSpeech: v0.6.1-0-g3df20fe So is the github documentation under "transcribe an audio file" out of date? https://github.com/mozilla/DeepSpeech

crowlogic

comment created time in a month

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha 912a980fe00b51b12f469706d457fc105d3a79ad

part 2 cleaned up

view details

push time in 5 months

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha 18b0683a14586918d49d08ba98208e72193fd1f1

part 2 works!

view details

push time in 5 months

pull request commentpjreddie/uwimg

Updated my fork for hw1

jeremydvoss

comment created time in 5 months

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha b3b800055d3100fb43a650e5214525a5fe3ce64c

part 2 not working

view details

push time in 5 months

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha a5cc504f0c2af1b988d40fcd2d38c9f6f28ac2ef

hw3 part 1

view details

push time in 5 months

push eventjeremydvoss/uwimg

Joseph Redmon

commit sha ec397bea45aa1002ff3876b7069afc3d13a5c2b4

adding hw3

view details

Joseph Redmon

commit sha 76a95b738eae060b3b630cd20ed6c667ff06ebea

panoramam tests

view details

Joseph Redmon

commit sha 7bb6e7fa0cea587abcd378d44c378bfea80d250c

path fixes

view details

Joseph Redmon

commit sha f6ac406156c90fba02d2f1682a77cda8d5d9e222

no hw6

view details

Joseph Redmon

commit sha 703d7791656c55e043f813a1bcf32d710afc71f5

added clamp test

view details

Joseph Redmon

commit sha a2625408a9e24a00ee29983cc07e2e5b41929e0d

BUT WHO TESTS THE TESTS???

view details

Joseph Redmon

commit sha 8af559b653dd786f328a04cd7454700b329de613

more precise tests

view details

Joseph Redmon

commit sha bd5e23b7520530c99ff2ea60da2d439275895186

Merge branch 'master' of https://github.com/pjreddie/uwimg

view details

Joseph Redmon

commit sha 9e258c3acea5b69f52e179f7a832b77948e98acc

maybe better...

view details

jeremydvoss

commit sha d94d119c717a28587eeb8934cfbe26b2813388d4

Merge branch 'master' into master

view details

push time in 5 months

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha 10d14e03f40330465bd8e11df94a8cf9f0dfdb29

hw2

view details

push time in 5 months

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha 4b91a3fd807d154cf73f6cc0541d6f36ae31e973

More accurate blur on sobel image

view details

push time in 5 months

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha 9cbc3d2eed2f0da78dc91c18fa817351b6db306a

Passes hw2

view details

push time in 5 months

push eventjeremydvoss/uwimg

Joseph Redmon

commit sha 2189103e3d32582d3b5c6699270f2878e716ef55

Update process_image.c

view details

Joseph Redmon

commit sha f3d4ba62676d375ba104f5feca1de54ab3d1b400

Update process_image.c

view details

Joseph Redmon

commit sha 9c4b4fed92cbeacb1866b815a52c4beff2865350

Update uwimg.py

view details

Joseph Redmon

commit sha 13009a719f0b44829d4dfe9e11bda8679092e1e3

Revert "Update uwimg.py" This reverts commit 9c4b4fed92cbeacb1866b815a52c4beff2865350.

view details

Joseph Redmon

commit sha 1a3d1ee2e111ff4afd1f5ab4cf531441ea2c2221

Revert "Update process_image.c" This reverts commit f3d4ba62676d375ba104f5feca1de54ab3d1b400.

view details

Joseph Redmon

commit sha 9ec82244274a57d4ad0c927405e8fe6b793e7f7f

Revert "Update process_image.c" This reverts commit 2189103e3d32582d3b5c6699270f2878e716ef55.

view details

Joseph Redmon

commit sha 988e832f379ddd1143242fe3d49004ddb9ee8960

Revert "scale image in library" This reverts commit a697571f5cbdd84719a1e91d9bfdae6e12da1a59.

view details

Joseph Redmon

commit sha 7a3f3e9988ab9822f6a5fa041faabc60f6367bcf

homework 2 stuff

view details

Joseph Redmon

commit sha 9e368837b722f886033f9e506ae8ded4272f6f16

python file, updated paths

view details

jeremydvoss

commit sha 134689c1024ba4b16cc3b59db4b9e7871e0d2bc6

Merge branch 'master' into master

view details

push time in 5 months

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha cb089fe7264e46112ff0de5cfc071d58bee1ebc5

Commited removal of scale image from hw0

view details

push time in 5 months

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha 68f18b456dda64ce642b69dbea5df94af4c26d70

hw1 working

view details

Jeremy Voss

commit sha aa3a0aba9b6783be21121285aed546a465d65004

test.c no is more specific

view details

push time in 5 months

push eventjeremydvoss/uwimg

Joseph Redmon

commit sha 78891acea0038f36f576eafcfc482dca6cf8afbf

homework 1

view details

Joseph Redmon

commit sha 4903e35f7c6032ed16293aa78b1d0cf6ca98b544

figure paths

view details

Joseph Redmon

commit sha a697571f5cbdd84719a1e91d9bfdae6e12da1a59

scale image in library

view details

Joseph Redmon

commit sha f526fd0a2b775e12288a250b5d6bf6588746ee2a

test code

view details

jeremydvoss

commit sha 117106e7758d93c81799fa8f0d410f8b1030b4f4

Merge branch 'master' into master

view details

push time in 6 months

PR opened pjreddie/uwimg

Updated my fork for hw1
+216 -14

0 comment

4 changed files

pr created time in 6 months

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha 5fe0cc5e8b4b387fb9097da0f196a3fac238c680

hw1 works, but not tests

view details

push time in 6 months

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha 1320f71103f68fc73ccdb44443dcc060771afa0e

Passes all tests and works

view details

push time in 6 months

push eventjeremydvoss/uwimg

Jeremy Voss

commit sha 55ff7680ffbafa5941cddb847b4fe24081549106

Only failing hsv reversal

view details

push time in 6 months

fork jeremydvoss/uwimg

Build your own computer vision library in C

fork in 6 months

more