bricewalker / hey-jetson Goto Github PK

I am interested in this project for doing emotion detection from speech.

Do we need to use flask for inferencing on jetson. I mean there are projects which do inferencing by direct deployment of model like here https://github.com/dusty-nv/jetson-inference

Any advice for emotion detection from speech for this project

Thanks

Hi brice,
I'm getting this error and I don't understand it at all. Some help would be much appreciated. I've installed all the dependencies and prepared the dataset etc... training starts too but getting this annoying error.

2019-07-17 14:57:26.395650: W tensorflow/core/framework/op_kernel.cc:1502] OP_REQUIRES failed at variable_ops.cc:104 : Already exists: Resource __per_step_3/training/Adam/gradients/bidirectional_7/while_1/strided_slice_3/Enter_grad/ArithmeticOptimizer/AddOpsRewrite_Add/tmp_var/N10tensorflow19TemporaryVariableOp6TmpVarE 2019-07-17 14:57:26.396630: W tensorflow/core/framework/op_kernel.cc:1502] OP_REQUIRES failed at variable_ops.cc:104 : Already exists: Resource __per_step_3/training/Adam/gradients/bidirectional_7/while_1/strided_slice_3/Enter_grad/ArithmeticOptimizer/AddOpsRewrite_Add/tmp_var/N10tensorflow19TemporaryVariableOp6TmpVarE 2019-07-17 14:57:26.397577: W tensorflow/core/framework/op_kernel.cc:1502] OP_REQUIRES failed at variable_ops.cc:104 : Already exists: Resource __per_step_3/training/Adam/gradients/bidirectional_7/while_1/strided_slice_3/Enter_grad/ArithmeticOptimizer/AddOpsRewrite_Add/tmp_var/N10tensorflow19TemporaryVariableOp6TmpVarE Traceback (most recent call last): File "train_model.py", line 842, in <module> spectrogram=True) # True for Spectrogram/False for MFCC File "train_model.py", line 536, in train_model callbacks=[checkpointer, terminator, logger, time_machiner, tensor_boarder], verbose=verbose) File "/home/nvidia/anaconda3/lib/python3.7/site-packages/keras/legacy/interfaces.py", line 91, in wrapper return func(*args, **kwargs) File "/home/nvidia/anaconda3/lib/python3.7/site-packages/keras/engine/training.py", line 1418, in fit_generator initial_epoch=initial_epoch) File "/home/nvidia/anaconda3/lib/python3.7/site-packages/keras/engine/training_generator.py", line 217, in fit_generator class_weight=class_weight) File "/home/nvidia/anaconda3/lib/python3.7/site-packages/keras/engine/training.py", line 1217, in train_on_batch outputs = self.train_function(ins) File "/home/nvidia/anaconda3/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py", line 2715, in __call__ return self._call(inputs) File "/home/nvidia/anaconda3/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py", line 2675, in _call fetched = self._callable_fn(*array_vals) File "/home/nvidia/anaconda3/lib/python3.7/site-packages/tensorflow/python/client/session.py", line 1458, in __call__ run_metadata_ptr) tensorflow.python.framework.errors_impl.AlreadyExistsError: Resource __per_step_3/training/Adam/gradients/bidirectional_7/while_1/strided_slice_3/Enter_grad/ArithmeticOptimizer/AddOpsRewrite_Add/tmp_var/N10tensorflow19TemporaryVariableOp6TmpVarE [[{{node training/Adam/gradients/bidirectional_7/while_1/strided_slice_3/Enter_grad/ArithmeticOptimizer/AddOpsRewrite_Add/tmp_var}}]]

I have Jetson nano, with L4T 4.2.
The pip install -r jetson_requirement.txt doesn't run to completion.

First package which is not found is apturl.

Hi Bricewalker,
This is a very interest project. Could you please share the hardware components?

like this project:

• https://arxiv.org/pdf/1811.00787.pdf
• https://github.com/Alexander-H-Liu/End-to-end-ASR-Pytorch

Hello @bricewalker ,

thanks for sharing this. Very good work. Can this be scaled to multigpu training?

Getting this error:

conda env create -f environment.yml
Solving environment: failed

ResolvePackageNotFound:

• openssl==1.1.1b=hfa6e2cd_2
• m2w64-flac==1.3.1=3
• icu==58.2=ha66f8fd_1
• h5py==2.9.0=py37h5e291fa_0
• python==3.7.3=h8c8aaf0_1
• qt==5.9.7=vc14h73c81de_0
• mkl-service==2.0.2=py37he774522_0
• cryptography==2.7=py37hb32ad35_0
• pyyaml==5.1=py37he774522_0
• winpty==0.4.3=4
• tensorflow-base==1.13.1=gpu_py37h871c8ca_0
• matplotlib==3.1.0=py37hc8f65d3_0
• wincertstore==0.2=py37_0
• mkl_fft==1.0.12=py37h14836fe_0
• grpcio==1.16.1=py37h351948d_1
• msys2-conda-epoch==20160418=1
• m2w64-speex==1.2rc2=3
• tensorboard==1.13.1=py37h33f27b4_0
• yaml==0.1.7=hc54c509_2
• protobuf==3.8.0=py37h33f27b4_0
• mkl_random==1.0.2=py37h343c172_0
• m2w64-libogg==1.3.2=3
• m2w64-gcc-libs==5.3.0=7
• m2w64-gcc-libgfortran==5.3.0=6
• m2w64-libsndfile==1.0.26=2
• libprotobuf==3.8.0=h7bd577a_0
• zlib==1.2.11=h62dcd97_3
• hdf5==1.10.4=h7ebc959_0
• pyqt==5.9.2=py37h6538335_2
• libsodium==1.0.16=h9d3ae62_0
• sip==4.19.8=py37h6538335_0
• pyrsistent==0.14.11=py37he774522_0
• m2w64-gcc-libs-core==5.3.0=7
• mistune==0.8.4=py37he774522_0
• sqlite==3.28.0=he774522_0
• m2w64-speexdsp==1.2rc3=3
• pyzmq==18.0.0=py37ha925a31_0
• pywinpty==0.5.5=py37_1000
• libpng==1.6.37=h2a8f88b_0
• m2w64-gmp==6.1.0=2
• icc_rt==2019.0.0=h0cc432a_1
• m2w64-libwinpthread-git==5.0.0.4634.697f757=2
• pandas==0.24.2=py37ha925a31_0
• pycrypto==2.6.1=py37hfa6e2cd_1002
• jpeg==9b=hb83a4c4_2
• mkl==2019.4=245
• m2w64-libvorbis==1.3.5=2
• tornado==6.0.2=py37he774522_0
• intel-openmp==2019.4=245
• cffi==1.12.3=py37h7a1dbc1_0
• statsmodels==0.9.0=py37h452e1ab_0
• numpy==1.16.4=py37h19fb1c0_0
• pyreadline==2.1=py37_1
• freetype==2.9.1=ha9979f8_1
• vs2015_runtime==14.15.26706=h3a45250_4
• win_inet_pton==1.1.0=py37_0
• scikit-learn==0.21.2=py37h6288b17_0
• kiwisolver==1.1.0=py37ha925a31_0
• vc==14.1=h0510ff6_4
• markupsafe==1.1.1=py37he774522_0
• tensorflow==1.13.1=gpu_py37h83e5d6a_0
• zeromq==4.3.1=h33f27b4_3
• numpy-base==1.16.4=py37hc3f5095_0
• scipy==1.2.1=py37h29ff71c_0

I'm trying to running this model usign a portuguese data corpus, but, I don't know how exactly use a PT-BR dataset because of the accents that have in portuguese, like:

ACCENTS = 'ãõçâêôáíóúàüóé'

how can I add these characters to this model?

ty

So I am trying to translate an audio file from the LibriSpeech data set using the run_inference() function in the make_predictions.py file by loading up model_11.h5, but the predicted translation is only 2 letters long. This is true for the other audio samples I have tried as well where I only get an output text of 2 letters.
The output array generated by the model's predict() method, which is saved under the variable "prediction", for one of the audio samples I've tried is in the shape of (1, 149, 29), but the tensor generated by the ctc_decode() method is in the shape of (1, 2). So I'm guessing that something is wrong with the decoder? I can't quite figure it out.

Is this an issue that you have encountered before? I've attached the python script I've used to run the inference along with the sample audio file and the output to the screen.

file.zip

2019-07-19 04:56:06.472125: I tensorflow/stream_executor/dso_loader.cc:153] successfully opened CUDA library libcublas.so.10 locally
3455/3455 [==============================] - 21350s 6s/step - loss: 356.8575 - val_loss: 237.5635
Epoch 2/30
1883/3455 [===============>..............] - ETA: 2:39:53 - loss: nan Batch 1882: Invalid loss, terminating training

Hello @bricewalker
Could you clarify what hardware setup you used for inference?

• Which nvidia jetson version?
• Did you use an external microphone connected to the jetson, or the integrated mic in the camera?
• Any other hardware that you connected to the jetson?
  Thanks in advance.

I've been getting problems with Keras regarding generic utility functions and other problems of modules when I try to adapt my own Jetson to get this to work.