Hi I loaded a noisy one channel .wav file to perform inference using <code class="notr

<a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/us

Real-time-processing Tf_lite inference program ERROR about dtln HOT 6 CLOSED

breizhn commented on September 28, 2024

Real-time-processing Tf_lite inference program ERROR

from dtln.

Comments (6)

breizhn commented on September 28, 2024 1

Should be fixed now.

from dtln.

cybermax-008 commented on September 28, 2024 1

Should be fixed now.

Thanks for the quick fix!
The audio file is denoised finally.

from dtln.

StuartIanNaylor commented on September 28, 2024

@breizhn I just did the convert_weights_to_tf_lite.py then on a model I had just trained that is also default.

        self.fs = 16000
        self.batchsize = 32
        self.len_samples = 15
        self.activation = 'sigmoid'
        self.numUnits = 128
        self.numLayer = 2
        self.blockLen = 512
        self.block_shift = 128
        self.dropout = 0.25
        self.lr = 1e-3
        self.max_epochs = 200
        self.encoder_size = 256
        self.eps = 1e-7

Then run

python real_time_processing_tf_lite.py
Traceback (most recent call last):
  File "real_time_processing_tf_lite.py", line 69, in <module>
    interpreter_1.set_tensor(input_details_1[0]['index'], in_mag)
  File "/home/stuart/DTLN/venv/lib/python3.8/site-packages/tflite_runtime/interpreter.py", line 705, in set_tensor
    self._interpreter.SetTensor(tensor_index, value)
ValueError: Cannot set tensor: Dimension mismatch. Got 3 but expected 4 for input 0.

Its tflite_runtime 2.7.0 trained on 2.7.1 @breizhn Am I doing something wrong or have things regressed?

Works with pretrained

interpreter_1 = tflite.Interpreter(model_path='./pretrained_model/model_1.tflite')
interpreter_1.allocate_tensors()
interpreter_2 = tflite.Interpreter(model_path='./pretrained_model/model_2.tflite')
interpreter_2.allocate_tensors()

But ones I just trained

interpreter_1 = tflite.Interpreter(model_path='./DTLN_2.tflite')
interpreter_1.allocate_tensors()
interpreter_2 = tflite.Interpreter(model_path='./DTLN_1.tflite')
interpreter_2.allocate_tensors()

I am guessing you have to train for a stateful model?
What setting do I need for modelTrainer.build_DTLN_model_stateful() ?

from dtln.