自定义语音识别模型转TF Lite部署Coral Dev Board：输入形状参数求助

Hey there! Let's work through this TFLite conversion issue step by step. The error pops up because your frozen graph's wav_data input has a dynamic (unknown) shape, and quantized TFLite models require fixed input dimensions to handle the quantization logic properly. Here's how to find the right shape and fix the command:

1. First, confirm the input tensor's shape from your frozen graph #

You can quickly inspect the wav_data tensor in your frozen graph with a short Python script. Since you're working with a legacy frozen graph, use TensorFlow's v1 compatibility APIs:

import tensorflow as tf

with tf.compat.v1.Session() as sess:
    # Load the frozen graph file
    with tf.io.gfile.GFile('frozen.pb', 'rb') as f:
        graph_def = tf.compat.v1.GraphDef()
        graph_def.ParseFromString(f.read())
        sess.graph.as_default()
        tf.import_graph_def(graph_def, name='')
    
    # Fetch the wav_data input tensor and print its shape
    input_tensor = sess.graph.get_tensor_by_name('wav_data:0')
    print(f"wav_data tensor shape: {input_tensor.shape}")

If the output shows something like () (scalar) or <unknown>, that confirms the shape isn't fixed in the graph—you'll need to define it based on your training setup.

2. Determine the correct input shape from your training pipeline #

The wav_data input is raw audio bytes, so its shape corresponds to the fixed length of audio clips you used during training. For example:

• If you trained on 1-second, 16kHz, 16-bit mono WAV files, each file is 16000 samples * 2 bytes/sample = 32000 bytes—so the shape is [32000].
• Adjust this based on your actual audio specs: calculate total bytes as sample_rate * duration * bytes_per_sample * num_channels.

If you're unsure, check your training preprocessing code—look for where you load or pad/truncate audio to a fixed length. That fixed length (in bytes) is your input shape.

3. Update your tflite_convert command #

Add the --input_shapes parameter with your confirmed shape. Using the 32000-byte example, your command becomes:

tflite_convert --output_file=model.tflite --graph_def_file=frozen.pb --input_arrays=wav_data --output_arrays=labels_softmax --inference_type=QUANTIZED_UINT8 --input_shapes=[32000]

Alternative: Use the TFLite Converter Python API (more flexible) #

If you prefer a script-based approach (which is easier to debug), use the Python converter:

import tensorflow as tf

# Initialize converter with your frozen graph
converter = tf.compat.v1.lite.TFLiteConverter.from_frozen_graph(
    graph_def_file='frozen.pb',
    input_arrays=['wav_data'],
    output_arrays=['labels_softmax'],
    input_shapes={'wav_data': [32000]}  # Replace with your actual shape
)

# Set quantization type
converter.inference_type = tf.compat.v1.lite.constants.QUANTIZED_UINT8

# Convert and save the model
tflite_model = converter.convert()
with open('model.tflite', 'wb') as f:
    f.write(tflite_model)

This should resolve the shape error and generate your quantized TFLite model ready for the Coral Dev Board.

内容的提问来源于stack exchange，提问作者davide85