search

kuroko1t / nne

convert a pytorch model to a model for edge device
Apache License 2.0
18 stars 3 forks source link

multiple input model convertion error for tflite #19

Closed kuroko1t closed 3 years ago

kuroko1t commented 3 years ago 
model = torch.nn.Transformer()                                                                                                                                                                                     
src = np.random.rand(10, 32, 512).astype(np.float32)                                                                                                                                                               
tgt = np.random.rand(20, 32, 512).astype(np.float32)                                                                                                                                                               

src_shape = (10, 32, 512) 
tgt_shape = (20, 32, 512)  

tflite_file = "resnet.tflite"                                                                                                                                                                                      
bm = nne.Benchmark(name='tflite')                                                                                                                                                                                  
nne.cv2tflite(model, (src_shape, tgt_shape), tflite_file)                                                                                                                                                          
tflite_model = nne.load_tflite(tflite_file)                                                                                                                                                                        
nne.infer_tflite(tflite_model, (src, tgt), bm=bm)

Even though the model's input shape is set to (10, 32, 512) , (20, 32, 512) , but when I checked interpreter.get_input_details(), it was as follows.

[{'name': 'serving_default_query.7:0', 'index': 0, 'shape': array([ 20,  32, 512], dtype=int32), 'shape_signature': array([ 20,  32, 512], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}, {'name': 'serving_default_input:0', 'index': 1, 'shape': array([ 10,  32, 512], dtype=int32), 'shape_signature': array([ 10,  32, 512], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}]

So, this caused dimension mismatch error 

ValueError: Cannot set tensor: Dimension mismatch. Got 10 but expected 20 for dimension 0 of input 0.
kuroko1t commented 3 years ago

maybe.. using onnx-tf 1.8.0, this bug was resolved.