Giters

ggerganov / ggml

Tensor library for machine learning

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

mnist convert from tensorflow to ggml

datduonguva opened this issue · comments

commented

in the file examples/mnist/mnist-cnn.py

    kernel1 = model.layers[0].weights[0].numpy()                                                    
    kernel1 = np.moveaxis(kernel1, [2,3], [0,1])                                                    
    kernel1 = kernel1.astype(np.float16)                                                            
    gguf_writer.add_tensor("kernel1", kernel1, raw_shape=(32, 1, 3, 3))                             
    ...                                                                                                
    kernel2 = model.layers[2].weights[0].numpy()                                                    
    kernel2 = np.moveaxis(kernel2, [0,1,2,3], [2,3,1,0])                                            
    kernel2 = kernel2.astype(np.float16)                                                            
    gguf_writer.add_tensor("kernel2", kernel2, raw_shape=(64, 32, 3, 3))

why are kernel1's axes and kernel2's axes moved by np.moveaxis differently?

kernel1 has shape (3, 3, 1, 32) so after np.moveaxis(kernel1, [2, 3], [0, 1]), its shape will be (1, 32, 3, 3), NOT (32, 1, 3, 3) as showned on the next line.

Wondering if this is a bug.
Thanks!