Created by: F0REacH

I am trying to replicate model similar to DeepSpeech2 and neon example ( [spectrogram ~4000 timesteps each of size 512] -> [1D or 2D invariant convolutions] -> [bidirectional GRU] -> [fully connected layer] -> [warp-ctc] )

No problem with:

[bidirectional GRU] -> [fully connected layer] -> [warp-ctc]

But I have no idea how to add CNN layer.

Docs says:

img_conv_layer - Convolution layer for image. Paddle only support square input currently and thus input image’s width equals height.

I am providing data as:

settings.input_types = [dense_vector_sequence(512)]

Non-square CNN inputs are really not supported or there is some way to load data?