In the example below TF and PT operate on different channels, what is the reasoning behind this? is the order of the data different in TF than that of PT?
pt_tensor = torch.rand((5,80,256)) tf_tensor = tf.random.uniform((5,80, 256)) tf_conv = tf.keras.layers.Conv1D(128,3) pt_conv = torch.nn.Conv1d(80, 128,3) tf_conv(tf_tensor).shape # shape = TensorShape([5, 78, 128]) pt_conv(pt_tensor).shape # shape = torch.Size([5, 128, 254])
How are tensors stored? Are PT tensor stored as
batch_size + [in_hight, in_width] and TF tensors stores as
batch_size + [in_width, in_height]?