Are variably shaped keras models supported?

[mnboos]

I’ve got a keras model here with unknown/unspecified shapes like

(None, None, None, 3)

for a Conv2D layer.

However, this does not work with TVM relay due to the following error:

  File "yolo.py", line 92, in to_onnx_tvm
    func, params = relay.frontend.from_keras(self.yolo_model, shape_dict)
  File "/home/martin/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/relay/frontend/keras.py", line 723, in from_keras
    keras_op_to_relay(inexpr, keras_layer, keras_layer.name + ':' + str(node_idx), etab)
  File "/home/martin/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/relay/frontend/keras.py", line 647, in keras_op_to_relay
    outs = _convert_map[op_name](inexpr, keras_layer, etab)
  File "/home/martin/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/relay/frontend/keras.py", line 214, in _convert_convolution
    pad_t, pad_b = _get_pad_pair(in_h, dilated_kernel_h, stride_h)
  File "/home/martin/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/relay/frontend/keras.py", line 23, in _get_pad_pair
    out1d = (input1d + stride1d - 1) // stride1d
TypeError: unsupported operand type(s) for +: 'NoneType' and 'int'

I’ve seen other questions here regarding this topic like:

[Relay] Shape inference with partial input shape information

I recently started to learn Relay by comparing it with NNVM. When Relay does shape inference, it requires complete input shape information from all the input tensors to succeed. For example, the following three cases would fail, while in NNVM they can run through. Is this because the add operator is actually a broadcast_add so that the incomplete shape info cannot be inferred, or the design of Relay has deliberately forbidden shape inference with partial input shape information? If it’s the latt…

Variable shape input support for code generation

Will TVM support variable input shape for code generation in future? In NMT and ASR, the network input size are not constant, it will be wonderful if TVM can support variable input length. @tqchen

It is a bit unclear to me, whether I’m doing something wrong or if this a bug. If it’s clear for you, that this cannot work, would it be possible to intentionally raise an error in such a case?

[kevinthesun]

Dynamic shape support is in progress, for both compilation and runtime. This error is related to keras converter, and it looks like only supporting static shape.

[mnboos]

Dynamic shape support is in progress, for both compilation and runtime.

Awesome, looking forward to it!