Can `slice` from relay support empty result?

Development
1

Hi

For strided_slice(data, begin, end, strides=None) , when begin=[0] and end=[0], strided_slice should return an empty array. Currently strided_slice will raise an error.

Can we adjust the code so strided_slice will return an empty array instead of raising an error?

2

I have the same problem and add judgment:

if end == 0:
    return _expr.const([],dtype = dtype)

A new error is encountered later, and I’m not sure if it’s related to this return value

3

There’s another piece of information,an error occurred while obtaining the end value:

def _stridedSlice():
     xxxx
     end = _get_list_param(params, inputs[2])
     xxxx

error info:
<class 'tvm.relay.expr.Call'> has no attribute name_hint

So,I change the code:

try:
    end = _get_list_param(params, inputs[2])
except (IndexError, KeyError, AttributeError):
    end = _infer_value(inputs[2], params).asnumpy().tolist()

Finally,I get the ‘end’ value of 0(end = [0]).But begin = [-1], after running ‘_transform_mask’ function, begin = [3] (because input shape is (4,),so [-1:] == [3:]).

I think the ‘end = _infer_value(inputs[2], params).asnumpy().tolist()’ is wrong. if not have ‘end’,the ‘enc’ value may be ‘len(inputs[0])’. example:

input.shape = (4,)
input[-1:] = input[-1:4] = input[3:4]
4

@kevinthesun As @lsy643 said, Can we adjust the code so strided_slice will return an empty array instead of raising an error?

Such as when ‘begin > end’ or ‘end == 0’ or etc, we return an empty array or ‘_expr.const([],dtype = dtype)’

5

https://github.com/apache/incubator-tvm/blob/e89b19d95d/topi/include/topi/transform.h#L620 begin == end case is already supported.

6

@kevinthesun My code has been updated recently, but still error:

tvm/src/relay/op/tensor/transform.cc: Check failed: begin_v < end_v (0 vs 0) :stride_slice get empty slice at axis 1

begin = [0,0]
end = [15,0]

When have an ‘end_v < begin_v error’ , it still seems to throw an error instead of continuing. It’s not just “end_v == begin_v”.

7

In this case, same change need to be made for relay: CHECK_LT(begin_v, end_v) → CHECK_LE(begin_v, end_v)

8

@heliqi @lsy643 I encountered a similar problem, did you solve it ？

this error:

strided_slice(%307, meta[relay.Constant][0], meta[relay.Constant][1], meta[relay.Constant][2], begin=[0, 3, 0, 0], end=[1, 0, 1, 512], strides=[1, 1, 1, 1]) 
an internal invariant was violated while typechecking your program [12:17:52]  
/root/worker/tvm/src/relay/op/tensor/transform.cc:2074: 
Check failed: begin_v <= end_v (3 vs. 0) : strided_slice get invalid slice at axis 1
9

What is the sematics of begin=3 and end=0 in the original framework? This relay node is illegal since it generates negative slice.

10

Test model is MobileNetV3, use TF to load saved model to test, it works right.

For tf.strided_slice(input_, begin, end, strides=None, …) , if begin > end, it will return empty list [] instead of raising an error.

11

Hi,

I got the exactly same error when I running TF SSD-ResNet34 model downloaded from MLperf model zoo https://zenodo.org/record/3345892/files/tf_ssd_resnet34_22.1.zip?download=1

fn (%map/TensorArrayUnstack/TensorArrayScatter/TensorArrayScatterV3_loop_var: List[static_tensor_float32_15130_4_t[]], %map/Const_loop_var: int32) {
  %0 = take(%map/Const_loop_var, 0);
  %1 = @tensor_array_read_float32_15130_4(%map/TensorArrayUnstack/TensorArrayScatter/TensorArrayScatterV3_loop_var, %0);
  %2 = @tensor_get_data_float32_15130_4(%1);
  strided_slice(%2, begin=[0, 3], end=[15130, 0], strides=[1, 1]) an internal invariant was violated while typechecking your program [09:38:40] /home/ww/2020/tvm_latest/incubator-tvm/src/relay/op/tensor/transform.cc:2102: Check failed: begin_v <= end_v (3 vs. 0) : strided_slice get invalid slice at axis 1

And I use ./incubator-tvm/tests/python/frontend/tensorflow/test_forward.py test_forward_ssd() to do the test. I just modify model_path, shape, in_node and out_node

model_path = './tf_ssd_resnet34_22.1/resnet34_tf.22.1.pb' 
data = np.random.uniform(0.0, 255.0, size=(1, 3, 1200, 1200)).astype('float32')
in_node = "image"
out_node = ['detection_bboxes', "detection_scores", "detection_classes"]
12

Hi, have you solve the problem? Could you give me some advice about the following error:

Check failed: begin_v <= end_v (3 vs. 0) : strided_slice get invalid slice at axis 1