@@ -2266,28 +2266,6 @@ def rewrite_conv2d_with_pad(g, ops):
2266
2266
return ops
2267
2267
2268
2268
2269
- def rewrite_const_sub_with_pack (g , ops ):
2270
- # slice op needs "begin" and "size" are const while tf fold_const can't fold const_sub with pack
2271
- pattern = \
2272
- OpTypePattern ("Pack" , name = "pack" , inputs = [
2273
- OpTypePattern ("Sub" , name = "sub" , inputs = [
2274
- OpTypePattern ("Const" ),
2275
- OpTypePattern ("Const" )
2276
- ])
2277
- ])
2278
- matcher = GraphMatcher (pattern )
2279
- match_results = list (matcher .match_ops (ops ))
2280
- for match in match_results :
2281
- sub = match .get_op ("sub" )
2282
- sub_res = sub .inputs [0 ].get_tensor_value () - sub .inputs [1 ].get_tensor_value ()
2283
- utils .make_sure (isinstance (sub_res , (int , float )), "pack input here should be a scalar" )
2284
- pack = match .get_op ("pack" )
2285
- np_val = np .array ([sub_res ]).astype (utils .map_onnx_to_numpy_type (g .get_dtype (pack .output [0 ])))
2286
- const = g .make_const (utils .make_name ("const_val" ), np_val )
2287
- g .replace_all_inputs (ops , pack .output [0 ], const .output [0 ])
2288
- return g .get_nodes ()
2289
-
2290
-
2291
2269
def tensorflow_onnx_mapping (g , continue_on_error , custom_op_handlers ):
2292
2270
mapped_op = collections .Counter ()
2293
2271
unmapped_op = collections .Counter ()
@@ -2494,7 +2472,7 @@ def process_tf_graph(tf_graph, continue_on_error=False, verbose=False, target=No
2494
2472
2495
2473
# pre-processing graph rewrites
2496
2474
# bi-directional re-writer should be placed after single directional re-writer
2497
- rewriters = [rewrite_transpose , rewrite_flatten , rewrite_const_sub_with_pack ,
2475
+ rewriters = [rewrite_transpose , rewrite_flatten ,
2498
2476
rewrite_random_uniform , rewrite_random_uniform_fold_const ,
2499
2477
rewrite_random_normal , rewrite_dropout ,
2500
2478
rewrite_leakyrelu , rewrite_conv2d_with_pad ,
0 commit comments