Switch to onnxoptimizer as onnx.optimizer is removed

Looks like multiple onnx-ml.proto copies are allowed in protobuf-lite.
See also [1] for the protobuf issue.

[1] onnx/optimizer#38

.gitignore

@@ -13,6 +13,7 @@
 /common/data.cpp
 /common/data.h
 /data/others.txt
+/data/*-opt*.onnx
 
 # CCS-generated files
 .launches/

transform.py

@@ -15,6 +15,7 @@
 
 import onnx
 import onnx.helper
+import onnxoptimizer
 import numpy as np
 
 from utils import (
@@ -264,22 +265,15 @@ def get_prev_node(n):
 Constants.LEA_BUFFER_SIZE = lea_buffer_size[args.target]
 
 onnx_opt_model_name = config['onnx_model'].replace('.onnx', '-opt.onnx')
-if os.path.exists(onnx_opt_model_name):
-    onnx_model = onnx.load(onnx_opt_model_name)
-else:
-    onnx_model = onnx.load(config['onnx_model'])
-    try:
-        import onnx.optimizer
-        # https://zhuanlan.zhihu.com/p/41255090
-        onnx_model = onnx.optimizer.optimize(onnx_model, [
-            'fuse_add_bias_into_conv',
-            'fuse_matmul_add_bias_into_gemm',
-        ])
-    except IndexError:
-        # Somehow the optimizer cannot handle models transformed from keras2onnx
-        pass
-    onnx_model = onnx.shape_inference.infer_shapes(onnx_model)
-    onnx.save_model(onnx_model, onnx_opt_model_name)
+onnx_model = onnx.load(config['onnx_model'])
+# https://zhuanlan.zhihu.com/p/41255090
+onnx_model = onnxoptimizer.optimize(onnx_model, [
+    'fuse_add_bias_into_conv',
+    'fuse_matmul_add_bias_into_gemm',
+])
+
+onnx_model = onnx.shape_inference.infer_shapes(onnx_model)
+onnx.save_model(onnx_model, onnx_opt_model_name)
 g = onnx_model.graph
 names = {}
 
@@ -301,8 +295,12 @@ def find_initializer(name):
             return initializer
 
 def replace_squeeze(node, inp):
-    axes_name = node.input[1]
-    axes = find_initializer(axes_name).int64_data
+    # Since opset 13, axes is an input instead of an attribute
+    try:
+        axes_name = node.input[1]
+        axes = find_initializer(axes_name).int64_data
+    except IndexError:
+        axes = get_attr(node, 'axes')
     new_dims = [dim for dim_idx, dim in enumerate(inp.dims) if dim_idx not in axes]
     # Repeated fields cannot be assigned directly
     # https://developers.google.com/protocol-buffers/docs/reference/python-generated#repeated-fields