(4.x) Merge 3.4

testdata/dnn/download_models.py

@@ -51,7 +51,8 @@ def verify(self):
                         break
                     sha.update(buf)
             print('  actual {}'.format(sha.hexdigest()))
-            return self.sha == sha.hexdigest()
+            self.sha_actual = sha.hexdigest()
+            return self.sha == self.sha_actual
         except Exception as e:
             print('  catch {}'.format(e))
 
@@ -82,7 +83,10 @@ def get(self):
 
         print(' done')
         print(' file {}'.format(self.filename))
-        return self.verify()
+        candidate_verify = self.verify()
+        if not candidate_verify:
+            self.handle_bad_download()
+        return candidate_verify
 
     def download(self):
         try:
@@ -112,6 +116,27 @@ def save(self, r):
                 print('>', end='')
                 sys.stdout.flush()
 
+    def handle_bad_download(self):
+        if os.path.exists(self.filename):
+            # rename file for further investigation
+            try:
+                # NB: using `self.sha_actual` may create unbounded number of files
+                rename_target = self.filename + '.invalid'
+                # TODO: use os.replace (Python 3.3+)
+                try:
+                    if os.path.exists(rename_target):  # avoid FileExistsError on Windows from os.rename()
+                        os.remove(rename_target)
+                finally:
+                    os.rename(self.filename, rename_target)
+                    print('  renaming invalid file to ' + rename_target)
+            except:
+                import traceback
+                traceback.print_exc()
+            finally:
+                if os.path.exists(self.filename):
+                    print('  deleting invalid file')
+                    os.remove(self.filename)
+
 
 def GDrive(gid):
     def download_gdrive(dst):

testdata/dnn/onnx/generate_onnx_models.py

@@ -6,6 +6,7 @@
 import torch.nn.functional as F
 import tensorflow as tf # version 2.5.0
 import tf2onnx # version 1.9.1
+import paddle # version 2.1.1
 import numpy as np
 import os.path
 import onnx
@@ -1159,6 +1160,19 @@ def forward(self, x):
 model = Scale()
 save_data_and_model("scale", x, model)
 
+class ScaleBroadcast(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super(ScaleBroadcast, self).__init__()
+
+    def forward(self, x0, x1, x2):
+        return torch.mul(torch.mul(x0, x1), x2)
+
+model = ScaleBroadcast()
+input_0 = Variable(torch.ones(2, 1, 4, 5, dtype=torch.float32))
+input_1 = Variable(torch.ones(1, 4, 1, dtype=torch.float32))
+input_2 = Variable(torch.ones(2, 1, 4, 1, dtype=torch.float32))
+save_data_and_model_multy_inputs("scale_broadcast", model, input_0, input_1, input_2)
+
 x = Variable(torch.randn(1, 3, 25))
 conv1d = nn.Conv1d(3, 2, kernel_size=3, padding=2, stride=2, dilation=2, bias=False)
 save_data_and_model("conv1d", x, conv1d)
@@ -1310,6 +1324,19 @@ def forward(self, x):
 save_data_and_model("average_pooling_dynamic_axes", input, ave_pool)
 postprocess_model("models/average_pooling_dynamic_axes.onnx", [[1, 3, 'height', 'width']])
 
+class DynamicBatch(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super(DynamicBatch, self).__init__()
+        self.pool = nn.MaxPool2d(2, stride=2)
+
+    def forward(self, x):
+        return torch.cat((self.pool(x), torch.ones(2, 3, 1, 2)))
+
+model = DynamicBatch()
+input_ = Variable(torch.ones(2, 3, 3, 4, dtype=torch.float32))
+save_data_and_model("dynamic_batch", input_, model, export_params=True)
+postprocess_model("models/dynamic_batch.onnx", [['batch_size', 3, 3, 4]])
+
 x = Variable(torch.randn(1, 3, 10))
 max_pool = nn.MaxPool1d(kernel_size=(5), stride=1, padding=2, dilation=1)
 save_data_and_model("maxpooling_1d", x, max_pool)
@@ -1443,4 +1470,27 @@ def cumsum_reverse(x):
 def cumsum_exclusive_1d_reverse(x):
     return tf.cumsum(x, exclusive=True, reverse=True)
 
-save_data_and_tf_function(cumsum_exclusive_1d_reverse, "cumsum_1d_exclusive_1_reverse", x)
+save_data_and_tf_function(cumsum_exclusive_1d_reverse, "cumsum_1d_exclusive_1_reverse", x)
+
+#paddle2onnx model
+class Resize_HumanSeg(paddle.nn.Layer):
+    def __init__(self, ):
+        super(Resize_HumanSeg, self).__init__()
+
+    def forward(self, x0):
+        x1 = paddle.nn.functional.interpolate(x0,size=[6,8],mode='bilinear',align_corners=False)
+        return x1
+
+def save_data_and_paddle_model(model, name, input_data):
+    model.eval()
+    np.save(os.path.join("data", "input_" + name + ".npy"), input_data.numpy())
+    output = model(input_data)
+    np.save(os.path.join("data", "output_" + name + ".npy"), output.numpy())
+    inputs = [paddle.static.InputSpec(shape=input_data.shape, dtype="float32")]
+    paddle.onnx.export(model, "models/" + name,
+                       input_spec=inputs,
+                       opset_version=11)
+
+input_shape = [1, 2, 3, 4]
+x = paddle.rand(input_shape, dtype="float32")
+save_data_and_paddle_model(Resize_HumanSeg(), "resize_humanseg", x)

testdata/dnn/onnx/models/scale_broadcast.onnx

@@ -0,0 +1,30 @@
+pytorch1.9:�

+
+
0
+
1
3Mul_0"Mul
+
+
3
+
2
4Mul_1"Multorch-jit-exportZ
+
0
+

+
+

+
+Z
+
1
+

+

+
+
Z
+
2
+

+
+

+
+
b
+
4
+

+
+

+
+B