fix bugs in ort examples

Signed-off-by: yuwenzho <[email protected]>

Author: yuwenzho

examples/onnxrt/image_recognition/unet/quantization/ptq_static/README.md

@@ -16,7 +16,7 @@ pip install -r requirements.txt
 ## 2. Prepare Model
 
 ```bash
-python prepare_model.py --input_model='CompVis/stable-diffusion-v1-4' --output_model='.'
+python prepare_model.py --input_model='CompVis/stable-diffusion-v1-4' --output_model='unet-export.onnx'
 ```
 
 # Run

examples/onnxrt/image_recognition/unet/quantization/ptq_static/main.py

@@ -32,7 +32,7 @@ class Dataloader:
     def __init__(self, batch_size):
         self.batch_size = batch_size
         shape = [[batch_size, 4, 64, 64], [batch_size], [batch_size, 77, 768]]
-        dtype = ['float32', 'int64', 'float32']
+        dtype = ['float32', 'float32', 'float32']
         self.dataset = []
         for idx in range(0, len(shape)):
             tensor = np.random.uniform(size=shape[idx])

examples/onnxrt/image_recognition/unet/quantization/ptq_static/prepare_model.py

@@ -1,5 +1,6 @@
 import argparse
 import os
+import shutil
 import subprocess
 
 
@@ -9,10 +10,28 @@ def parse_arguments():
     parser.add_argument("--output_model", type=str, required=True)
     return parser.parse_args()
 
+def move_and_rename_model(source_folder, destination_folder):
+    if not os.path.exists(source_folder):
+        raise RuntimeError("{} path is not exists".format(source_folder))
+    for file_name in os.listdir(source_folder):
+        source_file = os.path.join(source_folder, file_name)
+        destination_file = os.path.join(destination_folder, file_name)
+        
+        if os.path.isdir(source_file):
+            continue
+    
+        shutil.move(source_file, destination_file)
+
+        if file_name == "model.onnx":
+            new_file_name = "unet-export.onnx"
+            new_file_path = os.path.join(destination_folder, new_file_name)
+            os.rename(destination_file, new_file_path)
 
 def prepare_model(input_model, output_model):
     # Use [tf2onnx tool](https://github.com/onnx/tensorflow-onnx) to convert tflite to onnx model.
     print("\nexport model...")
+
+    export_file = "prepare_unet"
     subprocess.run(
         [
             "git",
@@ -34,11 +53,18 @@ def prepare_model(input_model, output_model):
             "--model_path",
             input_model,
             "--output_path",
-            output_model,
+            export_file,
         ],
         stdout=subprocess.PIPE,
         text=True,
     )
+
+    move_and_rename_model(os.path.join(export_file, "unet"), os.path.dirname(output_model))
+    try:
+        shutil.rmtree(export_file, ignore_errors=True)
+    except OSError as e:
+        raise e
+    
     assert os.path.exists(output_model), f"Export failed! {output_model} doesn't exist!"
 
 

examples/onnxrt/nlp/huggingface_model/text_classification/quantization/ptq_static/prepare_model.py

@@ -88,6 +88,6 @@ def export_onnx_model(args, model):
 
     if args.input_model == 'Intel/bart-large-mrpc':
         import os
-        os.system('python -m transformers.onnx --model=Intel/bart-large-mrpc --feature=sequence-classification bart-large-mrpc/')
+        os.system('python -m transformers.onnx --model=Intel/bart-large-mrpc --feature=sequence-classification --export_with_transformers bart-large-mrpc/')
     else:
         export_onnx_model(args, model)

examples/onnxrt/nlp/roberta/quantization/ptq_dynamic/prepare_model.py

@@ -25,6 +25,21 @@ def parse_arguments():
                         help='Maximum length of the sentence pairs')
     return parser.parse_args()
 
+def comment_out_line(filepath, code):
+    modified_lines = []
+
+    with open(filepath, 'r') as file:
+        lines = file.readlines()
+        file.seek(0)
+        for line in lines:
+            if re.match(code, line.strip()):
+                import pdb;pdb.set_trace()
+                line = "#" + line
+            modified_lines.append(line)
+    
+    with open(filepath, 'w') as file:
+        file.writelines(modified_lines)
+
 def prepare_model(input_model, output_model, task_name):
     print("\nexport model...")
     subprocess.run(
@@ -42,6 +57,10 @@ def prepare_model(input_model, output_model, task_name):
         text=True,
     )
 
+    # remove transformers min version check
+    comment_out_line("my_transformers/examples/pytorch/text-classification/run_glue.py", 
+                     r"check_min_version\(.*\)")
+
     subprocess.run(
         [
             "python",

examples/onnxrt/nlp/roberta/quantization/ptq_static/prepare_model.py

@@ -1,5 +1,5 @@
 import argparse
-import os
+import re
 import subprocess
 
 import torch
@@ -25,6 +25,21 @@ def parse_arguments():
                         help='Maximum length of the sentence pairs')
     return parser.parse_args()
 
+def comment_out_line(filepath, code):
+    modified_lines = []
+
+    with open(filepath, 'r') as file:
+        lines = file.readlines()
+        file.seek(0)
+        for line in lines:
+            if re.match(code, line.strip()):
+                import pdb;pdb.set_trace()
+                line = "#" + line
+            modified_lines.append(line)
+    
+    with open(filepath, 'w') as file:
+        file.writelines(modified_lines)
+
 def prepare_model(input_model, output_model, task_name):
     print("\nexport model...")
     subprocess.run(
@@ -42,6 +57,10 @@ def prepare_model(input_model, output_model, task_name):
         text=True,
     )
 
+    # remove transformers min version check
+    comment_out_line("my_transformers/examples/pytorch/text-classification/run_glue.py", 
+                     r"check_min_version\(.*\)")
+
     subprocess.run(
         [
             "python",

neural_compressor/adaptor/ox_utils/operators/lstm.py

@@ -128,10 +128,12 @@ def convert(self, convert_format):
 
         kwargs = {}
         for attribute in node.attribute:
-            kwargs.update(attribute_to_kwarg(attribute))
-        kwargs["domain"] = ms_domain
+            if attribute.name == "layout":
+                continue
+            kwarg = attribute_to_kwarg(attribute)
+            kwargs.update(kwarg)
 
         quant_lstm_name = node.name + "_quant"
-        quant_lstm_node = onnx.helper.make_node("DynamicQuantizeLSTM", inputs, node.output, quant_lstm_name, **kwargs)
+        quant_lstm_node = onnx.helper.make_node("DynamicQuantizeLSTM", inputs, node.output, quant_lstm_name, domain="com.microsoft", **kwargs)
         self.quantizer.remove_nodes.append(node)
         self.quantizer.new_nodes.append(quant_lstm_node)