check.py

# import onnxruntime as ort
# import numpy as np

# # 直接加载ONNX模型查看输入要求
# model_path = "/root/.cache/espnet_onnx/transformer_lm/full/default_encoder.onnx"
# try:
#     sess = ort.InferenceSession(model_path, providers=['CPUExecutionProvider'])
#     input_details = sess.get_inputs()
#     print("ONNX模型输入要求:")
#     for inp in input_details:
#         print(f"  名称: {inp.name}, 形状: {inp.shape}, 类型: {inp.type}")
# except Exception as e:
#     print(f"加载模型失败: {e}")




# import os
# import onnx
# import onnxruntime as ort
# import numpy as np

# # 检查ONNX模型文件
# model_path = "/root/.cache/espnet_onnx/transformer_lm/full/default_encoder.onnx"

# print("检查模型文件...")
# if os.path.exists(model_path):
#     model_size = os.path.getsize(model_path)
#     print(f"模型大小: {model_size} bytes")
    
#     # 加载模型查看结构
#     try:
#         model = onnx.load(model_path)
#         print(f"模型IR版本: {model.ir_version}")
#         print(f"生产者: {model.producer_name} {model.producer_version}")
#         print(f"模型输入: {len(model.graph.input)} 个")
#         print(f"模型输出: {len(model.graph.output)} 个")
#         print(f"节点数量: {len(model.graph.node)}")
        
#         # 查找Where节点
#         where_nodes = [node for node in model.graph.node if node.op_type == "Where"]
#         print(f"找到 {len(where_nodes)} 个Where节点")
        
#         for i, node in enumerate(where_nodes[:3]):  # 只显示前3个
#             print(f"  Where节点 {i}: {node.name}")
#             print(f"    输入: {[input for input in node.input]}")
#             print(f"    输出: {[output for output in node.output]}")
            
#     except Exception as e:
#         print(f"加载模型失败: {e}")
# else:
#     print(f"模型文件不存在: {model_path}")



import onnxruntime as ort
import numpy as np

model_path = "/root/.cache/espnet_onnx/transformer_lm/full/default_encoder.onnx"

print("=== 检查模型实际输入 ===")
sess = ort.InferenceSession(model_path, providers=['CPUExecutionProvider'])

# 详细检查输入
print("模型输入详细信息:")
for inp in sess.get_inputs():
    print(f"\n输入: {inp.name}")
    print(f"  形状: {inp.shape}")
    print(f"  类型: {inp.type}")
    
    # 打印每个维度
    for i, dim in enumerate(inp.shape):
        print(f"    维度[{i}]: {dim}")

# 尝试不同的输入名称
print("\n=== 尝试不同的输入名称 ===")

# 创建测试数据
batch_size = 1
time_frames = 100
n_mels = 80

dummy_feats = np.random.randn(batch_size, time_frames, n_mels).astype(np.float32)

# 获取所有可能的输入名称
input_names = [inp.name for inp in sess.get_inputs()]
print(f"模型接受的输入名称: {input_names}")

# 尝试所有可能的输入组合
test_inputs = []

# 常见的输入名称模式
common_names = [
    'feats', 'speech', 'input', 'x',
    'feats_length', 'speech_lengths', 'lengths', 'ilens'
]

for name in input_names:
    print(f"\n测试输入: {name}")
    
    # 根据名称猜测类型
    if 'length' in name.lower() or 'lens' in name.lower():
        # 可能是长度输入
        dummy_input = np.array([time_frames], dtype=np.int64)
    else:
        # 可能是特征输入
        dummy_input = dummy_feats
    
    try:
        outputs = sess.run(None, {name: dummy_input})
        print(f"  成功! 使用单一输入: {name}")
        print(f"  输出数量: {len(outputs)}")
        for i, out in enumerate(outputs):
            print(f"    输出{i}: {out.shape}")
        break
    except:
        print(f"  失败: 单一输入{name}")

# 尝试多输入
if len(input_names) > 1:
    print(f"\n尝试多输入组合: {input_names}")
    
    # 准备输入字典
    input_dict = {}
    for name in input_names:
        if 'length' in name.lower() or 'lens' in name.lower():
            input_dict[name] = np.array([time_frames], dtype=np.int64)
        else:
            input_dict[name] = dummy_feats
    
    try:
        outputs = sess.run(None, input_dict)
        print(f"  成功! 使用多输入")
        for i, out in enumerate(outputs):
            print(f"    输出{i}: {out.shape}")
    except Exception as e:
        print(f"  失败: {e}")