Description
Describe the bug
i want to use tensorrt to accelarate my model ,so i need to convert pb to oNNX，
first i use tensorflow 1.13.2 and estimator api to save pb
but my model is larger than 2G, it seems that tf2onnx not support --large_model
so i change the code in tensorflow 2.8
i also save model with estimator api
but when i convert the pb to onnx by command:
python -m tf2onnx.convert --saved-model ./export_model/1741318171 --output model.onnx --large_model
i got
function_optimizer: function_optimizer did nothing. time = 0.002ms.
Optimization results for grappler item: __inference_batch_normalization_1_cond_1_true_5630_491
function_optimizer: function_optimizer did nothing. time = 0.002ms.
function_optimizer: function_optimizer did nothing. time = 0.001ms.
2025-03-07 11:38:10,820 - INFO - Using tensorflow=2.8.0, onnx=1.12.0, tf2onnx=1.10.1/a37f29
2025-03-07 11:38:10,820 - INFO - Using opset <onnx, 9>
2025-03-07 11:38:10,965 - WARNING - Removing optional input keras_learning_phase that appears to be a keras learning phase parameter. Use --ignore_default to force this into an input.
2025-03-07 11:38:11,061 - WARNING - shape missmatch between caller and function: __inference_batch_normalization_6_cond_false_7047_frozen_4698
Traceback (most recent call last):
File "/data/software/app/anaconda3/envs/py39/lib/python3.9/site-packages/tf2onnx/tf_loader.py", line 772, in resolve_functions
func = function_def_to_graph(fdef, input_shapes=input_shapes)
File "/data/software/app/anaconda3/envs/py39/lib/python3.9/site-packages/tensorflow/python/framework/function_def_to_graph.py", line 88, in function_def_to_graph
output_tensor_names = [
File "/data/software/app/anaconda3/envs/py39/lib/python3.9/site-packages/tensorflow/python/framework/function_def_to_graph.py", line 89, in
nested_to_flat_tensor_name[fdef.ret[arg.name]]
KeyError: 'ReadVariableOp:value:0'
i do not know why ,but it seems that ReadVariableOp:value:0 is tf1 op?
System information
- centos 7
- TensorFlow Version:2.8
- Python version:3.9
- ONNX version1.12.0
-tf2onnx version 1.11.1
To Reproduce
the code i use
actor_like_emb = fc.numeric_column(
'actor_like_emb',
shape=[1024],
default_value=[0.0 for i in range(1024)],
dtype=tf.float32
)
params["img_feature_columns"]=["actor_like_emb"]
img_feature = tf.keras.layers.DenseFeatures(list(params["img_feature_columns"].values()))(features)
the loss part is
update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS) # 获取批量归一化的更新操作
optimizer = tf.compat.v1.train.ProximalAdagradOptimizer(
learning_rate=params["learning_rate"],
l2_regularization_strength=params["l2_reg"],
)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss, global_step=tf.compat.v1.train.get_global_step())
i use keras api and feature_column and compat.v1
really need help and suggestion
really thx
Screenshots
Additional context