Input layer not found in function 'cv::dnn::dnn4_v20210608::`anonymous-namespace'::TFImporter::connect'

[FrancescoMandru]

I'm trying to load my Tensorflow model with opencv DNN as follows:

1:

model.save('mymodel', save_format='tf')

2:

import tensorflow as tf
from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2

physical_devices = tf.config.list_physical_devices('GPU')
tf.config.set_visible_devices(physical_devices[0],'GPU')
tf.config.experimental.set_memory_growth(physical_devices[0], True)

loaded = tf.saved_model.load('mymodel')
infer = loaded.signatures['serving_default']

f = tf.function(infer).get_concrete_function(input_1=tf.TensorSpec(shape=[1, 256, 256, 1], 
                                                                        dtype=tf.float32))

f2 = convert_variables_to_constants_v2(f, aggressive_inlining=True)
graph_def = f2.graph.as_graph_def()

# Export frozen graph
with tf.io.gfile.GFile('frozen_graph.pb', 'wb') as f:
    f.write(graph_def.SerializeToString())

3:

net = cv2.dnn.readNet('frozen_graph.pb')
img = np.random.standard_normal((1, 256, 256, 1)).astype(np.float32)
print(img.shape)
net.setInput(img)
out = net.forward()
print(out.shape)

but I get the following error:

---------------------------------------------------------------------------
error                                     Traceback (most recent call last)
~\AppData\Local\Temp/ipykernel_13288/1971587229.py in <module>
      1 import cv2 as cv2
      2 
----> 3 net = cv2.dnn.readNet('frozen_graph.pb')
      4 img = np.random.standard_normal((1, 256, 256, 1)).astype(np.float32)
      5 print(img.shape)

error: OpenCV(4.5.3) C:\Users\runneradmin\AppData\Local\Temp\pip-req-build-c2l3r8zm\opencv\modules\dnn\src\tensorflow\tf_importer.cpp:2445: error: (-2:Unspecified error) Input layer not found: StatefulPartitionedCall/StatefulPartitionedCall/my_net/Decoderblock1/conv2d_transpose/stack/1 in function 'cv::dnn::dnn4_v20210608::`anonymous-namespace'::TFImporter::connect'

Tensorflow version: 2.5.0
OpenCV version 4.5.3

This is my frozen graph:
frozen_graph.zip

[d5423197]

I am having the same error. Opencv DNN is really difficult to use. I have tried to convert tf 2 model to onnx model. I can successfully use it in python onnx module. But I cannot read it using opencv dnn.

[FrancescoMandru]

I think that the Conv2DTranspose is not implemented yet and this may cause some problems during the conversion. I have also other issues when there is asymmetric padding applied by Tensorflow

[LaurentBerger]

@opencv-alalek why this issue is incomplete? I can reproduce error usng opencv-5.0 pre

cv2.error: OpenCV(5.0.0-pre) C:\lib\opencv\modules\dnn\src\tensorflow\tf_importer.cpp:2853: error: (-2:Unspecified error) Input layer not found: StatefulPartitionedCall/StatefulPartitionedCall/my_net/Decoderblock1/conv2d_transpose/stack/1 in function cv::dnn::dnn5_v20231225::anonymous-namespace'::TFImporter::connect'