tf2onnx.tfonnx:Tensorflow op [sequential_1_1/lstm_1/CudnnRNNV3: CudnnRNNV3] is not supported

[nassimus26]

New Operator

Having this TensorFlow model

nbr_frame = 10
img_width = 180
img_height = 150
img_size = (img_height, img_width)
input_shape = (3,) + img_size
full_input_shape = (nbr_frame,) + input_shape
print(full_input_shape)
np.random.seed(1234)
num_classes = 2
tf.keras.backend.set_image_data_format(
    'channels_first'
)
#vg19 = tf.keras.applications.vgg19.VGG19
#base_model = vg19(include_top=False,weights='imagenet',input_shape=(img_width, img_height,3))
base_model = tf.keras.applications.MobileNetV2(
    include_top=False, weights='imagenet', input_tensor=None,
    input_shape = input_shape,
    pooling=None, 
)
for layer in base_model.layers:
    layer.trainable = False
base_model.summary()
cnn = models.Sequential()
cnn.add(base_model)
cnn.add(layers.GlobalAveragePooling2D())
cnn.add(layers.Dropout(0.2))
base_model.trainable = False
# define LSTM model
model = models.Sequential()
print(full_input_shape)
model.add(layers.TimeDistributed(cnn, input_shape=full_input_shape))
model.add(layers.LSTM(nbr_frame, return_sequences=True))
model.add(layers.TimeDistributed(layers.Dense(nbr_frame, activation='relu')))
model.add(layers.Flatten())
model.add(layers.Dense(164, activation='relu', name="filter"))
model.add(layers.Dropout(0.2))
model.add(layers.Dense(24, activation='sigmoid', name="filter2"))
model.add(layers.Dropout(0.1))
model.add(layers.Dense(num_classes, activation="sigmoid", name="last"))
rms = optimizers.RMSprop()
metrics = [tf.keras.metrics.CategoricalAccuracy('accuracy', dtype=tf.float32)]
loss = tf.keras.losses.CategoricalCrossentropy()

model.compile(
    loss=loss,
    optimizer= rms,
    metrics=metrics
)
model.summary()

Running the conversion

input_signature = [tf.TensorSpec((None,) + full_input_shape, tf.float32, name='x')]
# Use from_function for tf functions
# model.output_names=['output']
onnx_model, _ = tf2onnx.convert.from_keras(model, input_signature, opset=18)
onnx.save(onnx_model, "fights.onnx")

Return this

WARNING:tf2onnx.shape_inference:Cannot infer shape for sequential_1_1/lstm_1/CudnnRNNV3: sequential_1_1/lstm_1/CudnnRNNV3:3,sequential_1_1/lstm_1/CudnnRNNV3:4
ERROR:tf2onnx.tfonnx:Tensorflow op [sequential_1_1/lstm_1/CudnnRNNV3: CudnnRNNV3] is not supported
ERROR:tf2onnx.tfonnx:Unsupported ops: Counter({'CudnnRNNV3': 1})

More about this operation here :

https://www.tensorflow.org/api_docs/java/org/tensorflow/op/core/CudnnRNNV3

[andrenatal]

I'm having the same issue.

[tosa-no-onchan]

model.add(layers.LSTM(nbr_frame, return_sequences=True))

How about this
model.add(layers.LSTM(nbr_frame, return_sequences=True,use_cudnn=False))

and

model.compile(
    ....
    jit_compile=False
)

https://keras.io/api/layers/recurrent_layers/lstm/
https://keras.io/guides/migrating_to_keras_3/

but, I wonder
onnxruntime.InferenceSession() load ONNX.model reports error?

[KirillMonster]

Fix: We use the processor during the model conversion

import tf2onnx
import tensorflow as tf
tf.config.set_visible_devices([], 'GPU') # use cpu
# !!! important, we need to upload the saved model.
model = tf.keras.models.load_model('result/model.keras')

model.output_names = ['output']
input_signature = [tf.TensorSpec((None, max_symbols), model.inputs[0].dtype, name='digit')]
onnx_model, _ = tf2onnx.convert.from_keras(model, input_signature, opset=15)
with open('result/model.onnx', "wb") as f:
	f.write(onnx_model.SerializeToString())