Unable to deploy tensorflow image classification model

Autonomous Machines Jetson & Embedded Systems Jetson TX2
#1

Hello guys,
I retrained a image classification model using tensorflow for poets(https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/)
The model i used is (https://github.com/aswinkumar2019/CNN-model-for-rupee-notes-except-2000rs-/retrained_graph.pb)
I converted it into uff format.The converted model is (https://github.com/aswinkumar2019/CNN-model-for-rupee-notes-except-2000rs-/model.uff)

When i try to infer the model using sample.py i get the error as shown in screenshot(https://github.com/aswinkumar2019/CNN-model-for-rupee-notes-except-2000rs-/send.png)

[TensorRT] ERROR: UffParser: Unsupported number of graph 0
[TensorRT] ERROR: Network must have at least one output
Traceback (most recent call last):
File “sample.py”, line 112, in
main()
File “sample.py”, line 98, in main
with build_engine(model_file) as engine:
AttributeError: enter

Is model retrained using tensorflow for poets not supported in tensorrt???
Please help me.
Regards,
Aswin Kumar

#2

Hi,

Thanks for you post.

Somehow we cannot open the link you shared above.
Would you mind to check it for us?

Thanks.

#3

https://github.com/aswinkumar2019/CNN-model-for-rupee-notes-except-2000rs-/blob/master/retrained_graph.pb

https://github.com/aswinkumar2019/CNN-model-for-rupee-notes-except-2000rs-/blob/master/send.png

Sorry guys these are the updated links

#4

https://github.com/aswinkumar2019/CNN-model-for-rupee-notes-except-2000rs-/blob/master/model.uff

This is the converted uff file

#5

Hello guys,
I thought that the sample.py file found in /usr/src/tensorrt/samples/python/end_to_end_tensorflow_mnist folder can be used with command ( python sample.py -d ./model.uff) to use our own tensorflow image classification model(model.uff) converted from the retrained_graph.pb file got from tensorflow for poets((https://codelabs.developers.google.com/codelabs/tensorflow-for-poets/)

Am i right or,i am doing something wrong???
Or,i should use nvidia DIGITS to train the model to be deployed in jetson devices???
If i am wrong,please give instructions to follow to train the model
Please help me
Regards,
Aswin Kumar

#6

Guys any update regarding the issue?
I am waiting for your response

#7

Hi,

Sorry for the late update.
This issue is caused by a non-supported layer.

We try to run your uff model with this sample:

import os
import ctypes
import tensorrt as trt

uff_filename = 'model.uff'
trt_filename = 'model.trt'

input_name  = 'Input'
output_name = 'MarkOutput_0'
input_dim   = [3, 224, 224]


# initialize
TRT_LOGGER = trt.Logger(trt.Logger.INFO)
trt.init_libnvinfer_plugins(TRT_LOGGER, '')
runtime = trt.Runtime(TRT_LOGGER)


if not os.path.isfile(trt_filename):
    with trt.Builder(TRT_LOGGER) as builder, builder.create_network() as network, trt.UffParser() as parser:
        builder.max_workspace_size = 1 << 28
        builder.max_batch_size = 1
        builder.fp16_mode = True

        parser.register_input(input_name, input_dim)
        parser.register_output(output_name)
        parser.parse(uff_filename, network)
        engine = builder.build_cuda_engine(network)

        buf = engine.serialize()
        with open(model.TRTbin, 'wb') as f:
            f.write(buf)

[TensorRT] ERROR: UffParser: Validator error: input_1/BottleneckInputPlaceholder: Unsupported operation _PlaceholderWithDefault
[TensorRT] ERROR: Network must have at least one output
Traceback (most recent call last):
  File "inference.py", line 30, in <module>
    buf = engine.serialize()
AttributeError: 'NoneType' object has no attribute 'serialize'

Is it possible to replace the PlaceholderWithDefault op with some supported operation listed here:
https://docs.nvidia.com/deeplearning/sdk/tensorrt-archived/tensorrt-601/tensorrt-support-matrix/index.html#supported-ops

Thanks.