Convert tensorflow frozen graph to uff error:Invalid scale mode, nbWeights: 32

[qiuyang163]

Hi,every one!Please help me.
I have a frozen graph about inception resnet V2,and I want to convert it to uff file.After I fiish the conversion,uff parser showed me an error below:

[TensorRT] ERROR: UffParser: Parser error: show_and_tell/main/image_text_sim_1/InceptionResnetV2/InceptionResnetV2/Conv2d_1a_3x3/BatchNorm/FusedBatchNorm: Invalid scale mode, nbWeights: 32
[TensorRT] ERROR: Network must have at least one output

here is my code:

OUTPUT_NODE = 'show_and_tell/main/image_text_sim_1/dot/MatMul'
BATCH = 8
HEIGHT = 299
WIDTH = 299
CHANNEL = 3
TEXT_SIZE = (BATCH, 20)
IMAGE_SIZE = (BATCH, HEIGHT, WIDTH, CHANNEL)

MAX_WORKSPACE = 2 << 30


def get_engine():
    logger = trt.Logger(trt.Logger.WARNING)
    builder = trt.Builder(logger)
    builder.max_batch_size = BATCH
    builder.max_workspace_size = MAX_WORKSPACE
    input_order = trt.tensorrt.UffInputOrder.NCHW
    with builder:
        network = builder.create_network()
        parser = trt.UffParser()
        parser.register_input(INPUT_NODE_IMAGE, IMAGE_SIZE, input_order)
        parser.register_input(INPUT_NODE_TEXT, TEXT_SIZE)
        parser.register_output(OUTPUT_NODE)
        parser.parse(uff_name, network)
        engine = builder.build_cuda_engine(network)
        return engine

An here is mu graph input definition,and FusedBatchNorm node definition:

import/show_and_tell/main/image_text_sim_1/InceptionResnetV2/InceptionResnetV2/Conv2d_1a_3x3/convolution
Operation: Conv2D
Attributes (6)
T
{"type":"DT_FLOAT"}
data_format
{"s":"NHWC"}
dilations
{"list":{"i":[1,1,1,1]}}
padding
{"s":"VALID"}
strides
{"list":{"i":[1,2,2,1]}}
use_cudnn_on_gpu
{"b":true}
Inputs (2)
import/show_and_tell/main/image_text_sim_1/map/TensorArrayStack/TensorArrayGatherV3 ?×299×299×3
import/InceptionResnetV2/Conv2d_1a_3x3/weights/read 3×3×3×32
Outputs (1)
import/show_and_tell/main/image_text_sim_1/InceptionResnetV2/InceptionResnetV2/Conv2d_1a_3x3/BatchNorm/FusedBatchNorm ?×149×149×32

import/show_and_tell/main/image_text_sim_1/InceptionResnetV2/InceptionResnetV2/Conv2d_1a_3x3/BatchNorm/FusedBatchNorm
Operation: FusedBatchNorm
Attributes (4)
T
{"type":"DT_FLOAT"}
data_format
{"s":"NHWC"}
epsilon
{"f":0.0010000000475}
is_training
{"b":false}
Inputs (5)
import/show_and_tell/main/image_text_sim_1/InceptionResnetV2/InceptionResnetV2/Conv2d_1a_3x3/convolution ?×149×149×32
import/InceptionResnetV2/Conv2d_1a_3x3/BatchNorm/beta/read 32
import/InceptionResnetV2/Conv2d_1a_3x3/BatchNorm/moving_mean/read 32
import/InceptionResnetV2/Conv2d_1a_3x3/BatchNorm/moving_variance/read 32
import/show_and_tell/main/image_text_sim_1/InceptionResnetV2/InceptionResnetV2/Conv2d_1a_3x3/BatchNorm/Const 32
Outputs (1)
import/show_and_tell/main/image_text_sim_1/InceptionResnetV2/InceptionResnetV2/Conv2d_1a_3x3/Relu ?×149×149×32

[rmccorm4]

Hi @qiuyang163,

I notice the input shape for FusedBatchNorm is ?×149×149×32, where I'm guessing the ? means a dynamic batch/dimension.

Dynamic shapes were added in the TensorRT 6.x release, and I think you ran this in TensorRT 5.x given the date - You can try upgrading to 6.0 and trying again to see if your problem is solved.

Otherwise, maybe one of these posts can help:

• https://devtalk.nvidia.com/default/topic/1063279/tensorrt/error-uffparser-parser-error-gender_resnet0_batchnorm0-mul-invalid-scale-mode-nbweights-3/post/5387684/#5387684
• https://devtalk.nvidia.com/default/topic/1045471/tensorrt/error-uffparser-parser-error-sub-invalid-scale-mode-nbweights-3/post/5304943/#5304943

[rmccorm4]

Closing since no response, feel free to open a new issue or ask to re-open this one.