- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
hi,
I convert caffemodel with order "./ModelOptimizer -p FP16 -w ResNet_262/xx.caffemodel -d ResNet_262/deploy_xx.prototxt -i -b 1 "; generate IR files in defalut direct——Artifacts. then I test this model by "./multi_output_sample -i armstrong.bmp -m xx_fp16.xml -d GPU", I get error as follow:
[ ERROR ] program creation failed: ../../../../src/gpu/pooling_gpu.cpp at line: 133
Error has occured for: pool5_7x7_s1_cldnn_output_postprocess
Best_kernel.empty()(true) should be false
Cannot find a proper kernel with this arguments
i don't konw which step is wrong, how can i use "fp16" peculiarity in my application?
thanks for answer!
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Someone will be following up with you soon. But in the meantime, please post your proto.txt and also your IR XML file. There may be an unsupported layer (unsupported by the GPU) in these files.
Thanks,
Shubha
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Thanks for your reply! I attached the xml and protoxt.
Maybe the error caused by last layer "Pooling", the last layer in my prototxt:
layer { name: "pool5_7x7_s1" type: "Pooling" bottom: "res5b5_" top: "pool5_7x7_s1" pooling_param { pool: AVE global_pooling: true } }
The corespondinig XML as follow:
<layer name="pool5_7x7_s1" type="Pooling" precision="FP16" id="610"> <pooling_data kernel-x="4" kernel-y="4" pad-x="0" pad-y="0" stride-x="1" stride-y="1" rounding-type="ceil" pool-method="avg"/> <input> <port id="1305"> <dim>1</dim> <dim>2048</dim> <dim>4</dim> <dim>4</dim> </port> </input> <output> <port id="1306"> <dim>1</dim> <dim>2048</dim> <dim>1</dim> <dim>1</dim> </port> </output> </layer> </layers>
I have another question: when I convert caffemodel to Intermediate Representation (IR) with precision(fp16), and i have trained the caffemodel with other version caffe, then how can get the factor (with the option of command line "-f") .
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi Hou,
Does your network work with FP32 well? To reproduce the issue we also need the .caffemodel file, could you please provide it?
Best wishes,
Anna
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Anna B. (Intel) wrote:
Hi Hou,
Does your network work with FP32 well? To reproduce the issue we also need the .caffemodel file, could you please provide it?
Best wishes,
Anna
Hi Anna,
yeah, my network work well with FP32! Because of limitation of the attachment' size, I split the model into four parts.
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page