Bug: incorrect output dimensions in yolo layer #60

New Issue

2020-06-28T10:47:27+02:00

mrhosseini commented

2020-06-28 10:47:27 +02:00

(Migrated from github.com)

In yolov3_tiny network dimensions of the first yolo layer is 13 x 13 x 255 and dimensions of the second one is 26 x 26 x 255. However, if we use a network configuration in which unlike yolov3_tiny the dimensions of the first yolo layer is bigger than the second one (like this), this line:

checkCuda( cudaMemcpy(predictions, dstData, output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost));

will fail with the error:

Cuda failure: invalid argument

The problem is that output_dim of the layer is different from the dimensions of the dstData and with debugging it can be found that for the first yolo layer the output_dim of it equals to the second one and vice versa. (Or we can say that the dimensions of dstData belongs the other layer).

I'm not sure if other parameters of the layer has the same problem.

In yolov3_tiny network dimensions of the first yolo layer is `13 x 13 x 255` and dimensions of the second one is `26 x 26 x 255`. However, if we use a network configuration in which unlike yolov3_tiny the dimensions of the first yolo layer is bigger than the second one ([like this](https://github.com/ceccocats/tkDNN/files/4776580/config.cfg.txt)),[ this line:](https://github.com/ceccocats/tkDNN/blob/6d9beb1ec5da217d8c6b1b2b51bfabf6766c3f95/src/Yolo.cpp#L123) ```C checkCuda( cudaMemcpy(predictions, dstData, output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost)); ``` will fail with the error: ``` Cuda failure: invalid argument ``` The problem is that `output_dim` of the layer is different from the dimensions of the `dstData` and with debugging it can be found that for the first yolo layer the `output_dim` of it equals to the second one and vice versa. (Or we can say that the dimensions of `dstData` belongs the other layer). I'm not sure if other parameters of the layer has the same problem.

mrhosseini commented

2020-06-28 12:12:42 +02:00

(Migrated from github.com)

After more debugging I found that, order of yolo layers in pluginFactory of NetworkRT after desrialization (here), is different from what we have in NetworkRT::buffersRT which is based on engineRT->getBindingDimensions() (here).

After more debugging I found that, order of yolo layers in `pluginFactory` of `NetworkRT` after desrialization ([here](https://github.com/ceccocats/tkDNN/blob/6d9beb1ec5da217d8c6b1b2b51bfabf6766c3f95/src/NetworkRT.cpp#L629)), is different from what we have in `NetworkRT::buffersRT` which is based on `engineRT->getBindingDimensions()` ([here](https://github.com/ceccocats/tkDNN/blob/6d9beb1ec5da217d8c6b1b2b51bfabf6766c3f95/src/NetworkRT.cpp#L185)).

mrhosseini commented

2020-06-28 12:28:14 +02:00

(Migrated from github.com)

Changing this line to the following will solve the problem. But I am not sure if this works for all the cases.

rt_out[i] = (dnnType *) netRT->buffersRT[netRT->pluginFactory->n_yolos - i] + netRT->buffersDIM[netRT->pluginFactory->n_yolos - i].tot() * bi;

Changing [this line](https://github.com/ceccocats/tkDNN/blob/6d9beb1ec5da217d8c6b1b2b51bfabf6766c3f95/src/Yolo3Detection.cpp#L95) to the following will solve the problem. But I am not sure if this works for all the cases. ```C rt_out[i] = (dnnType *) netRT->buffersRT[netRT->pluginFactory->n_yolos - i] + netRT->buffersDIM[netRT->pluginFactory->n_yolos - i].tot() * bi; ```

ceccocats commented

2020-07-06 14:08:26 +02:00

(Migrated from github.com)

Hi,
The order of buffersRt should always be:
Input
Output 0
Output 1
Etc.

If this really happen I need to check

Hi, The order of buffersRt should always be: Input Output 0 Output 1 Etc. If this really happen I need to check

mrhosseini commented

2020-07-06 16:17:45 +02:00

(Migrated from github.com)

The order of buffersRt should always be:
Input
Output 0
Output 1
Etc.

What happens is that order of yolo layers in NetworkRT::pluginFactory is different from bufferRT. Example for 2 yolo layers :

Input
Output 0 [26 x 26 x 18]
Output 1 [13 x 13 x 18]

in bufferRT but in pluginFactory:

Yolo 0 [13 x 13 x 18]
Yolo 1 [26 x 26 x 18]

> The order of buffersRt should always be: > Input > Output 0 > Output 1 > Etc. What happens is that order of yolo layers in `NetworkRT::pluginFactory` is different from `bufferRT`. Example for 2 yolo layers : ``` Input Output 0 [26 x 26 x 18] Output 1 [13 x 13 x 18] ``` in `bufferRT` but in `pluginFactory`: ``` Yolo 0 [13 x 13 x 18] Yolo 1 [26 x 26 x 18] ```

This repo is archived. You cannot comment on issues.

1 Participants

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: mmr/tkDNN#60