tiny yolo v3 conversion

[chrisrapson]

When I use your tool to convert my cfg and weight files to tensorflow models, there seems to be an error. Here is the resulting graph from tensorboard, and I think that the connection down the right hand side should connect to a convolutional block, not a maxpool block. Specifically, convolutional5.

AlexeyAB's explanation of route layers shows yolov2 has a similar connection to a convolutional block.

Also when I tried converting with this similar repo it connects to a convolutional block.

I'm not 100% sure I've understood route blocks correctly, but it could be that you count the input net1 as a block? If so, then you should increment the index (in this case from 8 to 9). Of course that only applies where the route has a positive number.

[sjain-stanford]

@chrisrapson Thanks for reporting this. Indeed, this is an indexing issue. The stack (in DW2TF) stores layers starting with net at index 0, whereas @AlexeyAB's comment clarifies that indexing excludes the first net (placeholder). When route layers (basically concat operation) use relative indexing (negative values), this was not an issue (e.g. yolo-v2). However with absolute indexing (positive values), we're off by one layer (e.g. tiny-yolo-v3).

I've identified a simple fix, but need to verify it doesn't break other use-cases. Will push shortly. Let me know if you see any other issues. Thanks again!

[sjain-stanford]

I've fixed this by excluding net layer from the stack. Relative indices should work as usual. Absolute indexing starts with the first true layer. Below is the revised yolov3-tiny graph; route layer concatenates convolutional5 and not maxpool4.

This fixes yolov3 and yolov3-tiny conversions, but has no impact on yolov2 (used relative indexing) or yolov2-tiny (no route layers). For those using DW2TF to convert yolov3 or yolov3-tiny, please use the latest release DW2TF-1.2.

@chrisrapson Closing this, please test and feel free to re-open if you see issues.