Rocal classification PR #30
Conversation
rocAL/rocAL/source/augmentations/geometry_augmentations/node_resize.cpp
Outdated
Show resolved
Hide resolved
rocAL/rocAL/source/augmentations/geometry_augmentations/node_resize.cpp
Outdated
Show resolved
Hide resolved
rocAL/rocAL/source/augmentations/geometry_augmentations/node_resize.cpp
Outdated
Show resolved
Hide resolved
| src_w_dims = _inputs[0]->info().get_roi_width_vec(); | ||
| src_h_dims = _inputs[0]->info().get_roi_height_vec(); | ||
| original_w_dims = _inputs[0]->info().get_original_roi_width_vec(); |
There was a problem hiding this comment.
We could have an if condition to check if it is resize shorter, then we can assign original roi width and height to the src_w_dims and src_h_dims and then call the adjust_out_roi_size function to maintain uniformity
Also add a comment
There was a problem hiding this comment.
Resolved in the latest commit. Will trigger a training to confirm convergence. Slightly doubtful whether std::lround in adjust_roi_size may cause difference in the resize dimensions.
| host_memory_padding_jpeg2k=0, hybrid_huffman_threshold=1000000, memory_stats=False, normalized_anchor=True, | ||
| normalized_shape=True, output_type=types.RGB, preserve=False, seed=1, split_stages=False, use_chunk_allocator=False, | ||
| use_fast_idct=False, device=None, decode_size_policy=types.USER_GIVEN_SIZE_ORIG, max_decoded_width=1000, max_decoded_height=1000): | ||
|
|
||
| reader = Pipeline._current_pipeline._reader | ||
| #Reader -> Randon BBox Crop -> ImageDecoderSlice | ||
| #Random crop parameters taken from pytorch's RandomResizedCrop default function arguments |
There was a problem hiding this comment.
Is it specifically used for classification?
There was a problem hiding this comment.
Checked with swetha and found that its used only for classification
| // Using original width and height instead of the decoded width and height for computing resize shorter dimensions | ||
| src_w_dims = _inputs[0]->info().get_original_roi_width_vec(); | ||
| src_h_dims = _inputs[0]->info().get_original_roi_height_vec(); | ||
| } |
There was a problem hiding this comment.
For all modes we should calculate using original width and height ryt?
Can we change this to default?
There was a problem hiding this comment.
Resolved in the latest commit
| @@ -90,7 +90,7 @@ Decoder::Status FusedCropTJDecoder::decode(unsigned char *input_buffer, size_t i | |||
| max_decoded_width * planes, | |||
| max_decoded_height, | |||
| tjpf, | |||
| TJFLAG_FASTDCT, &x1_diff, &crop_width_diff, | |||
| TJFLAG_ACCURATEDCT, &x1_diff, &crop_width_diff, | |||
There was a problem hiding this comment.
Are we not converging with TJFLAG_FASTDCT. Has that been checked?
Also was there any performance impact because of change from fast to accurate?
There was a problem hiding this comment.
Will be triggering a training to confirm convergence with FastDCT and use FastDCT if its converging. With regards to performance, I've not observed any performance drop with AccurateDCT algorithm
Changed required for classification training convergence