Daniil Lyakhov
Daniil Lyakhov
### Details: HF Transformer pruning support. Changes: Supported ops/ patterns mask propagation * Transpose * Special case for Self-Attention reshape * MatMul batch dimensions * MatMul with only masks on...
### Details: - *item1* - *...* ### Tickets: - *ticket-id*
### Changes Redundant input shape param was removed from `count_flops_and_weights_per_node` function. `count_filters_num` now always takes into account pruned linear layers. `count_flops_and_weights_per_node` now correctly computes statistics for model with pruned linear...
Updated config to train nncf compressed model in 2e2 with expected accuracy drop. Please see details in a ticket 84131
### Changes ### Reason for changes ### Related tickets ### Tests
### Changes Test for QAT after PTQ for torch classifications ### Reason for changes To check PTQ quantized model is trainable in torch ### Related tickets 129584 ### Results Testcase...
On top of #2515 ### Changes PTQ quantization aware training documentation ### Reason for changes To improve docs coverage ### Related tickets 129584
### Changes * PT QAT statistic collection is using `PTStatisticsAggregator` to aggregate statistics * `generate_collectors_for_range_init_statistics_collection` function is simplified * `get_input_shape_and_channel_idx` is introduced to remove code duplicates ### Reason for changes...
### Changes `port_id` and `parallel_edges_ids` are added to the graph for analysis  ### Reason for changes To test `parallel_edges_ids` and `port_id` in graph tests as well as the structural...
### Search before asking - [X] I have searched the YOLOv8 [issues](https://github.com/ultralytics/ultralytics/issues) and found no similar bug report. ### YOLOv8 Component Train, Val ### Bug #### Train ``` Ultralytics YOLOv8.2.5...