add

KyanChen · Nov 26, 2023 · b51e916 · b51e916
1 parent 4c732b4
commit b51e916
Show file tree

Hide file tree

Showing 7 changed files with 14 additions and 152 deletions.
diff --git a/README_zh-CN.md b/README_zh-CN.md
@@ -333,14 +333,19 @@ python zero_to_fp32.py . $SAVE_CHECKPOINT_NAME -t $CHECKPOINT_DIR  # $SAVE_CHECK
 
 这里我们列出了使用不同模型的资源消耗情况，供您参考。
 
-|         模型名称          |  骨干网络类型  | 图像尺寸 |       GPU       | Batch Size  |       加速策略       | 单卡显存占用  |
-|:---------------------:|:--------:| :------: |:---------------:|:-----------:|:----------------:|:-------:|
-| SAM-seg (Mask R-CNN)  | ViT-B/16 | 1024x1024 |   1x RTX 4090   |      8      |     AMP FP16     | 19.4 GB |
-| SAM-seg (Mask2Former) | ViT-B/16 | 1024x1024 |   1x RTX 4090   |      8      |     AMP FP16     | 21.5 GB |
-|        SAM-det        | ResNet50 | 1024x1024 |   1x RTX 4090   |      8      |       FP32       | 16.6 GB |
-|   RSPrompter-anchor   | ViT-B/16 | 1024x1024 |   1x RTX 4090   |      2      |     AMP FP16     | 20.9 GB |
-|   RSPrompter-query    | ViT-B/16 | 1024x1024 |   1x RTX 4090   |      1      |     AMP FP16     |   OOM   |
-|   RSPrompter-anchor   | ViT-B/16 | 1024x1024 |        1        | 8x RTX 4090 | DeepSpeed ZeRO-2 | 2.4 GB  |
+|         模型名称          |  骨干网络类型  | 图像尺寸 |           GPU            | Batch Size |       加速策略       | 单卡显存占用  |
+|:---------------------:|:--------:| :------: |:------------------------:|:----------:|:----------------:|:-------:|
+| SAM-seg (Mask R-CNN)  | ViT-B/16 | 1024x1024 |       1x RTX 4090        |     8      |     AMP FP16     | 19.4 GB |
+| SAM-seg (Mask2Former) | ViT-B/16 | 1024x1024 |       1x RTX 4090        |     8      |     AMP FP16     | 21.5 GB |
+|        SAM-det        | ResNet50 | 1024x1024 |       1x RTX 4090        |     8      |       FP32       | 16.6 GB |
+|   RSPrompter-anchor   | ViT-B/16 | 1024x1024 |       1x RTX 4090        |     2      |     AMP FP16     | 20.9 GB |
+|   RSPrompter-query    | ViT-B/16 | 1024x1024 |       1x RTX 4090        |     1      |     AMP FP16     |   OOM   |
+|   RSPrompter-anchor   | ViT-B/16 | 1024x1024 |       8x RTX 4090        |     4      | DeepSpeed ZeRO-2 | 2.4 GB  |
+
+
+### 4. dist_train.sh: Bad substitution的解决
+
+如果您在运行`dist_train.sh`时出现了`Bad substitution`的错误，请使用`bash dist_train.sh`来运行脚本。
 
 
 </details>

diff --git a/configs/rsprompter/_base_/rsprompter_anchor.py b/configs/rsprompter/_base_/rsprompter_anchor.py
@@ -313,49 +313,6 @@
 ]
 
 
-# runner_type = 'FlexibleRunner'
-# strategy = dict(
-#     type='DeepSpeedStrategy',
-#     fp16=dict(
-#         enabled=True,
-#         auto_cast=False,
-#         fp16_master_weights_and_grads=False,
-#         loss_scale=0,
-#         loss_scale_window=500,
-#         hysteresis=2,
-#         min_loss_scale=1,
-#         initial_scale_power=15,
-#     ),
-#     inputs_to_half=['inputs'],
-#     zero_optimization=dict(
-#         stage=2,
-#         allgather_partitions=True,
-#         allgather_bucket_size=2e8,
-#         reduce_scatter=True,
-#         reduce_bucket_size='auto',
-#         overlap_comm=True,
-#         contiguous_gradients=True,
-#     ),
-# )
-#
-# optim_wrapper = dict(
-#     type='DeepSpeedOptimWrapper',
-#     optimizer=dict(
-#         type='AdamW',
-#         lr=base_lr,
-#         weight_decay=0.05
-#     )
-# )
-
-optim_wrapper = dict(
-    type='AmpOptimWrapper',
-    dtype='float16',
-    optimizer=dict(
-        type='AdamW',
-        lr=base_lr,
-        weight_decay=0.05)
-)
-
 # Default setting for scaling LR automatically
 #   - `enable` means enable scaling LR automatically
 #       or not by default.

diff --git a/configs/rsprompter/_base_/rsprompter_query.py b/configs/rsprompter/_base_/rsprompter_query.py
@@ -314,42 +314,6 @@
     )
 ]
 
-
-runner_type = 'FlexibleRunner'
-strategy = dict(
-    type='DeepSpeedStrategy',
-    fp16=dict(
-        enabled=True,
-        auto_cast=False,
-        fp16_master_weights_and_grads=False,
-        loss_scale=0,
-        loss_scale_window=500,
-        hysteresis=2,
-        min_loss_scale=1,
-        initial_scale_power=15,
-    ),
-    gradient_clipping=0.1,
-    inputs_to_half=['inputs'],
-    zero_optimization=dict(
-        stage=2,
-        allgather_partitions=True,
-        allgather_bucket_size=2e8,
-        reduce_scatter=True,
-        reduce_bucket_size='auto',
-        overlap_comm=True,
-        contiguous_gradients=True,
-    ),
-)
-
-optim_wrapper = dict(
-    type='DeepSpeedOptimWrapper',
-    optimizer=dict(
-        type='AdamW',
-        lr=base_lr,
-        weight_decay=0.05
-    )
-)
-
 # Default setting for scaling LR automatically
 #   - `enable` means enable scaling LR automatically
 #       or not by default.

diff --git a/configs/rsprompter/_base_/samdet.py b/configs/rsprompter/_base_/samdet.py
@@ -270,16 +270,6 @@
     )
 ]
 
-optim_wrapper = dict(
-    type='OptimWrapper',
-    optimizer=dict(
-        type='AdamW',
-        lr=base_lr,
-        weight_decay=0.05,
-        eps=1e-8,
-        betas=(0.9, 0.999))
-)
-
 # Default setting for scaling LR automatically
 #   - `enable` means enable scaling LR automatically
 #       or not by default.

diff --git a/configs/rsprompter/_base_/samseg-mask2former.py b/configs/rsprompter/_base_/samseg-mask2former.py
@@ -304,52 +304,6 @@
     )
 ]
 
-runner_type = 'FlexibleRunner'
-strategy = dict(
-    type='DeepSpeedStrategy',
-    fp16=dict(
-        enabled=True,
-        auto_cast=False,
-        fp16_master_weights_and_grads=False,
-        loss_scale=0,
-        loss_scale_window=500,
-        hysteresis=2,
-        min_loss_scale=1,
-        initial_scale_power=15,
-    ),
-    gradient_clipping=0.1,
-    inputs_to_half=['inputs'],
-    zero_optimization=dict(
-        stage=2,
-        allgather_partitions=True,
-        allgather_bucket_size=2e8,
-        reduce_scatter=True,
-        reduce_bucket_size='auto',
-        overlap_comm=True,
-        contiguous_gradients=True,
-    ),
-)
-optim_wrapper = dict(
-    type='DeepSpeedOptimWrapper',
-    optimizer=dict(
-        type='AdamW',
-        lr=base_lr,
-        weight_decay=0.05
-    )
-)
-
-# runner_type = 'Runner'
-# optim_wrapper = dict(
-#     type='AmpOptimWrapper',
-#     dtype='float16',
-#     optimizer=dict(
-#         type='AdamW',
-#         lr=base_lr,
-#         weight_decay=0.05,
-#         eps=1e-8,
-#         betas=(0.9, 0.999))
-# )
-
 # Default setting for scaling LR automatically
 #   - `enable` means enable scaling LR automatically
 #       or not by default.

diff --git a/configs/rsprompter/_base_/samseg-maskrcnn.py b/configs/rsprompter/_base_/samseg-maskrcnn.py
@@ -290,15 +290,6 @@
 
 find_unused_parameters = True
 
-optim_wrapper = dict(
-    type='AmpOptimWrapper',
-    dtype='float16',
-    optimizer=dict(
-        type='AdamW',
-        lr=base_lr,
-        weight_decay=0.05
-    )
-)
 
 # Default setting for scaling LR automatically
 #   - `enable` means enable scaling LR automatically

diff --git a/configs/rsprompter/rsprompter_query-whu.py b/configs/rsprompter/rsprompter_query-whu.py
@@ -151,6 +151,7 @@
     ),
 )
 optim_wrapper = dict(
+    __delete__=True,
     type='DeepSpeedOptimWrapper',
     optimizer=dict(
         type='AdamW',