update params and fix CI

mindspore-lab · Mar 1, 2024 · 8bb78df · 8bb78df
1 parent d398539
commit 8bb78df
Show file tree

Hide file tree

Showing 4 changed files with 15 additions and 10 deletions.
diff --git a/configs/navit/navit_b16_384_ascend.yaml b/configs/navit/navit_b16_384_ascend.yaml
@@ -11,11 +11,11 @@ dataset: "imagenet"
 data_dir: "/path/to/imagenet"
 shuffle: True
 dataset_download: False
-batch_size: 12
+batch_size: 64
 drop_remainder: True
 patch_size: 16
-max_seq_length: 2048
-max_num_each_group: 40
+max_seq_length: 768
+max_num_each_group: 16
 
 # augmentation
 image_resize: 384
@@ -29,7 +29,7 @@ drop_path_rate: 0.1
 num_classes: 1000
 pretrained: False
 ckpt_path: ""
-keep_checkpoint_max: 3
+keep_checkpoint_max: 1
 ckpt_save_policy: "top_k"
 ckpt_save_dir: "./ckpt"
 epoch_size: 100

diff --git a/configs/navit/navit_b16_384_ascend_control.yaml b/configs/navit/navit_b16_384_ascend_control.yaml
@@ -11,7 +11,7 @@ dataset: "imagenet"
 data_dir: "/path/to/imagenet"
 shuffle: True
 dataset_download: False
-batch_size: 64
+batch_size: 116
 drop_remainder: True
 
 # augmentation
@@ -26,7 +26,7 @@ drop_path_rate: 0.1
 num_classes: 1000
 pretrained: False
 ckpt_path: ""
-keep_checkpoint_max: 3
+keep_checkpoint_max: 1
 ckpt_save_policy: "top_k"
 ckpt_save_dir: "./ckpt"
 epoch_size: 100

diff --git a/mindcv/models/navit.py b/mindcv/models/navit.py
@@ -139,7 +139,8 @@ def construct(self, x, context=None, token_mask=None):
             token_mask = ops.unsqueeze(token_mask, 1)
             attn = ops.masked_fill(attn, ~token_mask, -ms.numpy.inf)
 
-        attn = ops.softmax(attn, axis=-1)
+        dtype = attn.dtype
+        attn = ops.softmax(attn.to(ms.float32), axis=-1).to(dtype)
         attn = self.attn_drop(attn)
 
         out = self.attn_matmul_v(attn, v)

diff --git a/mindcv/utils/top_k.py b/mindcv/utils/top_k.py
@@ -1,7 +1,11 @@
 import numpy as np
 
-from mindspore.train.metrics.metric import _check_onehot_data, rearrange_inputs
-from mindspore.train.metrics.topk import TopKCategoricalAccuracy
+try:
+    from mindspore.train.metrics.metric import rearrange_inputs
+    from mindspore.train.metrics.topk import TopKCategoricalAccuracy
+except ImportError:  # MS Version < 2.0
+    from mindspore.nn.metrics.metric import rearrange_inputs
+    from mindspore.nn.metrics.topk import TopKCategoricalAccuracy
 
 
 class TopKCategoricalAccuracyForTokenData(TopKCategoricalAccuracy):
@@ -23,7 +27,7 @@ def update(self, *inputs):
         y_pred = y_pred[inds]
         y = y[inds]
 
-        if y_pred.ndim == y.ndim and _check_onehot_data(y):
+        if y_pred.ndim == y.ndim:
             y = y.argmax(axis=1)
         indices = np.argsort(-y_pred, axis=1)[:, : self.k]
         repeated_y = y.reshape(-1, 1).repeat(self.k, axis=1)