mirrors
/
PaddleOCR
mirror of https://github.com/PaddlePaddle/PaddleOCR.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

[TIPC]update tipc scripts and rm fluid api (#11098) 

* [TIPC]update xpu tipc script

* update tipc scripts and remove fluid api
pull/11144/head
Kai Song 2023-10-19 10:45:10 +08:00 committed by GitHub
parent 673c95262b
commit 344b7594e4
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 72 additions and 56 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

84
ppocr/modeling/heads/rec_robustscanner_head.py
View File

@ -11,7 +11,6 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
""" """
This code is refer from: This code is refer from:
https://github.com/open-mmlab/mmocr/blob/main/mmocr/models/textrecog/encoders/channel_reduction_encoder.py https://github.com/open-mmlab/mmocr/blob/main/mmocr/models/textrecog/encoders/channel_reduction_encoder.py
@ -28,6 +27,7 @@ from paddle import ParamAttr
import paddle.nn as nn import paddle.nn as nn
import paddle.nn.functional as F import paddle.nn.functional as F
class BaseDecoder(nn.Layer): class BaseDecoder(nn.Layer):
def __init__(self, **kwargs): def __init__(self, **kwargs):
super().__init__() super().__init__()
@ -48,9 +48,11 @@ class BaseDecoder(nn.Layer):
self.train_mode = train_mode self.train_mode = train_mode
if train_mode: if train_mode:
return self.forward_train(feat, out_enc, label, valid_ratios, word_positions) return self.forward_train(feat, out_enc, label, valid_ratios,
word_positions)
return self.forward_test(feat, out_enc, valid_ratios, word_positions) return self.forward_test(feat, out_enc, valid_ratios, word_positions)
class ChannelReductionEncoder(nn.Layer): class ChannelReductionEncoder(nn.Layer):
"""Change the channel number with a one by one convoluational layer. """Change the channel number with a one by one convoluational layer.
@ -59,14 +61,16 @@ class ChannelReductionEncoder(nn.Layer):
out_channels (int): Number of output channels. out_channels (int): Number of output channels.
""" """
def __init__(self, def __init__(self, in_channels, out_channels, **kwargs):
in_channels,
out_channels,
**kwargs):
super(ChannelReductionEncoder, self).__init__() super(ChannelReductionEncoder, self).__init__()
self.layer = nn.Conv2D( self.layer = nn.Conv2D(
in_channels, out_channels, kernel_size=1, stride=1, padding=0, weight_attr=nn.initializer.XavierNormal()) in_channels,
out_channels,
kernel_size=1,
stride=1,
padding=0,
weight_attr=nn.initializer.XavierNormal())
def forward(self, feat): def forward(self, feat):
""" """
@ -84,8 +88,8 @@ def masked_fill(x, mask, value):
y = paddle.full(x.shape, value, x.dtype) y = paddle.full(x.shape, value, x.dtype)
return paddle.where(mask, y, x) return paddle.where(mask, y, x)
class DotProductAttentionLayer(nn.Layer):
class DotProductAttentionLayer(nn.Layer):
def __init__(self, dim_model=None): def __init__(self, dim_model=None):
super().__init__() super().__init__()
@ -99,7 +103,7 @@ class DotProductAttentionLayer(nn.Layer):
logits = paddle.reshape(logits, [n, c, h, w]) logits = paddle.reshape(logits, [n, c, h, w])
if valid_ratios is not None: if valid_ratios is not None:
# cal mask of attention weight # cal mask of attention weight
with paddle.fluid.framework._stride_in_no_check_dy2st_diff(): with paddle.base.framework._stride_in_no_check_dy2st_diff():
for i, valid_ratio in enumerate(valid_ratios): for i, valid_ratio in enumerate(valid_ratios):
valid_width = min(w, int(w * valid_ratio + 0.5)) valid_width = min(w, int(w * valid_ratio + 0.5))
if valid_width < w: if valid_width < w:
@ -113,6 +117,7 @@ class DotProductAttentionLayer(nn.Layer):
glimpse = paddle.transpose(glimpse, (0, 2, 1)) glimpse = paddle.transpose(glimpse, (0, 2, 1))
return glimpse return glimpse
class SequenceAttentionDecoder(BaseDecoder): class SequenceAttentionDecoder(BaseDecoder):
"""Sequence attention decoder for RobustScanner. """Sequence attention decoder for RobustScanner.
@ -181,8 +186,8 @@ class SequenceAttentionDecoder(BaseDecoder):
self.prediction = None self.prediction = None
if not self.return_feature: if not self.return_feature:
pred_num_classes = num_classes - 1 pred_num_classes = num_classes - 1
self.prediction = nn.Linear( self.prediction = nn.Linear(dim_model if encode_value else
dim_model if encode_value else dim_input, pred_num_classes) dim_input, pred_num_classes)
def forward_train(self, feat, out_enc, targets, valid_ratios): def forward_train(self, feat, out_enc, targets, valid_ratios):
""" """
@ -243,12 +248,13 @@ class SequenceAttentionDecoder(BaseDecoder):
seq_len = self.max_seq_len seq_len = self.max_seq_len
batch_size = feat.shape[0] batch_size = feat.shape[0]
decode_sequence = (paddle.ones((batch_size, seq_len), dtype='int64') * self.start_idx) decode_sequence = (paddle.ones(
(batch_size, seq_len), dtype='int64') * self.start_idx)
outputs = [] outputs = []
for i in range(seq_len): for i in range(seq_len):
step_out = self.forward_test_step(feat, out_enc, decode_sequence, step_out = self.forward_test_step(feat, out_enc, decode_sequence, i,
i, valid_ratios) valid_ratios)
outputs.append(step_out) outputs.append(step_out)
max_idx = paddle.argmax(step_out, axis=1, keepdim=False) max_idx = paddle.argmax(step_out, axis=1, keepdim=False)
if i < seq_len - 1: if i < seq_len - 1:
@ -306,7 +312,6 @@ class SequenceAttentionDecoder(BaseDecoder):
class PositionAwareLayer(nn.Layer): class PositionAwareLayer(nn.Layer):
def __init__(self, dim_model, rnn_layers=2): def __init__(self, dim_model, rnn_layers=2):
super().__init__() super().__init__()
@ -384,16 +389,16 @@ class PositionAttentionDecoder(BaseDecoder):
self.embedding = nn.Embedding(self.max_seq_len + 1, self.dim_model) self.embedding = nn.Embedding(self.max_seq_len + 1, self.dim_model)
self.position_aware_module = PositionAwareLayer( self.position_aware_module = PositionAwareLayer(self.dim_model,
self.dim_model, rnn_layers) rnn_layers)
self.attention_layer = DotProductAttentionLayer() self.attention_layer = DotProductAttentionLayer()
self.prediction = None self.prediction = None
if not self.return_feature: if not self.return_feature:
pred_num_classes = num_classes - 1 pred_num_classes = num_classes - 1
self.prediction = nn.Linear( self.prediction = nn.Linear(dim_model if encode_value else
dim_model if encode_value else dim_input, pred_num_classes) dim_input, pred_num_classes)
def _get_position_index(self, length, batch_size): def _get_position_index(self, length, batch_size):
position_index_list = [] position_index_list = []
@ -403,7 +408,8 @@ class PositionAttentionDecoder(BaseDecoder):
batch_position_index = paddle.stack(position_index_list, axis=0) batch_position_index = paddle.stack(position_index_list, axis=0)
return batch_position_index return batch_position_index
def forward_train(self, feat, out_enc, targets, valid_ratios, position_index): def forward_train(self, feat, out_enc, targets, valid_ratios,
position_index):
""" """
Args: Args:
feat (Tensor): Tensor of shape :math:`(N, D_i, H, W)`. feat (Tensor): Tensor of shape :math:`(N, D_i, H, W)`.
@ -434,9 +440,9 @@ class PositionAttentionDecoder(BaseDecoder):
query = paddle.transpose(query, (0, 2, 1)) query = paddle.transpose(query, (0, 2, 1))
key = paddle.reshape(position_out_enc, (n, c_enc, h * w)) key = paddle.reshape(position_out_enc, (n, c_enc, h * w))
if self.encode_value: if self.encode_value:
value = paddle.reshape(out_enc,(n, c_enc, h * w)) value = paddle.reshape(out_enc, (n, c_enc, h * w))
else: else:
value = paddle.reshape(feat,(n, c_feat, h * w)) value = paddle.reshape(feat, (n, c_feat, h * w))
attn_out = self.attention_layer(query, key, value, h, w, valid_ratios) attn_out = self.attention_layer(query, key, value, h, w, valid_ratios)
attn_out = paddle.transpose(attn_out, (0, 2, 1)) # [n, len_q, dim_v] attn_out = paddle.transpose(attn_out, (0, 2, 1)) # [n, len_q, dim_v]
@ -472,9 +478,9 @@ class PositionAttentionDecoder(BaseDecoder):
query = paddle.transpose(query, (0, 2, 1)) query = paddle.transpose(query, (0, 2, 1))
key = paddle.reshape(position_out_enc, (n, c_enc, h * w)) key = paddle.reshape(position_out_enc, (n, c_enc, h * w))
if self.encode_value: if self.encode_value:
value = paddle.reshape(out_enc,(n, c_enc, h * w)) value = paddle.reshape(out_enc, (n, c_enc, h * w))
else: else:
value = paddle.reshape(feat,(n, c_feat, h * w)) value = paddle.reshape(feat, (n, c_feat, h * w))
attn_out = self.attention_layer(query, key, value, h, w, valid_ratios) attn_out = self.attention_layer(query, key, value, h, w, valid_ratios)
attn_out = paddle.transpose(attn_out, (0, 2, 1)) # [n, len_q, dim_v] attn_out = paddle.transpose(attn_out, (0, 2, 1)) # [n, len_q, dim_v]
@ -484,8 +490,8 @@ class PositionAttentionDecoder(BaseDecoder):
return self.prediction(attn_out) return self.prediction(attn_out)
class RobustScannerFusionLayer(nn.Layer):
class RobustScannerFusionLayer(nn.Layer):
def __init__(self, dim_model, dim=-1): def __init__(self, dim_model, dim=-1):
super(RobustScannerFusionLayer, self).__init__() super(RobustScannerFusionLayer, self).__init__()
@ -500,6 +506,7 @@ class RobustScannerFusionLayer(nn.Layer):
output = F.glu(output, self.dim) output = F.glu(output, self.dim)
return output return output
class RobustScannerDecoder(BaseDecoder): class RobustScannerDecoder(BaseDecoder):
"""Decoder for RobustScanner. """Decoder for RobustScanner.
@ -561,8 +568,7 @@ class RobustScannerDecoder(BaseDecoder):
padding_idx=padding_idx, padding_idx=padding_idx,
dropout=hybrid_decoder_dropout, dropout=hybrid_decoder_dropout,
encode_value=encode_value, encode_value=encode_value,
return_feature=True return_feature=True)
)
# init position decoder # init position decoder
self.position_decoder = PositionAttentionDecoder( self.position_decoder = PositionAttentionDecoder(
@ -573,9 +579,7 @@ class RobustScannerDecoder(BaseDecoder):
max_seq_len=max_seq_len, max_seq_len=max_seq_len,
mask=mask, mask=mask,
encode_value=encode_value, encode_value=encode_value,
return_feature=True return_feature=True)
)
self.fusion_module = RobustScannerFusionLayer( self.fusion_module = RobustScannerFusionLayer(
self.dim_model if encode_value else dim_input) self.dim_model if encode_value else dim_input)
@ -584,7 +588,8 @@ class RobustScannerDecoder(BaseDecoder):
self.prediction = nn.Linear(dim_model if encode_value else dim_input, self.prediction = nn.Linear(dim_model if encode_value else dim_input,
pred_num_classes) pred_num_classes)
def forward_train(self, feat, out_enc, target, valid_ratios, word_positions): def forward_train(self, feat, out_enc, target, valid_ratios,
word_positions):
""" """
Args: Args:
feat (Tensor): Tensor of shape :math:`(N, D_i, H, W)`. feat (Tensor): Tensor of shape :math:`(N, D_i, H, W)`.
@ -599,8 +604,8 @@ class RobustScannerDecoder(BaseDecoder):
Returns: Returns:
Tensor: A raw logit tensor of shape :math:`(N, T, C-1)`. Tensor: A raw logit tensor of shape :math:`(N, T, C-1)`.
""" """
hybrid_glimpse = self.hybrid_decoder.forward_train( hybrid_glimpse = self.hybrid_decoder.forward_train(feat, out_enc,
feat, out_enc, target, valid_ratios) target, valid_ratios)
position_glimpse = self.position_decoder.forward_train( position_glimpse = self.position_decoder.forward_train(
feat, out_enc, target, valid_ratios, word_positions) feat, out_enc, target, valid_ratios, word_positions)
@ -625,7 +630,8 @@ class RobustScannerDecoder(BaseDecoder):
seq_len = self.max_seq_len seq_len = self.max_seq_len
batch_size = feat.shape[0] batch_size = feat.shape[0]
decode_sequence = (paddle.ones((batch_size, seq_len), dtype='int64') * self.start_idx) decode_sequence = (paddle.ones(
(batch_size, seq_len), dtype='int64') * self.start_idx)
position_glimpse = self.position_decoder.forward_test( position_glimpse = self.position_decoder.forward_test(
feat, out_enc, valid_ratios, word_positions) feat, out_enc, valid_ratios, word_positions)
@ -649,8 +655,10 @@ class RobustScannerDecoder(BaseDecoder):
return outputs return outputs
class RobustScannerHead(nn.Layer): class RobustScannerHead(nn.Layer):
def __init__(self, def __init__(
self,
out_channels, # 90 + unknown + start + padding out_channels, # 90 + unknown + start + padding
in_channels, in_channels,
enc_outchannles=128, enc_outchannles=128,
@ -670,7 +678,7 @@ class RobustScannerHead(nn.Layer):
in_channels=in_channels, out_channels=enc_outchannles) in_channels=in_channels, out_channels=enc_outchannles)
# decoder module # decoder module
self.decoder =RobustScannerDecoder( self.decoder = RobustScannerDecoder(
num_classes=out_channels, num_classes=out_channels,
dim_input=in_channels, dim_input=in_channels,
dim_model=enc_outchannles, dim_model=enc_outchannles,
@ -697,8 +705,8 @@ class RobustScannerHead(nn.Layer):
if self.training: if self.training:
label = targets[0] # label label = targets[0] # label
label = paddle.to_tensor(label, dtype='int64') label = paddle.to_tensor(label, dtype='int64')
final_out = self.decoder( final_out = self.decoder(inputs, out_enc, label, valid_ratios,
inputs, out_enc, label, valid_ratios, word_positions) word_positions)
if not self.training: if not self.training:
final_out = self.decoder( final_out = self.decoder(
inputs, inputs,

4
test_tipc/test_train_inference_python_npu.sh
View File

@ -43,6 +43,10 @@ if [ $modelname == "rec_r31_sar" ] || [ $modelname == "rec_mtb_nrtr" ]; then
sed -i "s/gpu_list:0|0,1/gpu_list:0,1/g" $FILENAME sed -i "s/gpu_list:0|0,1/gpu_list:0,1/g" $FILENAME
sed -i "s/Global.use_npu:True|True/Global.use_npu:True/g" $FILENAME sed -i "s/Global.use_npu:True|True/Global.use_npu:True/g" $FILENAME
fi fi
if [ $modelname == "ch_ppocr_mobile_v2_0_rec_FPGM" ]; then
sed -i '18s/$/ -o Global.use_gpu=False/' $FILENAME
sed -i '32s/$/ Global.use_gpu=False/' $FILENAME
fi
# replace training config file # replace training config file
grep -n 'tools/.*yml' $FILENAME | cut -d ":" -f 1 \ grep -n 'tools/.*yml' $FILENAME | cut -d ":" -f 1 \

4
test_tipc/test_train_inference_python_xpu.sh
View File

@ -43,6 +43,10 @@ if [ $modelname == "rec_r31_sar" ] || [ $modelname == "rec_mtb_nrtr" ]; then
sed -i "s/gpu_list:0|0,1/gpu_list:0,1/g" $FILENAME sed -i "s/gpu_list:0|0,1/gpu_list:0,1/g" $FILENAME
sed -i "s/Global.use_xpu:True|True/Global.use_xpu:True/g" $FILENAME sed -i "s/Global.use_xpu:True|True/Global.use_xpu:True/g" $FILENAME
fi fi
if [ $modelname == "ch_ppocr_mobile_v2_0_rec_FPGM" ]; then
sed -i '18s/$/ -o Global.use_gpu=False/' $FILENAME
sed -i '32s/$/ Global.use_gpu=False/' $FILENAME
fi
# replace training config file # replace training config file
grep -n 'tools/.*yml' $FILENAME | cut -d ":" -f 1 \ grep -n 'tools/.*yml' $FILENAME | cut -d ":" -f 1 \