Add PyTorch checkpoint and NCNN INT8 model
Browse files- README.md +58 -0
- best.pt +3 -0
- ncnn/charset.json +0 -0
- ncnn/model.ncnn.bin +3 -0
- ncnn/model.ncnn.param +79 -0
README.md
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: pytorch
|
| 4 |
+
pipeline_tag: image-classification
|
| 5 |
+
tags:
|
| 6 |
+
- handwritten-chinese-character-recognition
|
| 7 |
+
- casia-hwdb
|
| 8 |
+
- mobilenet-v2
|
| 9 |
+
- ncnn
|
| 10 |
+
- int8
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# Handwritten
|
| 14 |
+
|
| 15 |
+
MobileNetV2 handwritten Chinese character recognizer trained on CASIA HWDB1.1.
|
| 16 |
+
The classifier covers 3,755 GB2312 level-1 Chinese characters with 64×64
|
| 17 |
+
grayscale input.
|
| 18 |
+
|
| 19 |
+
## Results
|
| 20 |
+
|
| 21 |
+
| Metric | Value |
|
| 22 |
+
|---|---:|
|
| 23 |
+
| PyTorch top-1 | 95.47% |
|
| 24 |
+
| PyTorch top-10 | 99.58% |
|
| 25 |
+
| Best epoch | 27 |
|
| 26 |
+
| NCNN INT8 size | 4.1 MB |
|
| 27 |
+
|
| 28 |
+
## Files
|
| 29 |
+
|
| 30 |
+
- `best.pt`: PyTorch model state, training configuration, epoch, and top-1 score.
|
| 31 |
+
- `ncnn/`: ready-to-run NCNN INT8 model and output charset.
|
| 32 |
+
|
| 33 |
+
The checkpoint SHA-256 is
|
| 34 |
+
`151f32b03c49dbed77f3b8076035d28ce050d70a1a102e12581274a44cb10460`.
|
| 35 |
+
|
| 36 |
+
## Load the checkpoint
|
| 37 |
+
|
| 38 |
+
Clone the source repository and place this checkpoint at
|
| 39 |
+
`runs/mbv2_phase1_aug/best.pt`:
|
| 40 |
+
|
| 41 |
+
```python
|
| 42 |
+
import torch
|
| 43 |
+
|
| 44 |
+
from src.models import build_model
|
| 45 |
+
|
| 46 |
+
checkpoint = torch.load("best.pt", map_location="cpu", weights_only=False)
|
| 47 |
+
model = build_model(checkpoint["cfg"]["model"], num_classes=3755)
|
| 48 |
+
model.load_state_dict(checkpoint["model"])
|
| 49 |
+
model.eval()
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
Source and full training/deployment instructions:
|
| 53 |
+
https://github.com/Ismantic/Handwritten
|
| 54 |
+
|
| 55 |
+
## Data
|
| 56 |
+
|
| 57 |
+
CASIA HWDB1.1 is not redistributed in this repository. Obtain it through the
|
| 58 |
+
official CASIA channel and comply with its license.
|
best.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:151f32b03c49dbed77f3b8076035d28ce050d70a1a102e12581274a44cb10460
|
| 3 |
+
size 16875467
|
ncnn/charset.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
ncnn/model.ncnn.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:86a4802e3c8b2ddf1180f3de361546517d708674381e58b3d10b766dc8fba49e
|
| 3 |
+
size 4288512
|
ncnn/model.ncnn.param
ADDED
|
@@ -0,0 +1,79 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
7767517
|
| 2 |
+
77 87
|
| 3 |
+
Input in0 0 1 in0
|
| 4 |
+
Convolution convclip_0 1 1 in0 1 0=32 1=3 3=2 4=1 5=1 6=288 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 5 |
+
ConvolutionDepthWise convdwclip_0 1 1 1 2 0=32 1=3 4=1 5=1 6=288 7=32 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 6 |
+
Convolution conv_2 1 1 2 3 0=16 1=1 5=1 6=512 8=102
|
| 7 |
+
Convolution convclip_1 1 1 3 4 0=96 1=1 5=1 6=1536 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 8 |
+
ConvolutionDepthWise convdwclip_1 1 1 4 5 0=96 1=3 3=2 4=1 5=1 6=864 7=96 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 9 |
+
Convolution conv_4 1 1 5 6 0=24 1=1 5=1 6=2304 8=2
|
| 10 |
+
Split splitncnn_0 1 2 6 7 8
|
| 11 |
+
Convolution convclip_2 1 1 8 9 0=144 1=1 5=1 6=3456 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 12 |
+
ConvolutionDepthWise convdwclip_2 1 1 9 10 0=144 1=3 4=1 5=1 6=1296 7=144 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 13 |
+
Convolution conv_6 1 1 10 11 0=24 1=1 5=1 6=3456 8=2
|
| 14 |
+
BinaryOp add_0 2 1 7 11 12
|
| 15 |
+
Convolution convclip_3 1 1 12 13 0=144 1=1 5=1 6=3456 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 16 |
+
ConvolutionDepthWise convdwclip_3 1 1 13 14 0=144 1=3 3=2 4=1 5=1 6=1296 7=144 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 17 |
+
Convolution conv_8 1 1 14 15 0=32 1=1 5=1 6=4608 8=2
|
| 18 |
+
Split splitncnn_1 1 2 15 16 17
|
| 19 |
+
Convolution convclip_4 1 1 17 18 0=192 1=1 5=1 6=6144 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 20 |
+
ConvolutionDepthWise convdwclip_4 1 1 18 19 0=192 1=3 4=1 5=1 6=1728 7=192 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 21 |
+
Convolution conv_10 1 1 19 20 0=32 1=1 5=1 6=6144 8=2
|
| 22 |
+
BinaryOp add_1 2 1 16 20 21
|
| 23 |
+
Split splitncnn_2 1 2 21 22 23
|
| 24 |
+
Convolution convclip_5 1 1 23 24 0=192 1=1 5=1 6=6144 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 25 |
+
ConvolutionDepthWise convdwclip_5 1 1 24 25 0=192 1=3 4=1 5=1 6=1728 7=192 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 26 |
+
Convolution conv_12 1 1 25 26 0=32 1=1 5=1 6=6144 8=2
|
| 27 |
+
BinaryOp add_2 2 1 22 26 27
|
| 28 |
+
Convolution convclip_6 1 1 27 28 0=192 1=1 5=1 6=6144 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 29 |
+
ConvolutionDepthWise convdwclip_6 1 1 28 29 0=192 1=3 3=2 4=1 5=1 6=1728 7=192 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 30 |
+
Convolution conv_14 1 1 29 30 0=64 1=1 5=1 6=12288 8=2
|
| 31 |
+
Split splitncnn_3 1 2 30 31 32
|
| 32 |
+
Convolution convclip_7 1 1 32 33 0=384 1=1 5=1 6=24576 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 33 |
+
ConvolutionDepthWise convdwclip_7 1 1 33 34 0=384 1=3 4=1 5=1 6=3456 7=384 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 34 |
+
Convolution conv_16 1 1 34 35 0=64 1=1 5=1 6=24576 8=2
|
| 35 |
+
BinaryOp add_3 2 1 31 35 36
|
| 36 |
+
Split splitncnn_4 1 2 36 37 38
|
| 37 |
+
Convolution convclip_8 1 1 38 39 0=384 1=1 5=1 6=24576 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 38 |
+
ConvolutionDepthWise convdwclip_8 1 1 39 40 0=384 1=3 4=1 5=1 6=3456 7=384 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 39 |
+
Convolution conv_18 1 1 40 41 0=64 1=1 5=1 6=24576 8=2
|
| 40 |
+
BinaryOp add_4 2 1 37 41 42
|
| 41 |
+
Split splitncnn_5 1 2 42 43 44
|
| 42 |
+
Convolution convclip_9 1 1 44 45 0=384 1=1 5=1 6=24576 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 43 |
+
ConvolutionDepthWise convdwclip_9 1 1 45 46 0=384 1=3 4=1 5=1 6=3456 7=384 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 44 |
+
Convolution conv_20 1 1 46 47 0=64 1=1 5=1 6=24576 8=2
|
| 45 |
+
BinaryOp add_5 2 1 43 47 48
|
| 46 |
+
Convolution convclip_10 1 1 48 49 0=384 1=1 5=1 6=24576 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 47 |
+
ConvolutionDepthWise convdwclip_10 1 1 49 50 0=384 1=3 4=1 5=1 6=3456 7=384 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 48 |
+
Convolution conv_22 1 1 50 51 0=96 1=1 5=1 6=36864 8=2
|
| 49 |
+
Split splitncnn_6 1 2 51 52 53
|
| 50 |
+
Convolution convclip_11 1 1 53 54 0=576 1=1 5=1 6=55296 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 51 |
+
ConvolutionDepthWise convdwclip_11 1 1 54 55 0=576 1=3 4=1 5=1 6=5184 7=576 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 52 |
+
Convolution conv_24 1 1 55 56 0=96 1=1 5=1 6=55296 8=2
|
| 53 |
+
BinaryOp add_6 2 1 52 56 57
|
| 54 |
+
Split splitncnn_7 1 2 57 58 59
|
| 55 |
+
Convolution convclip_12 1 1 59 60 0=576 1=1 5=1 6=55296 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 56 |
+
ConvolutionDepthWise convdwclip_12 1 1 60 61 0=576 1=3 4=1 5=1 6=5184 7=576 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 57 |
+
Convolution conv_26 1 1 61 62 0=96 1=1 5=1 6=55296 8=2
|
| 58 |
+
BinaryOp add_7 2 1 58 62 63
|
| 59 |
+
Convolution convclip_13 1 1 63 64 0=576 1=1 5=1 6=55296 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 60 |
+
ConvolutionDepthWise convdwclip_13 1 1 64 65 0=576 1=3 3=2 4=1 5=1 6=5184 7=576 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 61 |
+
Convolution conv_28 1 1 65 66 0=160 1=1 5=1 6=92160 8=2
|
| 62 |
+
Split splitncnn_8 1 2 66 67 68
|
| 63 |
+
Convolution convclip_14 1 1 68 69 0=960 1=1 5=1 6=153600 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 64 |
+
ConvolutionDepthWise convdwclip_14 1 1 69 70 0=960 1=3 4=1 5=1 6=8640 7=960 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 65 |
+
Convolution conv_30 1 1 70 71 0=160 1=1 5=1 6=153600 8=2
|
| 66 |
+
BinaryOp add_8 2 1 67 71 72
|
| 67 |
+
Split splitncnn_9 1 2 72 73 74
|
| 68 |
+
Convolution convclip_15 1 1 74 75 0=960 1=1 5=1 6=153600 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 69 |
+
ConvolutionDepthWise convdwclip_15 1 1 75 76 0=960 1=3 4=1 5=1 6=8640 7=960 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 70 |
+
Convolution conv_32 1 1 76 77 0=160 1=1 5=1 6=153600 8=2
|
| 71 |
+
BinaryOp add_9 2 1 73 77 78
|
| 72 |
+
Convolution convclip_16 1 1 78 79 0=960 1=1 5=1 6=153600 8=102 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 73 |
+
ConvolutionDepthWise convdwclip_16 1 1 79 80 0=960 1=3 4=1 5=1 6=8640 7=960 8=101 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 74 |
+
Convolution conv_34 1 1 80 81 0=320 1=1 5=1 6=307200 8=102
|
| 75 |
+
Convolution convclip_17 1 1 81 82 0=576 1=1 5=1 6=184320 8=2 9=3 -23310=2,0.000000e+00,6.000000e+00
|
| 76 |
+
Pooling gap_0 1 1 82 83 0=1 4=1
|
| 77 |
+
Reshape reshape_72 1 1 83 84 0=1 1=1 2=-1
|
| 78 |
+
Flatten flatten_73 1 1 84 85
|
| 79 |
+
InnerProduct linear_36 1 1 85 out0 0=3755 1=1 2=2162880 8=2
|