Added t2t-vit-14 model (#2967)

Anna Grebneva · web-flow · commit fe52bbc117de · 2022-01-27T23:40:35.000+03:00
diff --git a/demos/classification_demo/python/README.md b/demos/classification_demo/python/README.md
@@ -78,6 +78,7 @@ omz_converter --list models.lst
 * squeezenet1.0
 * squeezenet1.1
 * swin-tiny-patch4-window7-224
+* t2t-vit-14
 * vgg16
 * vgg19
 
diff --git a/demos/classification_demo/python/models.lst b/demos/classification_demo/python/models.lst
@@ -46,5 +46,6 @@ shufflenet-v2-x1.0
 squeezenet1.0
 squeezenet1.1
 swin-tiny-patch4-window7-224
+t2t-vit-14
 vgg16
 vgg19
diff --git a/models/public/device_support.md b/models/public/device_support.md
@@ -119,6 +119,7 @@
 | ssd_mobilenet_v1_fpn_coco | YES | YES | YES |
 | ssdlite_mobilenet_v2 | YES | YES | YES |
 | swin-tiny-patch4-window7-224 | YES | YES | YES |
+| t2t-vit-14 | YES | | |
 | text-recognition-resnet-fc | YES | YES | YES |
 | ultra-lightweight-face-detection-rfb-320 | YES | YES | YES |
 | ultra-lightweight-face-detection-slim-320 | YES | YES | YES |
diff --git a/models/public/index.md b/models/public/index.md
@@ -53,6 +53,7 @@ You can download models and convert them into Inference Engine format (\*.xml +
 | SqueezeNet v1.0             | Caffe\*                            | [squeezenet1.0](./squeezenet1.0/README.md)| 57.684%/80.38%| 1.737 | 1.248 |
 | SqueezeNet v1.1             | Caffe\*| [squeezenet1.1](./squeezenet1.1/README.md)| 58.382%/81% | 0.785 | 1.236 |
 | Swin Transformer Tiny, window size=7| PyTorch\*                  | [swin-tiny-patch4-window7-224](./swin-tiny-patch4-window7-224/README.md) | 81.38%/95.51% | 9.0280 | 28.8173 |
+| T2T-ViT, transformer layers number=14| PyTorch\*                 | [t2t-vit-14](./t2t-vit-14/README.md) | 81.44%/95.66% | 9.5451 | 21.5498 |
 | VGG 16                      | Caffe\*                            | [vgg16](./vgg16/README.md) | 70.968%/89.878% | 30.974 | 138.358 |
 | VGG 19                      | Caffe\*              | [vgg19](./vgg19/README.md) | 71.062%/89.832% | 39.3 | 143.667  |
 
diff --git a/models/public/t2t-vit-14/README.md b/models/public/t2t-vit-14/README.md
@@ -0,0 +1,98 @@
+# t2t-vit-14
+
+## Use Case and High-Level Description
+
+The `t2t-vit-14` model is a variant of the Tokens-To-Token Vision Transformer(T2T-ViT) pre-trained on ImageNet dataset for image classification task. T2T-ViT progressively tokenize the image to tokens and has an efficient backbone. T2T-ViT consists of two main components: 1) a layer-wise "Tokens-to-Token module" to model the local structure information of the image and reduce the length of tokens progressively; 2) an efficient "T2T-ViT backbone" to draw the global attention relation on tokens from the T2T module. The model has 14 transformer layers in T2T-ViT backbone with 384 hidden dimensions.
+
+More details provided in the [paper](https://arxiv.org/abs/2101.11986) and [repository](https://github.com/yitu-opensource/T2T-ViT).
+
+## Specification
+
+| Metric                          | Value          |
+|---------------------------------|----------------|
+| Type                            | Classification |
+| GFlops                          | 9.5451         |
+| MParams                         | 21.5498        |
+| Source framework                | PyTorch\*      |
+
+## Accuracy
+
+| Metric | Value  |
+| ------ | ------ |
+| Top 1  | 81.44% |
+| Top 5  | 95.66% |
+
+## Input
+
+### Original Model
+
+Image, name: `image`, shape: `1, 3, 224, 224`, format: `B, C, H, W`, where:
+
+- `B` - batch size
+- `C` - number of channels
+- `H` - image height
+- `W` - image width
+
+Expected color order: `RGB`.
+Mean values - [123.675, 116.28, 103.53], scale values - [58.395, 57.12, 57.375].
+
+### Converted Model
+
+Image, name: `image`, shape: `1, 3, 224, 224`, format: `B, C, H, W`, where:
+
+- `B` - batch size
+- `C` - number of channels
+- `H` - image height
+- `W` - image width
+
+Expected color order: `BGR`.
+
+## Output
+
+### Original Model
+
+Object classifier according to ImageNet classes, name: `probs`,  shape: `1, 1000`, output data format is `B, C`, where:
+
+- `B` - batch size
+- `C` - vector of probabilities for all dataset classes in logits format
+
+### Converted Model
+
+Object classifier according to ImageNet classes, name: `probs`,  shape: `1, 1000`, output data format is `B, C`, where:
+
+- `B` - batch size
+- `C` - vector of probabilities for all dataset classes in logits format
+
+## Download a Model and Convert it into Inference Engine Format
+
+You can download models and if necessary convert them into Inference Engine format using the [Model Downloader and other automation tools](../../../tools/model_tools/README.md) as shown in the examples below.
+
+An example of using the Model Downloader:
+```
+omz_downloader --name <model_name>
+```
+
+An example of using the Model Converter:
+```
+omz_converter --name <model_name>
+```
+
+## Legal Information
+
+The original model is distributed under the following
+[license](https://raw.githubusercontent.com/yitu-opensource/T2T-ViT/main/LICENSE):
+
+```
+The Clear BSD License
+
+Copyright (c) [2012]-[2021] Shanghai Yitu Technology Co., Ltd.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification, are permitted (subject to the limitations in the disclaimer below) provided that the following conditions are met:
+
+* Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
+* Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
+* Neither the name of Shanghai Yitu Technology Co., Ltd. nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
+
+NO EXPRESS OR IMPLIED LICENSES TO ANY PARTY'S PATENT RIGHTS ARE GRANTED BY THIS LICENSE. THIS SOFTWARE IS PROVIDED BY SHANGHAI YITU TECHNOLOGY CO., LTD. AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SHANGHAI YITU TECHNOLOGY CO., LTD. OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+```
diff --git a/models/public/t2t-vit-14/accuracy-check.yml b/models/public/t2t-vit-14/accuracy-check.yml
@@ -0,0 +1,31 @@
+models:
+  - name: t2t-vit-14
+
+    launchers:
+      - framework: dlsdk
+        adapter: classification
+
+    datasets:
+      - name: imagenet_1000_classes
+        reader: pillow_imread
+
+        preprocessing:
+          - type: resize
+            size: 256
+            aspect_ratio_scale: greater
+            use_pillow: True
+            interpolation: BICUBIC
+          - type: crop
+            size: 224
+            use_pillow: True
+          - type: rgb_to_bgr
+
+        metrics:
+          - name: accuracy@top1
+            type: accuracy
+            top_k: 1
+            reference: 0.8144
+          - name: accuracy@top5
+            type: accuracy
+            top_k: 5
+            reference: 0.9566
diff --git a/models/public/t2t-vit-14/model.py b/models/public/t2t-vit-14/model.py
@@ -0,0 +1,25 @@
+# Copyright (c) 2022 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from torch import load
+from models.t2t_vit import t2t_vit_14
+
+
+def create_model(weights):
+    model = t2t_vit_14()
+
+    checkpoint = load(weights, map_location='cpu')['state_dict_ema']
+    model.load_state_dict(checkpoint)
+
+    return model
diff --git a/models/public/t2t-vit-14/model.yml b/models/public/t2t-vit-14/model.yml
@@ -0,0 +1,77 @@
+# Copyright (c) 2022 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+description: >-
+  The "t2t-vit-14" model is a variant of the Tokens-To-Token Vision Transformer(T2T-ViT)
+  pre-trained on ImageNet dataset for image classification task. T2T-ViT progressively
+  tokenize the image to tokens and has an efficient backbone. T2T-ViT consists of
+  two main components: 1) a layer-wise "Tokens-to-Token module" to model the local
+  structure information of the image and reduce the length of tokens progressively;
+  2) an efficient "T2T-ViT backbone" to draw the global attention relation on tokens
+  from the T2T module. The model has 14 transformer layers in T2T-ViT backbone with
+  384 hidden dimensions.
+
+  More details provided in the paper <https://arxiv.org/abs/2101.11986> and repository
+  <https://github.com/yitu-opensource/T2T-ViT>.
+task_type: classification
+files:
+  - name: timm-0.5.4-py3-none-any.whl
+    size: 431537
+    checksum: e8f1967a8e2029fe21a43875132b4b123227b718abc35725d7f2b9fd0ef2062884ac3dd558570b51a780aad89bc375d6
+    source: https://files.pythonhosted.org/packages/49/65/a83208746dc9c0d70feff7874b49780ff110810feb528df4b0ecadcbee60/timm-0.5.4-py3-none-any.whl
+  - name: models/t2t_vit.py
+    size: 12499
+    checksum: caca4a1eced1616d403a3e582ebb30f79c9ab08be1b8789f4ddc5c34c55dccd0c95557062017e9794ff4e9b3bf017a63
+    source: https://raw.githubusercontent.com/yitu-opensource/T2T-ViT/2c25580c6b6968bba043be3c8f5d581428c6716d/models/t2t_vit.py
+  - name: models/token_transformer.py
+    size: 2326
+    checksum: a0e33192063f3ada47a504f5381c7f4c196ddae531b0514096ea06ee29bfa94a9716ffc8410fdd58bca07c732c0b4ef9
+    source: https://raw.githubusercontent.com/yitu-opensource/T2T-ViT/2c25580c6b6968bba043be3c8f5d581428c6716d/models/token_transformer.py
+  - name: models/token_performer.py
+    size: 2370
+    checksum: 77fd18c505f7780dd4f5a1149a208ef4b56b486bb9d89d05132321b32969ae79194ce5f49c8d194969176c1e9a4bedf6
+    source: https://raw.githubusercontent.com/yitu-opensource/T2T-ViT/2c25580c6b6968bba043be3c8f5d581428c6716d/models/token_performer.py
+  - name: models/transformer_block.py
+    size: 3286
+    checksum: e0ae48049b38c15664267d1e6a469f2427159d938feb9d438c37f6b922b65a959fe5172b8dee099fd3ae25a63c52489f
+    source: https://raw.githubusercontent.com/yitu-opensource/T2T-ViT/2c25580c6b6968bba043be3c8f5d581428c6716d/models/transformer_block.py
+  - name: 81.5_T2T_ViT_14.pth
+    size: 86212823
+    checksum: 98c8ae06a1e9997f3e83ad2dfbac49985bf7bd7723eb64e4b8462338163be02657998f0bd9f8997367417003caeaf7a1
+    source: https://github.com/yitu-opensource/T2T-ViT/releases/download/main/81.5_T2T_ViT_14.pth.tar
+postprocessing:
+  - $type: unpack_archive
+    format: zip
+    file: timm-0.5.4-py3-none-any.whl
+conversion_to_onnx_args:
+  - --model-path=$dl_dir
+  - --model-path=$config_dir
+  - --model-name=create_model
+  - --import-module=model
+  - --model-param=weights=r"$dl_dir/81.5_T2T_ViT_14.pth"
+  - --input-shape=1,3,224,224
+  - --input-names=image
+  - --output-names=probs
+  - --output-file=$conv_dir/t2t-vit-14.onnx
+  - --opset_version=12
+model_optimizer_args:
+  - --input_shape=[1,3,224,224]
+  - --input=image
+  - --input_model=$conv_dir/t2t-vit-14.onnx
+  - --mean_values=image[123.675,116.28,103.53]
+  - --scale_values=image[58.395,57.12,57.375]
+  - --reverse_input_channels
+  - --output=probs
+framework: pytorch
+license: https://raw.githubusercontent.com/yitu-opensource/T2T-ViT/main/LICENSE
diff --git a/tools/accuracy_checker/configs/t2t-vit-14.yml b/tools/accuracy_checker/configs/t2t-vit-14.yml
@@ -0,0 +1 @@
+../../../models/public/t2t-vit-14/accuracy-check.yml
diff --git a/tools/model_tools/src/openvino/model_zoo/internal_scripts/pytorch_to_onnx.py b/tools/model_tools/src/openvino/model_zoo/internal_scripts/pytorch_to_onnx.py
@@ -82,6 +82,7 @@ def parse_args():
                         help='Data type for inputs')
     parser.add_argument('--conversion-param', type=model_parameter, default=[], action='append',
                         help='Additional parameter for export')
+    parser.add_argument('--opset_version', type=int, default=11, help='The ONNX opset version')
     return parser.parse_args()
 
 
@@ -118,7 +119,8 @@ def load_model(model_name, weights, model_paths, module_name, model_params):
 
 
 @torch.no_grad()
-def convert_to_onnx(model, input_shapes, output_file, input_names, output_names, inputs_dtype, conversion_params):
+def convert_to_onnx(model, input_shapes, output_file, input_names, output_names, inputs_dtype, conversion_params,
+                    opset_version):
     """Convert PyTorch model to ONNX and check the resulting onnx model"""
 
     output_file.parent.mkdir(parents=True, exist_ok=True)
@@ -127,7 +129,7 @@ def convert_to_onnx(model, input_shapes, output_file, input_names, output_names,
         torch.zeros(input_shape, dtype=INPUT_DTYPE_TO_TORCH[inputs_dtype])
         for input_shape in input_shapes)
     model(*dummy_inputs)
-    torch.onnx.export(model, dummy_inputs, str(output_file), verbose=False, opset_version=11,
+    torch.onnx.export(model, dummy_inputs, str(output_file), verbose=False, opset_version=opset_version,
                       input_names=input_names.split(','), output_names=output_names.split(','), **conversion_params)
 
     model = onnx.load(str(output_file))
@@ -144,7 +146,8 @@ def main():
     model = load_model(args.model_name, args.weights,
                        args.model_paths, args.import_module, dict(args.model_param))
 
-    convert_to_onnx(model, args.input_shapes, args.output_file, args.input_names, args.output_names, args.inputs_dtype, dict(args.conversion_param))
+    convert_to_onnx(model, args.input_shapes, args.output_file, args.input_names, args.output_names, args.inputs_dtype,
+                    dict(args.conversion_param), args.opset_version)
 
 
 if __name__ == '__main__':

Original file line number	Diff line number	Diff line change
`@@ -0,0 +1 @@`
	`1`	`+../../../models/public/t2t-vit-14/accuracy-check.yml`