Merge pull request #7 from ARM-software/ds_cnn_large_clustered

Added DS-CNN Large Clustered

MODEL_LIST.md

@@ -13,6 +13,8 @@
 | [DNN Large INT8](models/keyword_spotting/dnn_large/tflite_int8) | INT8 | TensorFlow Lite |
 | [DNN Medium INT8](models/keyword_spotting/dnn_medium/tflite_int8) | INT8 | TensorFlow Lite |
 | [DNN Small INT8](models/keyword_spotting/dnn_small/tflite_int8) | INT8 | TensorFlow Lite |
+| [DS-CNN Clustered FP32](models/keyword_spotting/ds_cnn_large/tflite_clustered_fp32) | FP32 | TensorFlow Lite |
+| [DS-CNN Clustered INT8](models/keyword_spotting/ds_cnn_large/tflite_clustered_int8) | INT8 | TensorFlow Lite |
 | [DS-CNN Large INT8](models/keyword_spotting/ds_cnn_large/tflite_int8) | INT8 | TensorFlow Lite |
 | [DS-CNN Medium INT8](models/keyword_spotting/ds_cnn_medium/tflite_int8) | INT8 | TensorFlow Lite |
 | [DS-CNN Small INT8](models/keyword_spotting/ds_cnn_small/tflite_int8) | INT8 | TensorFlow Lite |

models/keyword_spotting/ds_cnn_large/tflite_clustered_fp32/README.md

@@ -0,0 +1,50 @@
+# DS-CNN Clustered FP32
+
+## Description
+This is a clustered (32 clusters, kmeans++ centroid initialization) and retrained (fine-tuned) FP32 version of the DS-CNN Large model developed by Arm from the Hello Edge paper. Code for the original DS-CNN implementation can be found here: https://github.com/ARM-software/ML-KWS-for-MCU. The original model was converted to Keras and optimized using the Clustering API in TensorFlow Model Optimization Toolkit.
+
+## License
+[Apache-2.0](https://spdx.org/licenses/Apache-2.0.html)
+
+## Network Information
+| Network Information |  Value         |
+|---------------------|------------------|
+|  Framework          | TensorFlow Lite |
+|  SHA-1 Hash         | a6f4fa3e253125499a1e77aa1010c943e72af568 |
+|  Size (Bytes)       | 1649816 |
+|  Provenance         | The original model (before clustering and quantization) is a pretrained checkpoint based on https://github.com/ARM-software/ML-KWS-for-MCU |
+|  Paper              | https://arxiv.org/abs/1711.07128 |
+
+## Accuracy
+Dataset: Google Speech Commands
+
+| Metric | Value |
+|--------|-------|
+| Top 1 Accuracy | 0.9506 |
+
+## Performance
+| Platform | Optimized |
+| -------- | ---------- |
+|   CPU    |      :heavy_check_mark:      |
+|   GPU    |      :heavy_check_mark:      |
+
+### Key
+ - :heavy_check_mark: - Optimized for the platform.
+ - :heavy_minus_sign: - Not optimized, but will run on the platform.
+ - :heavy_multiplication_x: - Not optimized and will not run on the platform.
+
+## Optimizations
+| Optimization |  Value  |
+|-----------------|---------|
+| Number of Clusters | 32 |
+| Cluster Initialization | K-Means |
+
+## Network Inputs
+| Input Node Name |  Shape  | Description |
+|-----------------|---------|-------------|
+| input_4 | (1, 1, 49, 10) | The input is a processed MFCCs of shape (1,1,49,10) |
+
+## Network Outputs
+| Output Node Name |  Shape  | Description |
+|------------------|---------|-------------|
+| Identity | (1, 12) | The probability on 12 keywords. |

models/keyword_spotting/ds_cnn_large/tflite_clustered_fp32/definition.yaml

@@ -0,0 +1,46 @@
+benchmark:
+  SpeechCommands:
+    top_1_accuracy: 0.9506
+description: 'This is a clustered (32 clusters, kmeans++ centroid initialization) and
+  retrained (fine-tuned) FP32 version of the DS-CNN Large model developed by Arm from the Hello Edge paper.
+  Code for the original DS-CNN implementation can be found here: https://github.com/ARM-software/ML-KWS-for-MCU. 
+  The original model was converted to Keras and optimized using the Clustering API in TensorFlow Model Optimization Toolkit.'
+license:
+- Apache-2.0
+network:
+  file_size_bytes: 1649816
+  filename: ds_cnn_clustered_fp32.tflite
+  framework: TensorFlow Lite
+  hash:
+    algorithm: sha1
+    value: a6f4fa3e253125499a1e77aa1010c943e72af568
+  provenance: The original model (before clustering and quantization) is a pretrained checkpoint based on https://github.com/ARM-software/ML-KWS-for-MCU
+network_parameters:
+  input_nodes:
+  - description: The input is a processed MFCCs of shape (1,1,49,10)
+    example_input:
+      path: models/keyword_spotting/ds_cnn_large/tflite_clustered_fp32/testing_input/input_4
+    name: input_4
+    shape:
+    - 1
+    - 1
+    - 49
+    - 10
+    type: float32
+  output_nodes:
+  - description: The probability on 12 keywords.
+    name: Identity
+    shape:
+    - 1
+    - 12
+    test_output_path: models/keyword_spotting/ds_cnn_large/tflite_clustered_fp32/testing_output/Identity
+operators:
+  TensorFlow Lite:
+  - AVERAGE_POOL_2D
+  - CONV_2D
+  - DEPTHWISE_CONV_2D
+  - FULLY_CONNECTED
+  - RELU
+  - RESHAPE
+  - SOFTMAX
+paper: https://arxiv.org/abs/1711.07128

models/keyword_spotting/ds_cnn_large/tflite_clustered_int8/README.md

@@ -0,0 +1,52 @@
+# DS-CNN Clustered INT8
+
+## Description
+This is a clustered (32 clusters, kmeans++ centroid initialization), retrained (fine-tuned) and fully quantized version (INT8) of the DS-CNN Large model developed by Arm from the Hello Edge paper. Code for the original DS-CNN implementation can be found here: https://github.com/ARM-software/ML-KWS-for-MCU. 
+The original model was converted to Keras, optimized using the Clustering API in TensorFlow Model Optimization Toolkit, and quantized using post-training quantization in the TF Lite Converter.
+
+## License
+[Apache-2.0](https://spdx.org/licenses/Apache-2.0.html)
+
+## Network Information
+| Network Information |  Value         |
+|---------------------|------------------|
+|  Framework          | TensorFlow Lite |
+|  SHA-1 Hash         | 367391990618e4e9a821a847dcbf3fe129212282 |
+|  Size (Bytes)       | 526192 |
+|  Provenance         | The original model (before clustering) is a pretrained checkpoint based on https://github.com/ARM-software/ML-KWS-for-MCU |
+|  Paper              | https://arxiv.org/abs/1711.07128 |
+
+## Accuracy
+Dataset: Google Speech Commands
+
+| Metric | Value |
+|--------|-------|
+| Top 1 Accuracy | 0.94701 |
+
+## Performance
+| Platform | Optimized |
+| -------- | ---------- |
+|   CPU    |      :heavy_check_mark:      |
+|   GPU    |      :heavy_check_mark:      |
+
+### Key
+ - :heavy_check_mark: - Optimized for the platform.
+ - :heavy_minus_sign: - Not optimized, but will run on the platform.
+ - :heavy_multiplication_x: - Not optimized and will not run on the platform.
+
+## Optimizations
+| Optimization |  Value  |
+|-----------------|---------|
+| Quantization | INT8 |
+| Number of Clusters | 32 |
+| Cluster Initialization | K-Means |
+
+## Network Inputs
+| Input Node Name |  Shape  | Description |
+|-----------------|---------|-------------|
+| input_2 | (1, 1, 49, 10) | The input is a processed MFCCs of shape (1,1,49,10) |
+
+## Network Outputs
+| Output Node Name |  Shape  | Description |
+|------------------|---------|-------------|
+| Identity | (1, 12) | The probability on 12 keywords. |

models/keyword_spotting/ds_cnn_large/tflite_clustered_int8/definition.yaml

@@ -0,0 +1,46 @@
+benchmark:
+  SpeechCommands:
+    top_1_accuracy: 0.94701
+description: 'This is a clustered (32 clusters, kmeans++ centroid initialization), retrained (fine-tuned) and fully quantized version (INT8)
+  of the DS-CNN Large model developed by Arm from the Hello Edge paper. Code for the original DS-CNN implementation 
+  can be found here: https://github.com/ARM-software/ML-KWS-for-MCU. 
+  The original model was converted to Keras, optimized using the Clustering API in TensorFlow Model Optimization Toolkit, and quantized using post-training quantization in the TF Lite Converter.'
+license:
+- Apache-2.0
+network:
+  file_size_bytes: 526192
+  filename: ds_cnn_clustered_int8.tflite
+  framework: TensorFlow Lite
+  hash:
+    algorithm: sha1
+    value: 367391990618e4e9a821a847dcbf3fe129212282
+  provenance: The original model (before clustering) is a pretrained checkpoint based on https://github.com/ARM-software/ML-KWS-for-MCU
+network_parameters:
+  input_nodes:
+  - description: The input is a processed MFCCs of shape (1,1,49,10)
+    example_input:
+      path: models/keyword_spotting/ds_cnn_large/tflite_clustered_int8/testing_input/input_2
+    name: input_2
+    shape:
+    - 1
+    - 1
+    - 49
+    - 10
+    type: int8
+  output_nodes:
+  - description: The probability on 12 keywords.
+    name: Identity
+    shape:
+    - 1
+    - 12
+    test_output_path: models/keyword_spotting/ds_cnn_large/tflite_clustered_int8/testing_output/Identity
+operators:
+  TensorFlow Lite:
+  - AVERAGE_POOL_2D
+  - CONV_2D
+  - DEPTHWISE_CONV_2D
+  - FULLY_CONNECTED
+  - RELU
+  - RESHAPE
+  - SOFTMAX
+paper: https://arxiv.org/abs/1711.07128