Release deployment artifacts from CompressionKit
Browse files- README.md +10 -0
- codec_spec.json +19 -1
- decoder.keras +1 -1
- encoder.keras +1 -1
- model_card.json +29 -0
README.md
CHANGED
|
@@ -61,6 +61,16 @@ _PRD here is faithfulness (vs the recorded input); see **Fidelity & Robustness**
|
|
| 61 |
### Bitrate
|
| 62 |
|
| 63 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
## Usage
|
| 65 |
|
| 66 |
### Python (compressionkit runtime)
|
|
|
|
| 61 |
### Bitrate
|
| 62 |
|
| 63 |
|
| 64 |
+
## Encoder Precision Parity
|
| 65 |
+
|
| 66 |
+
Difference from FP32 reconstruction on a disjoint real-data holdout; lower is better.
|
| 67 |
+
|
| 68 |
+
| Encoder | P90 PRD | Worst PRD | Status |
|
| 69 |
+
|---|---:|---:|---|
|
| 70 |
+
| FP16 | 0.00% | 8.56% | recommended |
|
| 71 |
+
| INT16X8 | 5.15% | 17.46% | recommended |
|
| 72 |
+
| INT8 | 10.77% | 42.82% | higher error; use knowingly |
|
| 73 |
+
|
| 74 |
## Usage
|
| 75 |
|
| 76 |
### Python (compressionkit runtime)
|
codec_spec.json
CHANGED
|
@@ -14,7 +14,7 @@
|
|
| 14 |
320,
|
| 15 |
1
|
| 16 |
],
|
| 17 |
-
"encoder_input_dtype": "
|
| 18 |
},
|
| 19 |
"preprocessing_contract": {
|
| 20 |
"format_version": 1,
|
|
@@ -52,6 +52,24 @@
|
|
| 52 |
"int16x8_header": "encoder_int16x8.h",
|
| 53 |
"int8_tflite": "encoder.tflite",
|
| 54 |
"int8_header": "encoder.h",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
"keras": "encoder.keras",
|
| 56 |
"input_shape": [
|
| 57 |
null,
|
|
|
|
| 14 |
320,
|
| 15 |
1
|
| 16 |
],
|
| 17 |
+
"encoder_input_dtype": "float32"
|
| 18 |
},
|
| 19 |
"preprocessing_contract": {
|
| 20 |
"format_version": 1,
|
|
|
|
| 52 |
"int16x8_header": "encoder_int16x8.h",
|
| 53 |
"int8_tflite": "encoder.tflite",
|
| 54 |
"int8_header": "encoder.h",
|
| 55 |
+
"variants": {
|
| 56 |
+
"float32": {
|
| 57 |
+
"tflite": "encoder_float32.tflite",
|
| 58 |
+
"input_dtype": "float32"
|
| 59 |
+
},
|
| 60 |
+
"fp16": {
|
| 61 |
+
"tflite": "encoder_fp16.tflite",
|
| 62 |
+
"input_dtype": "float32"
|
| 63 |
+
},
|
| 64 |
+
"int16x8": {
|
| 65 |
+
"tflite": "encoder_int16x8.tflite",
|
| 66 |
+
"input_dtype": "float32"
|
| 67 |
+
},
|
| 68 |
+
"int8": {
|
| 69 |
+
"tflite": "encoder.tflite",
|
| 70 |
+
"input_dtype": "int8"
|
| 71 |
+
}
|
| 72 |
+
},
|
| 73 |
"keras": "encoder.keras",
|
| 74 |
"input_shape": [
|
| 75 |
null,
|
decoder.keras
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 663387
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a5d69284ded81a973349a4c080a9754b921ca712bf5cea362bf9dbcf450029a2
|
| 3 |
size 663387
|
encoder.keras
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 239292
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8e153f8ac315dd501b2c3d3354adc132ad03c2fa1c3e3a92a1105888cffe028b
|
| 3 |
size 239292
|
model_card.json
CHANGED
|
@@ -1112,5 +1112,34 @@
|
|
| 1112 |
}
|
| 1113 |
}
|
| 1114 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1115 |
}
|
| 1116 |
}
|
|
|
|
| 1112 |
}
|
| 1113 |
}
|
| 1114 |
}
|
| 1115 |
+
},
|
| 1116 |
+
"encoder_precision_report": {
|
| 1117 |
+
"FP16": {
|
| 1118 |
+
"code_index_match_fraction_median": 1.0,
|
| 1119 |
+
"encoder_input_saturation_fraction_max": 0.0,
|
| 1120 |
+
"encoder_input_saturation_fraction_p90": 0.0,
|
| 1121 |
+
"frames_checked": 2048,
|
| 1122 |
+
"latent_prd_percent_p90": 0.02462919144008737,
|
| 1123 |
+
"reconstruction_prd_percent_max": 8.555098757135863,
|
| 1124 |
+
"reconstruction_prd_percent_p90": 0.0
|
| 1125 |
+
},
|
| 1126 |
+
"INT16X8": {
|
| 1127 |
+
"code_index_match_fraction_median": 1.0,
|
| 1128 |
+
"encoder_input_saturation_fraction_max": 0.0,
|
| 1129 |
+
"encoder_input_saturation_fraction_p90": 0.0,
|
| 1130 |
+
"frames_checked": 2048,
|
| 1131 |
+
"latent_prd_percent_p90": 0.5581851404504072,
|
| 1132 |
+
"reconstruction_prd_percent_max": 17.462837096401103,
|
| 1133 |
+
"reconstruction_prd_percent_p90": 5.151539496067478
|
| 1134 |
+
},
|
| 1135 |
+
"INT8": {
|
| 1136 |
+
"code_index_match_fraction_median": 0.85,
|
| 1137 |
+
"encoder_input_saturation_fraction_max": 0.0,
|
| 1138 |
+
"encoder_input_saturation_fraction_p90": 0.0,
|
| 1139 |
+
"frames_checked": 2048,
|
| 1140 |
+
"latent_prd_percent_p90": 3.734960589655774,
|
| 1141 |
+
"reconstruction_prd_percent_max": 42.82338073841767,
|
| 1142 |
+
"reconstruction_prd_percent_p90": 10.7728172259162
|
| 1143 |
+
}
|
| 1144 |
}
|
| 1145 |
}
|