Upload two_stream_attn_v1_2layer_ld_cong_finetune_20260515T134706Z
Browse files- README.md +20 -20
- config.json +26 -28
README.md
CHANGED
|
@@ -17,7 +17,7 @@ metrics:
|
|
| 17 |
- accuracy
|
| 18 |
- f1
|
| 19 |
model-index:
|
| 20 |
-
- name:
|
| 21 |
results:
|
| 22 |
- task:
|
| 23 |
type: gesture-recognition
|
|
@@ -26,12 +26,12 @@ model-index:
|
|
| 26 |
type: IPN-Hand
|
| 27 |
metrics:
|
| 28 |
- type: accuracy
|
| 29 |
-
value: 0.
|
| 30 |
- type: f1
|
| 31 |
-
value: 0.
|
| 32 |
---
|
| 33 |
|
| 34 |
-
#
|
| 35 |
|
| 36 |
A real-time hand gesture classifier trained on
|
| 37 |
a Hybrid Jester+IPN gesture dataset (Jester dynamic classes + IPN pointing classes).
|
|
@@ -43,7 +43,7 @@ standard webcam using MediaPipe for landmark extraction.
|
|
| 43 |
## Model Description
|
| 44 |
|
| 45 |
- **Architecture**: EnhancedTwoStreamLSTM (BiLSTM h=96×2, MHA 8 heads, proj=96, mean+max pool, MLP gate)
|
| 46 |
-
- **Parameters**:
|
| 47 |
- **Input**: `(batch, 16, 147)`
|
| 48 |
— 16-frame sliding window at 30 FPS ≈ 533 ms
|
| 49 |
- **Output**: Softmax logits over 10 gesture classes
|
|
@@ -139,7 +139,7 @@ from maestro.infrastructure.model.checkpoint_loader import load_inference_artifa
|
|
| 139 |
# Download the artifact (cached after first call)
|
| 140 |
local_path = hf_hub_download(
|
| 141 |
repo_id="ntsrigaud/maestro-lstm-hybrid",
|
| 142 |
-
filename="
|
| 143 |
)
|
| 144 |
|
| 145 |
# Load the artifact (includes model, class labels, and feature schema)
|
|
@@ -168,10 +168,10 @@ with torch.no_grad():
|
|
| 168 |
## Training Strategy
|
| 169 |
|
| 170 |
Two-phase transfer learning pipeline:
|
| 171 |
-
- **Phase 1 (pretraining):** backbone pretrained on external checkpoint `
|
| 172 |
- **Phase 2 (fine-tuning):** head replaced and model adapted on Hybrid Jester+IPN 10-gesture vocabulary.
|
| 173 |
- **Stage A (frozen backbone):** 10 epoch(s) head-only warmup.
|
| 174 |
-
- **Stage B (full model):** up to
|
| 175 |
- **Stage B retention defences:** replay_max_samples_per_class=500, distillation_weight=0.0, replay_ce_weight=0.0, backbone_lr_multiplier=0.1, ewc_weight=N/A, gpm_components=0, forgetting_penalty_weight=0.5.
|
| 176 |
|
| 177 |
## Training Configuration
|
|
@@ -199,23 +199,23 @@ Two-phase transfer learning pipeline:
|
|
| 199 |
|
| 200 |
| Metric | Value |
|
| 201 |
|--------|-------|
|
| 202 |
-
| Accuracy |
|
| 203 |
-
| Macro F1 | 95.
|
| 204 |
|
| 205 |
### Per-Class Recall
|
| 206 |
|
| 207 |
| Class | Recall |
|
| 208 |
|-------|--------|
|
| 209 |
-
| `fist` |
|
| 210 |
-
| `swiping_right` |
|
| 211 |
-
| `swiping_left` | 98.
|
| 212 |
-
| `swiping_down` |
|
| 213 |
-
| `swiping_up` |
|
| 214 |
-
| `zooming_in_full_hand` | 97.
|
| 215 |
-
| `zooming_out_full_hand` |
|
| 216 |
| `point_one` | 97.4% |
|
| 217 |
-
| `point_two` |
|
| 218 |
-
| `unknown` |
|
| 219 |
|
| 220 |
## Comparison with Previous Architecture
|
| 221 |
|
|
@@ -226,7 +226,7 @@ Two-phase transfer learning pipeline:
|
|
| 226 |
| Feature projection | No | **Yes (→96)** |
|
| 227 |
| Temporal pooling | Mean only | **Mean + Max** |
|
| 228 |
| Cross-stream fusion | Concat only | **2-layer MLP gate** |
|
| 229 |
-
| Parameters | ~182 K | ~
|
| 230 |
|
| 231 |
## Limitations and Risks
|
| 232 |
|
|
|
|
| 17 |
- accuracy
|
| 18 |
- f1
|
| 19 |
model-index:
|
| 20 |
+
- name: two_stream_attn_v1_2layer_ld_cong_finetune_20260515T134706Z
|
| 21 |
results:
|
| 22 |
- task:
|
| 23 |
type: gesture-recognition
|
|
|
|
| 26 |
type: IPN-Hand
|
| 27 |
metrics:
|
| 28 |
- type: accuracy
|
| 29 |
+
value: 0.9606
|
| 30 |
- type: f1
|
| 31 |
+
value: 0.9587
|
| 32 |
---
|
| 33 |
|
| 34 |
+
# two_stream_attn_v1_2layer_ld_cong_finetune_20260515T134706Z
|
| 35 |
|
| 36 |
A real-time hand gesture classifier trained on
|
| 37 |
a Hybrid Jester+IPN gesture dataset (Jester dynamic classes + IPN pointing classes).
|
|
|
|
| 43 |
## Model Description
|
| 44 |
|
| 45 |
- **Architecture**: EnhancedTwoStreamLSTM (BiLSTM h=96×2, MHA 8 heads, proj=96, mean+max pool, MLP gate)
|
| 46 |
+
- **Parameters**: 1,208,554
|
| 47 |
- **Input**: `(batch, 16, 147)`
|
| 48 |
— 16-frame sliding window at 30 FPS ≈ 533 ms
|
| 49 |
- **Output**: Softmax logits over 10 gesture classes
|
|
|
|
| 139 |
# Download the artifact (cached after first call)
|
| 140 |
local_path = hf_hub_download(
|
| 141 |
repo_id="ntsrigaud/maestro-lstm-hybrid",
|
| 142 |
+
filename="two_stream_attn_v1_2layer_ld_cong_finetune_20260515T134706Z_inference.pt",
|
| 143 |
)
|
| 144 |
|
| 145 |
# Load the artifact (includes model, class labels, and feature schema)
|
|
|
|
| 168 |
## Training Strategy
|
| 169 |
|
| 170 |
Two-phase transfer learning pipeline:
|
| 171 |
+
- **Phase 1 (pretraining):** backbone pretrained on external checkpoint `two_stream_attn_v1_2layer_pretrain_20260515T125437Z.pt` to learn generic gesture dynamics.
|
| 172 |
- **Phase 2 (fine-tuning):** head replaced and model adapted on Hybrid Jester+IPN 10-gesture vocabulary.
|
| 173 |
- **Stage A (frozen backbone):** 10 epoch(s) head-only warmup.
|
| 174 |
+
- **Stage B (full model):** up to 66 epoch(s) joint fine-tuning with scheduler/early stopping.
|
| 175 |
- **Stage B retention defences:** replay_max_samples_per_class=500, distillation_weight=0.0, replay_ce_weight=0.0, backbone_lr_multiplier=0.1, ewc_weight=N/A, gpm_components=0, forgetting_penalty_weight=0.5.
|
| 176 |
|
| 177 |
## Training Configuration
|
|
|
|
| 199 |
|
| 200 |
| Metric | Value |
|
| 201 |
|--------|-------|
|
| 202 |
+
| Accuracy | 96.1% |
|
| 203 |
+
| Macro F1 | 95.9% |
|
| 204 |
|
| 205 |
### Per-Class Recall
|
| 206 |
|
| 207 |
| Class | Recall |
|
| 208 |
|-------|--------|
|
| 209 |
+
| `fist` | 97.3% |
|
| 210 |
+
| `swiping_right` | 97.1% |
|
| 211 |
+
| `swiping_left` | 98.3% |
|
| 212 |
+
| `swiping_down` | 98.0% |
|
| 213 |
+
| `swiping_up` | 98.2% |
|
| 214 |
+
| `zooming_in_full_hand` | 97.0% |
|
| 215 |
+
| `zooming_out_full_hand` | 95.1% |
|
| 216 |
| `point_one` | 97.4% |
|
| 217 |
+
| `point_two` | 95.1% |
|
| 218 |
+
| `unknown` | 85.7% |
|
| 219 |
|
| 220 |
## Comparison with Previous Architecture
|
| 221 |
|
|
|
|
| 226 |
| Feature projection | No | **Yes (→96)** |
|
| 227 |
| Temporal pooling | Mean only | **Mean + Max** |
|
| 228 |
| Cross-stream fusion | Concat only | **2-layer MLP gate** |
|
| 229 |
+
| Parameters | ~182 K | ~1,208,554 |
|
| 230 |
|
| 231 |
## Limitations and Risks
|
| 232 |
|
config.json
CHANGED
|
@@ -1,7 +1,7 @@
|
|
| 1 |
{
|
| 2 |
-
"model_version": "
|
| 3 |
"model_config": {
|
| 4 |
-
"model_name": "
|
| 5 |
"input_size": 147,
|
| 6 |
"hidden_size": 96,
|
| 7 |
"num_layers": 2,
|
|
@@ -32,33 +32,33 @@
|
|
| 32 |
}
|
| 33 |
},
|
| 34 |
"evaluation": {
|
| 35 |
-
"test_accuracy": 0.
|
| 36 |
-
"test_macro_f1": 0.
|
| 37 |
-
"test_loss": 0.
|
| 38 |
-
"calibration_ece": 0.
|
| 39 |
"per_class_recall": {
|
| 40 |
-
"fist": 0.
|
| 41 |
-
"swiping_right": 0.
|
| 42 |
-
"swiping_left": 0.
|
| 43 |
-
"swiping_down": 0.
|
| 44 |
-
"swiping_up": 0.
|
| 45 |
-
"zooming_in_full_hand": 0.
|
| 46 |
-
"zooming_out_full_hand": 0.
|
| 47 |
"point_one": 0.973753280839895,
|
| 48 |
-
"point_two": 0.
|
| 49 |
-
"unknown": 0.
|
| 50 |
},
|
| 51 |
"per_class_precision": {
|
| 52 |
-
"fist": 0.
|
| 53 |
-
"swiping_right": 0.
|
| 54 |
-
"swiping_left": 0.
|
| 55 |
-
"swiping_down": 0.
|
| 56 |
-
"swiping_up": 0.
|
| 57 |
-
"zooming_in_full_hand": 0.
|
| 58 |
-
"zooming_out_full_hand": 0.
|
| 59 |
-
"point_one": 0.
|
| 60 |
-
"point_two": 0.
|
| 61 |
-
"unknown": 0.
|
| 62 |
}
|
| 63 |
},
|
| 64 |
"class_labels": [
|
|
@@ -73,7 +73,7 @@
|
|
| 73 |
"point_two",
|
| 74 |
"unknown"
|
| 75 |
],
|
| 76 |
-
"created_at": "2026-05-15T13:
|
| 77 |
"gesture_command_mapping": {
|
| 78 |
"commands": {
|
| 79 |
"swiping_up": "start_presentation",
|
|
@@ -83,8 +83,6 @@
|
|
| 83 |
"zooming_in_full_hand": "zoom_in_view",
|
| 84 |
"zooming_out_full_hand": "zoom_out_view",
|
| 85 |
"fist": "erase_annotations",
|
| 86 |
-
"pinch": "activate_laser_pointer",
|
| 87 |
-
"click": "mouse_click",
|
| 88 |
"unknown": "no_action"
|
| 89 |
},
|
| 90 |
"modes": {
|
|
|
|
| 1 |
{
|
| 2 |
+
"model_version": "two_stream_attn_v1_2layer_ld_cong_finetune_20260515T134706Z",
|
| 3 |
"model_config": {
|
| 4 |
+
"model_name": "two_stream_attn_v1_2layer_ld_cong",
|
| 5 |
"input_size": 147,
|
| 6 |
"hidden_size": 96,
|
| 7 |
"num_layers": 2,
|
|
|
|
| 32 |
}
|
| 33 |
},
|
| 34 |
"evaluation": {
|
| 35 |
+
"test_accuracy": 0.960553633217993,
|
| 36 |
+
"test_macro_f1": 0.9587371203998121,
|
| 37 |
+
"test_loss": 0.404887427927576,
|
| 38 |
+
"calibration_ece": 0.033548892410568715,
|
| 39 |
"per_class_recall": {
|
| 40 |
+
"fist": 0.9725085910652921,
|
| 41 |
+
"swiping_right": 0.9706546275395034,
|
| 42 |
+
"swiping_left": 0.983177570093458,
|
| 43 |
+
"swiping_down": 0.980276134122288,
|
| 44 |
+
"swiping_up": 0.9816700610997964,
|
| 45 |
+
"zooming_in_full_hand": 0.9704545454545455,
|
| 46 |
+
"zooming_out_full_hand": 0.9514767932489452,
|
| 47 |
"point_one": 0.973753280839895,
|
| 48 |
+
"point_two": 0.9510869565217391,
|
| 49 |
+
"unknown": 0.8567901234567902
|
| 50 |
},
|
| 51 |
"per_class_precision": {
|
| 52 |
+
"fist": 0.9433333333333334,
|
| 53 |
+
"swiping_right": 0.9728506787330317,
|
| 54 |
+
"swiping_left": 0.9813432835820896,
|
| 55 |
+
"swiping_down": 0.9613152804642167,
|
| 56 |
+
"swiping_up": 0.9620758483033932,
|
| 57 |
+
"zooming_in_full_hand": 0.9510022271714922,
|
| 58 |
+
"zooming_out_full_hand": 0.9740820734341252,
|
| 59 |
+
"point_one": 0.9298245614035088,
|
| 60 |
+
"point_two": 0.958904109589041,
|
| 61 |
+
"unknown": 0.9559228650137741
|
| 62 |
}
|
| 63 |
},
|
| 64 |
"class_labels": [
|
|
|
|
| 73 |
"point_two",
|
| 74 |
"unknown"
|
| 75 |
],
|
| 76 |
+
"created_at": "2026-05-15T13:52:14.109098+00:00",
|
| 77 |
"gesture_command_mapping": {
|
| 78 |
"commands": {
|
| 79 |
"swiping_up": "start_presentation",
|
|
|
|
| 83 |
"zooming_in_full_hand": "zoom_in_view",
|
| 84 |
"zooming_out_full_hand": "zoom_out_view",
|
| 85 |
"fist": "erase_annotations",
|
|
|
|
|
|
|
| 86 |
"unknown": "no_action"
|
| 87 |
},
|
| 88 |
"modes": {
|