Upload two_stream_attn_v1_finetune_20260515T104743Z
Browse files- README.md +23 -23
- config.json +30 -30
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).
|
|
@@ -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)
|
|
@@ -171,7 +171,7 @@ Two-phase transfer learning pipeline:
|
|
| 171 |
- **Phase 1 (pretraining):** backbone pretrained on external checkpoint `two_stream_attn_v1_20260513T155730Z.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
|
| 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, gpm_components=0, forgetting_penalty_weight=0.5.
|
| 176 |
|
| 177 |
## Training Configuration
|
|
@@ -184,38 +184,38 @@ Two-phase transfer learning pipeline:
|
|
| 184 |
| Projection dim | 96 |
|
| 185 |
| Num layers | 4 |
|
| 186 |
| MHA heads | 8 (head dim: 24) |
|
| 187 |
-
| Dropout | 0.
|
| 188 |
| Learning rate | 3e-05 |
|
| 189 |
-
| Weight decay | 0.
|
| 190 |
| Batch size | 128 |
|
| 191 |
-
| Max epochs |
|
| 192 |
-
| Early stopping patience |
|
| 193 |
| Label smoothing | 0.05 |
|
| 194 |
| Class weighting | disabled |
|
| 195 |
| Max samples per class | 3000 |
|
| 196 |
-
| LR scheduler | ReduceLROnPlateau (factor=0.5, patience=
|
| 197 |
|
| 198 |
## Evaluation Results (Test Set)
|
| 199 |
|
| 200 |
| Metric | Value |
|
| 201 |
|--------|-------|
|
| 202 |
-
| Accuracy | 95.
|
| 203 |
-
| Macro F1 | 95.
|
| 204 |
|
| 205 |
### Per-Class Recall
|
| 206 |
|
| 207 |
| Class | Recall |
|
| 208 |
|-------|--------|
|
| 209 |
-
| `fist` | 96.
|
| 210 |
-
| `swiping_right` | 95.
|
| 211 |
-
| `swiping_left` | 98.
|
| 212 |
-
| `swiping_down` | 96.
|
| 213 |
-
| `swiping_up` | 97.
|
| 214 |
-
| `zooming_in_full_hand` | 97.
|
| 215 |
-
| `zooming_out_full_hand` | 93.
|
| 216 |
-
| `point_one` |
|
| 217 |
-
| `point_two` |
|
| 218 |
-
| `unknown` | 87.
|
| 219 |
|
| 220 |
## Comparison with Previous Architecture
|
| 221 |
|
|
|
|
| 17 |
- accuracy
|
| 18 |
- f1
|
| 19 |
model-index:
|
| 20 |
+
- name: two_stream_attn_v1_finetune_20260515T104743Z
|
| 21 |
results:
|
| 22 |
- task:
|
| 23 |
type: gesture-recognition
|
|
|
|
| 26 |
type: IPN-Hand
|
| 27 |
metrics:
|
| 28 |
- type: accuracy
|
| 29 |
+
value: 0.9566
|
| 30 |
- type: f1
|
| 31 |
+
value: 0.9556
|
| 32 |
---
|
| 33 |
|
| 34 |
+
# two_stream_attn_v1_finetune_20260515T104743Z
|
| 35 |
|
| 36 |
A real-time hand gesture classifier trained on
|
| 37 |
a Hybrid Jester+IPN gesture dataset (Jester dynamic classes + IPN pointing 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_finetune_20260515T104743Z_inference.pt",
|
| 143 |
)
|
| 144 |
|
| 145 |
# Load the artifact (includes model, class labels, and feature schema)
|
|
|
|
| 171 |
- **Phase 1 (pretraining):** backbone pretrained on external checkpoint `two_stream_attn_v1_20260513T155730Z.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 80 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, gpm_components=0, forgetting_penalty_weight=0.5.
|
| 176 |
|
| 177 |
## Training Configuration
|
|
|
|
| 184 |
| Projection dim | 96 |
|
| 185 |
| Num layers | 4 |
|
| 186 |
| MHA heads | 8 (head dim: 24) |
|
| 187 |
+
| Dropout | 0.4 |
|
| 188 |
| Learning rate | 3e-05 |
|
| 189 |
+
| Weight decay | 0.001 |
|
| 190 |
| Batch size | 128 |
|
| 191 |
+
| Max epochs | 80 |
|
| 192 |
+
| Early stopping patience | 20 |
|
| 193 |
| Label smoothing | 0.05 |
|
| 194 |
| Class weighting | disabled |
|
| 195 |
| Max samples per class | 3000 |
|
| 196 |
+
| LR scheduler | ReduceLROnPlateau (factor=0.5, patience=10) |
|
| 197 |
|
| 198 |
## Evaluation Results (Test Set)
|
| 199 |
|
| 200 |
| Metric | Value |
|
| 201 |
|--------|-------|
|
| 202 |
+
| Accuracy | 95.7% |
|
| 203 |
+
| Macro F1 | 95.6% |
|
| 204 |
|
| 205 |
### Per-Class Recall
|
| 206 |
|
| 207 |
| Class | Recall |
|
| 208 |
|-------|--------|
|
| 209 |
+
| `fist` | 96.2% |
|
| 210 |
+
| `swiping_right` | 95.5% |
|
| 211 |
+
| `swiping_left` | 98.7% |
|
| 212 |
+
| `swiping_down` | 96.6% |
|
| 213 |
+
| `swiping_up` | 97.6% |
|
| 214 |
+
| `zooming_in_full_hand` | 97.5% |
|
| 215 |
+
| `zooming_out_full_hand` | 93.9% |
|
| 216 |
+
| `point_one` | 97.4% |
|
| 217 |
+
| `point_two` | 94.3% |
|
| 218 |
+
| `unknown` | 87.7% |
|
| 219 |
|
| 220 |
## Comparison with Previous Architecture
|
| 221 |
|
config.json
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
{
|
| 2 |
-
"model_version": "
|
| 3 |
"model_config": {
|
| 4 |
"model_name": "two_stream_attn_v1",
|
| 5 |
"input_size": 147,
|
| 6 |
"hidden_size": 96,
|
| 7 |
"num_layers": 4,
|
| 8 |
-
"dropout": 0.
|
| 9 |
"num_classes": 10
|
| 10 |
},
|
| 11 |
"feature_schema": {
|
|
@@ -16,10 +16,10 @@
|
|
| 16 |
"window_step": 3
|
| 17 |
},
|
| 18 |
"training_config": {
|
| 19 |
-
"epochs":
|
| 20 |
"batch_size": 128,
|
| 21 |
"learning_rate": 3e-05,
|
| 22 |
-
"weight_decay": 0.
|
| 23 |
"grad_clip_norm": 1.0,
|
| 24 |
"seed": 42,
|
| 25 |
"label_smoothing": 0.05,
|
|
@@ -27,38 +27,38 @@
|
|
| 27 |
"max_samples_per_class": 3000,
|
| 28 |
"scheduler": {
|
| 29 |
"factor": 0.5,
|
| 30 |
-
"patience":
|
| 31 |
"min_lr": 1e-06
|
| 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.
|
| 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-
|
| 77 |
"gesture_command_mapping": {
|
| 78 |
"commands": {
|
| 79 |
"swiping_up": "start_presentation",
|
|
|
|
| 1 |
{
|
| 2 |
+
"model_version": "two_stream_attn_v1_finetune_20260515T104743Z",
|
| 3 |
"model_config": {
|
| 4 |
"model_name": "two_stream_attn_v1",
|
| 5 |
"input_size": 147,
|
| 6 |
"hidden_size": 96,
|
| 7 |
"num_layers": 4,
|
| 8 |
+
"dropout": 0.4,
|
| 9 |
"num_classes": 10
|
| 10 |
},
|
| 11 |
"feature_schema": {
|
|
|
|
| 16 |
"window_step": 3
|
| 17 |
},
|
| 18 |
"training_config": {
|
| 19 |
+
"epochs": 80,
|
| 20 |
"batch_size": 128,
|
| 21 |
"learning_rate": 3e-05,
|
| 22 |
+
"weight_decay": 0.001,
|
| 23 |
"grad_clip_norm": 1.0,
|
| 24 |
"seed": 42,
|
| 25 |
"label_smoothing": 0.05,
|
|
|
|
| 27 |
"max_samples_per_class": 3000,
|
| 28 |
"scheduler": {
|
| 29 |
"factor": 0.5,
|
| 30 |
+
"patience": 10,
|
| 31 |
"min_lr": 1e-06
|
| 32 |
}
|
| 33 |
},
|
| 34 |
"evaluation": {
|
| 35 |
+
"test_accuracy": 0.9566320645905421,
|
| 36 |
+
"test_macro_f1": 0.9555633825064029,
|
| 37 |
+
"test_loss": 0.4212703935896512,
|
| 38 |
+
"calibration_ece": 0.030699590615060227,
|
| 39 |
"per_class_recall": {
|
| 40 |
+
"fist": 0.9621993127147767,
|
| 41 |
+
"swiping_right": 0.9548532731376975,
|
| 42 |
+
"swiping_left": 0.9869158878504672,
|
| 43 |
+
"swiping_down": 0.9664694280078896,
|
| 44 |
+
"swiping_up": 0.9755600814663951,
|
| 45 |
+
"zooming_in_full_hand": 0.975,
|
| 46 |
+
"zooming_out_full_hand": 0.9388185654008439,
|
| 47 |
+
"point_one": 0.973753280839895,
|
| 48 |
+
"point_two": 0.9429347826086957,
|
| 49 |
+
"unknown": 0.8765432098765432
|
| 50 |
},
|
| 51 |
"per_class_precision": {
|
| 52 |
+
"fist": 0.9790209790209791,
|
| 53 |
+
"swiping_right": 0.9701834862385321,
|
| 54 |
+
"swiping_left": 0.9777777777777777,
|
| 55 |
+
"swiping_down": 0.9551656920077972,
|
| 56 |
+
"swiping_up": 0.9618473895582329,
|
| 57 |
+
"zooming_in_full_hand": 0.9407894736842105,
|
| 58 |
+
"zooming_out_full_hand": 0.973741794310722,
|
| 59 |
+
"point_one": 0.8918269230769231,
|
| 60 |
+
"point_two": 0.9719887955182073,
|
| 61 |
+
"unknown": 0.9441489361702128
|
| 62 |
}
|
| 63 |
},
|
| 64 |
"class_labels": [
|
|
|
|
| 73 |
"point_two",
|
| 74 |
"unknown"
|
| 75 |
],
|
| 76 |
+
"created_at": "2026-05-15T10:57:07.231228+00:00",
|
| 77 |
"gesture_command_mapping": {
|
| 78 |
"commands": {
|
| 79 |
"swiping_up": "start_presentation",
|