Tonic commited on
Commit
0d770c2
·
verified ·
1 Parent(s): b6bb7e0

esp32/audioset: update export_nano_tflite_job.py

Browse files
Files changed (1) hide show
  1. export_nano_tflite_job.py +94 -88
export_nano_tflite_job.py CHANGED
@@ -1,88 +1,94 @@
1
- # /// script
2
- # requires-python = ">=3.11"
3
- # dependencies = ["torch", "numpy", "onnx", "onnxscript", "onnx2tf", "onnx-graphsurgeon", "onnxsim", "psutil", "ai-edge-litert", "sng4onnx", "pandas", "flatbuffers", "joblib", "termcolor", "requests", "tqdm", "packaging", "tensorflow-cpu", "tf-keras", "huggingface_hub>=0.30"]
4
- # ///
5
- """Convert the trained nano96 ESP32 checkpoint to an int8 TFLite model + model_data.h.
6
-
7
- Waits for esp32/nano96_best.pt in Tonic/challenge5-acoustic-code (uploaded by the
8
- train_nano job), then: best.pt -> ONNX (1x3x96x96, static) -> SavedModel (onnx2tf)
9
- -> TFLite dynamic-range int8 -> C header model_data.h for firmware/node.
10
-
11
- Uploads to code repo: esp32/model_nano_96.tflite, esp32/model_nano_96_dynamic_int8.onnx,
12
- esp32/model_data.h
13
- """
14
- import os, sys, time, importlib.util
15
-
16
- tok = os.environ["HF_TOKEN"]
17
- from huggingface_hub import HfApi, hf_hub_download
18
- api = HfApi(token=tok)
19
- CODE = "Tonic/challenge5-acoustic-code"
20
-
21
- # 1) wait for the nano checkpoint (train_nano job uploads it at the end)
22
- print("waiting for esp32/nano96_best.pt ...", flush=True)
23
- ckpt = None
24
- for i in range(90): # up to ~45 min
25
- try:
26
- ckpt = hf_hub_download(CODE, "esp32/nano96_best.pt", repo_type="model", token=tok)
27
- break
28
- except Exception:
29
- time.sleep(30)
30
- if ckpt is None:
31
- raise SystemExit("nano96_best.pt never appeared")
32
- print("ckpt:", os.path.getsize(ckpt) // 1024, "KB", flush=True)
33
-
34
- import torch
35
- ck = torch.load(ckpt, map_location="cpu", weights_only=False)
36
- sd = ck.get("state_dict", ck) if isinstance(ck, dict) else ck
37
-
38
- # 2) rebuild the nano arch from the trainer script in the same repo
39
- tj = hf_hub_download(CODE, "train_nano_job.py", repo_type="model", token=tok)
40
- spec = importlib.util.spec_from_file_location("train_nano_job", tj)
41
- tn = importlib.util.module_from_spec(spec)
42
- spec.loader.exec_module(tn)
43
- model = tn.build_nano()
44
- model.load_state_dict(sd, strict=True)
45
- model.eval()
46
- print("nano params:", sum(p.numel() for p in model.parameters()), flush=True)
47
-
48
- # 3) ONNX (static 1x3x96x96 — TFLite-Micro friendly)
49
- onnx_path = "/tmp/model_nano_96.onnx"
50
- torch.onnx.export(model, torch.zeros(1, 3, 96, 96), onnx_path, opset_version=17, dynamo=False,
51
- input_names=["spectrogram"], output_names=["wake_logit", "sound_class_logits"])
52
- print("onnx exported", flush=True)
53
-
54
- # 4) ONNX -> SavedModel -> TFLite dynamic int8
55
- import onnx2tf
56
- sm_dir = "/tmp/saved_model_nano"
57
- onnx2tf.convert(input_onnx_file_path=onnx_path, output_folder_path=sm_dir,
58
- not_use_onnxsim=True, verbosity="error")
59
- print("onnx2tf done", flush=True)
60
- import tensorflow as tf
61
- conv = tf.lite.TFLiteConverter.from_saved_model(sm_dir)
62
- conv.optimizations = [tf.lite.Optimize.DEFAULT]
63
- tfl = conv.convert()
64
- tfl_path = "/tmp/model_nano_96.tflite"
65
- open(tfl_path, "wb").write(tfl)
66
- print("tflite dynamic-int8:", os.path.getsize(tfl_path) // 1024, "KB", flush=True)
67
-
68
- # 5) C header for the firmware (xxd -i compatible names used by node.ino)
69
- with open(tfl_path, "rb") as f:
70
- data = f.read()
71
- lines = []
72
- for i in range(0, len(data), 12):
73
- lines.append(" " + ", ".join(f"0x{b:02x}" for b in data[i:i + 12]) + ",")
74
- hdr = ("// generated by export_nano_tflite_job.py — do not edit\n"
75
- "#pragma once\n#include <cstdint>\n"
76
- "const unsigned char model_tflite[] = {\n" + "\n".join(lines) + "\n};\n"
77
- f"const unsigned int model_tflite_len = {len(data)};\n")
78
- hdr_path = "/tmp/model_data.h"
79
- open(hdr_path, "w").write(hdr)
80
- print("model_data.h:", os.path.getsize(hdr_path) // 1024, "KB", flush=True)
81
-
82
- # 6) upload artifacts
83
- for p, rp in [(tfl_path, "esp32/model_nano_96.tflite"),
84
- (onnx_path, "esp32/model_nano_96_dynamic_int8.onnx"),
85
- (hdr_path, "esp32/model_data.h")]:
86
- api.upload_file(path_or_fileobj=p, path_in_repo=rp, repo_id=CODE, repo_type="model",
87
- token=tok, commit_message=f"esp32: {rp} ({os.path.getsize(p)//1024} KB)")
88
- print("UPLOADED esp32 nano tflite + model_data.h", flush=True)
 
 
 
 
 
 
 
1
+ # /// script
2
+ # requires-python = ">=3.11"
3
+ # dependencies = ["torch", "numpy", "onnx", "onnxscript", "onnx2tf", "onnx-graphsurgeon", "onnxsim", "psutil", "ai-edge-litert", "sng4onnx", "pandas", "flatbuffers", "joblib", "termcolor", "requests", "tqdm", "packaging", "tensorflow-cpu", "tf-keras", "huggingface_hub>=0.30"]
4
+ # ///
5
+ """Convert the trained nano96 ESP32 checkpoint to an int8 TFLite model + model_data.h.
6
+
7
+ Waits for esp32/nano96_best.pt in Tonic/challenge5-acoustic-code (uploaded by the
8
+ train_nano job), then: best.pt -> ONNX (1x3x96x96, static) -> SavedModel (onnx2tf)
9
+ -> TFLite dynamic-range int8 -> C header model_data.h for firmware/node.
10
+
11
+ Uploads to code repo: esp32/model_nano_96.tflite, esp32/model_nano_96_dynamic_int8.onnx,
12
+ esp32/model_data.h
13
+ """
14
+ import os, sys, time, importlib.util
15
+ import numpy as np
16
+
17
+ tok = os.environ["HF_TOKEN"]
18
+ from huggingface_hub import HfApi, hf_hub_download
19
+ api = HfApi(token=tok)
20
+ CODE = "Tonic/challenge5-acoustic-code"
21
+
22
+ # 1) wait for the nano checkpoint (train_nano job uploads it at the end)
23
+ print("waiting for esp32/nano96_best.pt ...", flush=True)
24
+ ckpt = None
25
+ for i in range(90): # up to ~45 min
26
+ try:
27
+ ckpt = hf_hub_download(CODE, "esp32/nano96_best.pt", repo_type="model", token=tok)
28
+ break
29
+ except Exception:
30
+ time.sleep(30)
31
+ if ckpt is None:
32
+ raise SystemExit("nano96_best.pt never appeared")
33
+ print("ckpt:", os.path.getsize(ckpt) // 1024, "KB", flush=True)
34
+
35
+ import torch
36
+ ck = torch.load(ckpt, map_location="cpu", weights_only=False)
37
+ sd = ck.get("state_dict", ck) if isinstance(ck, dict) else ck
38
+
39
+ # 2) rebuild the nano arch from the trainer script in the same repo
40
+ tj = hf_hub_download(CODE, "train_nano_job.py", repo_type="model", token=tok)
41
+ spec = importlib.util.spec_from_file_location("train_nano_job", tj)
42
+ tn = importlib.util.module_from_spec(spec)
43
+ spec.loader.exec_module(tn)
44
+ model = tn.build_nano()
45
+ model.load_state_dict(sd, strict=True)
46
+ model.eval()
47
+ print("nano params:", sum(p.numel() for p in model.parameters()), flush=True)
48
+
49
+ # 3) ONNX (static 1x3x96x96 — TFLite-Micro friendly)
50
+ onnx_path = "/tmp/model_nano_96.onnx"
51
+ torch.onnx.export(model, torch.zeros(1, 3, 96, 96), onnx_path, opset_version=17, dynamo=False,
52
+ input_names=["spectrogram"], output_names=["wake_logit", "sound_class_logits"])
53
+ print("onnx exported", flush=True)
54
+
55
+ # 4) ONNX -> SavedModel -> TFLite dynamic int8
56
+ import onnx2tf
57
+ import onnx2tf.utils.common_functions as _cf
58
+ import onnx2tf.onnx2tf as _o2t
59
+ _cf.download_test_image_data = lambda *a, **k: np.zeros((1, 96, 96, 3), dtype=np.float32)
60
+ if hasattr(_o2t, "download_test_image_data"):
61
+ _o2t.download_test_image_data = _cf.download_test_image_data
62
+ sm_dir = "/tmp/saved_model_nano"
63
+ onnx2tf.convert(input_onnx_file_path=onnx_path, output_folder_path=sm_dir,
64
+ not_use_onnxsim=True, verbosity="error")
65
+ print("onnx2tf done", flush=True)
66
+ import tensorflow as tf
67
+ conv = tf.lite.TFLiteConverter.from_saved_model(sm_dir)
68
+ conv.optimizations = [tf.lite.Optimize.DEFAULT]
69
+ tfl = conv.convert()
70
+ tfl_path = "/tmp/model_nano_96.tflite"
71
+ open(tfl_path, "wb").write(tfl)
72
+ print("tflite dynamic-int8:", os.path.getsize(tfl_path) // 1024, "KB", flush=True)
73
+
74
+ # 5) C header for the firmware (xxd -i compatible names used by node.ino)
75
+ with open(tfl_path, "rb") as f:
76
+ data = f.read()
77
+ lines = []
78
+ for i in range(0, len(data), 12):
79
+ lines.append(" " + ", ".join(f"0x{b:02x}" for b in data[i:i + 12]) + ",")
80
+ hdr = ("// generated by export_nano_tflite_job.py — do not edit\n"
81
+ "#pragma once\n#include <cstdint>\n"
82
+ "const unsigned char model_tflite[] = {\n" + "\n".join(lines) + "\n};\n"
83
+ f"const unsigned int model_tflite_len = {len(data)};\n")
84
+ hdr_path = "/tmp/model_data.h"
85
+ open(hdr_path, "w").write(hdr)
86
+ print("model_data.h:", os.path.getsize(hdr_path) // 1024, "KB", flush=True)
87
+
88
+ # 6) upload artifacts
89
+ for p, rp in [(tfl_path, "esp32/model_nano_96.tflite"),
90
+ (onnx_path, "esp32/model_nano_96_dynamic_int8.onnx"),
91
+ (hdr_path, "esp32/model_data.h")]:
92
+ api.upload_file(path_or_fileobj=p, path_in_repo=rp, repo_id=CODE, repo_type="model",
93
+ token=tok, commit_message=f"esp32: {rp} ({os.path.getsize(p)//1024} KB)")
94
+ print("UPLOADED esp32 nano tflite + model_data.h", flush=True)