Tonic commited on
Commit
d4d33f9
·
verified ·
1 Parent(s): 5e5009d

esp32: add onnx-graphsurgeon dep to export_job.py

Browse files
Files changed (1) hide show
  1. export_job.py +84 -84
export_job.py CHANGED
@@ -1,84 +1,84 @@
1
- # /// script
2
- # requires-python = ">=3.11"
3
- # dependencies = ["torch", "torchvision", "pillow", "numpy", "onnx", "onnxscript", "onnx2tf", "tensorflow-cpu", "tf-keras", "huggingface_hub>=0.30", "timm", "datasets>=4.0"]
4
- # ///
5
- """HF-job credibility gate: convert the wave-3 winner checkpoint (best.pt) to a
6
- quantized TFLite model for ESP32 TFLite-Micro and upload the artifact to the code repo.
7
-
8
- Path: best.pt -> ONNX (torch.onnx) -> SavedModel (onnx2tf, float) -> TFLite with
9
- dynamic-range int8 weights (Optimize.DEFAULT). TFLite-Micro runs this fine on ESP32.
10
- """
11
- import os, sys, io
12
- import numpy as np
13
- import torch
14
- from huggingface_hub import HfApi, hf_hub_download
15
-
16
- tok = os.environ["HF_TOKEN"]
17
- api = HfApi(token=tok)
18
- ckpt = hf_hub_download("Tonic/challenge5-acoustic-squeezenet1_1-s7", "best.pt",
19
- repo_type="model", token=tok)
20
- print("ckpt:", os.path.getsize(ckpt) // 1024, "KB", flush=True)
21
-
22
- here = os.path.dirname(os.path.abspath(__file__))
23
- print("step: torch.load", flush=True)
24
- ck = torch.load(ckpt, map_location="cpu", weights_only=False)
25
- print("step: ckpt loaded", flush=True)
26
- sd = ck.get("state_dict", ck) if isinstance(ck, dict) else ck
27
- arch = ck.get("arch", "squeezenet1_1") if isinstance(ck, dict) else "squeezenet1_1"
28
-
29
- import importlib.util
30
- jt_path = hf_hub_download("Tonic/challenge5-acoustic-code", "job_train.py", repo_type="model", token=tok)
31
- spec = importlib.util.spec_from_file_location("job_train", jt_path)
32
- jt = importlib.util.module_from_spec(spec)
33
- print("step: exec job_train", flush=True)
34
- spec.loader.exec_module(jt)
35
- print("step: job_train imported", flush=True)
36
- backbone, feat = jt.build_backbone(arch, pretrained=False)
37
- class WakeCNN(torch.nn.Module):
38
- def __init__(self, b, f):
39
- super().__init__()
40
- self.backbone = b
41
- self.head_wake = torch.nn.Linear(f, 1)
42
- self.head_class = torch.nn.Linear(f, 4)
43
- def forward(self, x):
44
- f = self.backbone(x)
45
- return self.head_wake(f).squeeze(-1), self.head_class(f)
46
- model = WakeCNN(backbone, feat)
47
- sd = {k: v for k, v in sd.items() if k.startswith("backbone.") or k.startswith("head_")}
48
- print("step: load_state_dict", flush=True)
49
- model.load_state_dict(sd, strict=True)
50
- print("step: state loaded", flush=True)
51
- model.eval()
52
- print("arch:", arch, "params:", sum(p.numel() for p in model.parameters()), flush=True)
53
-
54
- SIZE = 224
55
- onnx_path = "/tmp/model_full_224.onnx"
56
- torch.onnx.export(model, torch.zeros(1, 3, SIZE, SIZE), onnx_path, opset_version=17, dynamo=False,
57
- input_names=["spectrogram"], output_names=["wake_logit", "sound_class_logits"],
58
- dynamic_axes={"spectrogram": {0: "batch"}})
59
- print("onnx exported", flush=True)
60
-
61
- print("step: import onnx2tf", flush=True)
62
- import onnx2tf
63
- sm_dir = "/tmp/saved_model"
64
- onnx2tf.convert(input_onnx_file_path=onnx_path, output_folder_path=sm_dir,
65
- not_use_onnxsim=True, verbosity="error")
66
- print("onnx2tf done", flush=True)
67
-
68
- print("step: import tensorflow", flush=True)
69
- import tensorflow as tf
70
- print("step: tf imported", flush=True)
71
- conv = tf.lite.TFLiteConverter.from_saved_model(sm_dir)
72
- conv.optimizations = [tf.lite.Optimize.DEFAULT] # dynamic-range int8 weights
73
- tfl = conv.convert()
74
- out = "/tmp/model_full_224_dynamic_int8.tflite"
75
- open(out, "wb").write(tfl)
76
- print("tflite dynamic-int8:", os.path.getsize(out) // 1024, "KB", flush=True)
77
-
78
- api.upload_file(path_or_fileobj=out, path_in_repo="esp32/model_full_224_dynamic_int8.tflite",
79
- repo_id="Tonic/challenge5-acoustic-code", repo_type="model", token=tok,
80
- commit_message=f"esp32: dynamic-int8 tflite of squeezenet1_1-s7 ({os.path.getsize(out)//1024} KB)")
81
- api.upload_file(path_or_fileobj=onnx_path, path_in_repo="esp32/model_full_224.onnx",
82
- repo_id="Tonic/challenge5-acoustic-code", repo_type="model", token=tok,
83
- commit_message="esp32: onnx of squeezenet1_1-s7")
84
- print("UPLOADED esp32/ artifacts", flush=True)
 
1
+ # /// script
2
+ # requires-python = ">=3.11"
3
+ # dependencies = ["torch", "torchvision", "pillow", "numpy", "onnx", "onnxscript", "onnx2tf", "onnx-graphsurgeon", "onnxsim", "tensorflow-cpu", "tf-keras", "huggingface_hub>=0.30", "timm", "datasets>=4.0"]
4
+ # ///
5
+ """HF-job credibility gate: convert the wave-3 winner checkpoint (best.pt) to a
6
+ quantized TFLite model for ESP32 TFLite-Micro and upload the artifact to the code repo.
7
+
8
+ Path: best.pt -> ONNX (torch.onnx) -> SavedModel (onnx2tf, float) -> TFLite with
9
+ dynamic-range int8 weights (Optimize.DEFAULT). TFLite-Micro runs this fine on ESP32.
10
+ """
11
+ import os, sys, io
12
+ import numpy as np
13
+ import torch
14
+ from huggingface_hub import HfApi, hf_hub_download
15
+
16
+ tok = os.environ["HF_TOKEN"]
17
+ api = HfApi(token=tok)
18
+ ckpt = hf_hub_download("Tonic/challenge5-acoustic-squeezenet1_1-s7", "best.pt",
19
+ repo_type="model", token=tok)
20
+ print("ckpt:", os.path.getsize(ckpt) // 1024, "KB", flush=True)
21
+
22
+ here = os.path.dirname(os.path.abspath(__file__))
23
+ print("step: torch.load", flush=True)
24
+ ck = torch.load(ckpt, map_location="cpu", weights_only=False)
25
+ print("step: ckpt loaded", flush=True)
26
+ sd = ck.get("state_dict", ck) if isinstance(ck, dict) else ck
27
+ arch = ck.get("arch", "squeezenet1_1") if isinstance(ck, dict) else "squeezenet1_1"
28
+
29
+ import importlib.util
30
+ jt_path = hf_hub_download("Tonic/challenge5-acoustic-code", "job_train.py", repo_type="model", token=tok)
31
+ spec = importlib.util.spec_from_file_location("job_train", jt_path)
32
+ jt = importlib.util.module_from_spec(spec)
33
+ print("step: exec job_train", flush=True)
34
+ spec.loader.exec_module(jt)
35
+ print("step: job_train imported", flush=True)
36
+ backbone, feat = jt.build_backbone(arch, pretrained=False)
37
+ class WakeCNN(torch.nn.Module):
38
+ def __init__(self, b, f):
39
+ super().__init__()
40
+ self.backbone = b
41
+ self.head_wake = torch.nn.Linear(f, 1)
42
+ self.head_class = torch.nn.Linear(f, 4)
43
+ def forward(self, x):
44
+ f = self.backbone(x)
45
+ return self.head_wake(f).squeeze(-1), self.head_class(f)
46
+ model = WakeCNN(backbone, feat)
47
+ sd = {k: v for k, v in sd.items() if k.startswith("backbone.") or k.startswith("head_")}
48
+ print("step: load_state_dict", flush=True)
49
+ model.load_state_dict(sd, strict=True)
50
+ print("step: state loaded", flush=True)
51
+ model.eval()
52
+ print("arch:", arch, "params:", sum(p.numel() for p in model.parameters()), flush=True)
53
+
54
+ SIZE = 224
55
+ onnx_path = "/tmp/model_full_224.onnx"
56
+ torch.onnx.export(model, torch.zeros(1, 3, SIZE, SIZE), onnx_path, opset_version=17, dynamo=False,
57
+ input_names=["spectrogram"], output_names=["wake_logit", "sound_class_logits"],
58
+ dynamic_axes={"spectrogram": {0: "batch"}})
59
+ print("onnx exported", flush=True)
60
+
61
+ print("step: import onnx2tf", flush=True)
62
+ import onnx2tf
63
+ sm_dir = "/tmp/saved_model"
64
+ onnx2tf.convert(input_onnx_file_path=onnx_path, output_folder_path=sm_dir,
65
+ not_use_onnxsim=True, verbosity="error")
66
+ print("onnx2tf done", flush=True)
67
+
68
+ print("step: import tensorflow", flush=True)
69
+ import tensorflow as tf
70
+ print("step: tf imported", flush=True)
71
+ conv = tf.lite.TFLiteConverter.from_saved_model(sm_dir)
72
+ conv.optimizations = [tf.lite.Optimize.DEFAULT] # dynamic-range int8 weights
73
+ tfl = conv.convert()
74
+ out = "/tmp/model_full_224_dynamic_int8.tflite"
75
+ open(out, "wb").write(tfl)
76
+ print("tflite dynamic-int8:", os.path.getsize(out) // 1024, "KB", flush=True)
77
+
78
+ api.upload_file(path_or_fileobj=out, path_in_repo="esp32/model_full_224_dynamic_int8.tflite",
79
+ repo_id="Tonic/challenge5-acoustic-code", repo_type="model", token=tok,
80
+ commit_message=f"esp32: dynamic-int8 tflite of squeezenet1_1-s7 ({os.path.getsize(out)//1024} KB)")
81
+ api.upload_file(path_or_fileobj=onnx_path, path_in_repo="esp32/model_full_224.onnx",
82
+ repo_id="Tonic/challenge5-acoustic-code", repo_type="model", token=tok,
83
+ commit_message="esp32: onnx of squeezenet1_1-s7")
84
+ print("UPLOADED esp32/ artifacts", flush=True)