File size: 4,458 Bytes
c031767 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 | # /// script
# requires-python = ">=3.11"
# dependencies = ["torch", "torchvision", "pillow", "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", "timm", "datasets>=4.0"]
# ///
"""HF-job credibility gate: convert the wave-3 winner checkpoint (best.pt) to a
quantized TFLite model for ESP32 TFLite-Micro and upload the artifact to the code repo.
Path: best.pt -> ONNX (torch.onnx) -> SavedModel (onnx2tf, float) -> TFLite with
dynamic-range int8 weights (Optimize.DEFAULT). TFLite-Micro runs this fine on ESP32.
"""
import os, sys, io
import numpy as np
import torch
from huggingface_hub import HfApi, hf_hub_download
tok = os.environ["HF_TOKEN"]
api = HfApi(token=tok)
ckpt = hf_hub_download("Tonic/challenge5-acoustic-squeezenet1_1-s7", "best.pt",
repo_type="model", token=tok)
print("ckpt:", os.path.getsize(ckpt) // 1024, "KB", flush=True)
here = os.path.dirname(os.path.abspath(__file__))
print("step: torch.load", flush=True)
ck = torch.load(ckpt, map_location="cpu", weights_only=False)
print("step: ckpt loaded", flush=True)
sd = ck.get("state_dict", ck) if isinstance(ck, dict) else ck
arch = ck.get("arch", "squeezenet1_1") if isinstance(ck, dict) else "squeezenet1_1"
import importlib.util
jt_path = hf_hub_download("Tonic/challenge5-acoustic-code", "job_train.py", repo_type="model", token=tok)
spec = importlib.util.spec_from_file_location("job_train", jt_path)
jt = importlib.util.module_from_spec(spec)
print("step: exec job_train", flush=True)
spec.loader.exec_module(jt)
print("step: job_train imported", flush=True)
backbone, feat = jt.build_backbone(arch, pretrained=False)
class WakeCNN(torch.nn.Module):
def __init__(self, b, f):
super().__init__()
self.backbone = b
self.head_wake = torch.nn.Linear(f, 1)
self.head_class = torch.nn.Linear(f, 4)
def forward(self, x):
f = self.backbone(x)
return self.head_wake(f).squeeze(-1), self.head_class(f)
model = WakeCNN(backbone, feat)
sd = {k: v for k, v in sd.items() if k.startswith("backbone.") or k.startswith("head_")}
print("step: load_state_dict", flush=True)
model.load_state_dict(sd, strict=True)
print("step: state loaded", flush=True)
model.eval()
print("arch:", arch, "params:", sum(p.numel() for p in model.parameters()), flush=True)
SIZE = 224
onnx_path = "/tmp/model_full_224.onnx"
torch.onnx.export(model, torch.zeros(1, 3, SIZE, SIZE), onnx_path, opset_version=17, dynamo=False,
input_names=["spectrogram"], output_names=["wake_logit", "sound_class_logits"],
dynamic_axes={"spectrogram": {0: "batch"}})
print("onnx exported", flush=True)
print("step: import onnx2tf", flush=True)
import onnx2tf
import onnx2tf.utils.common_functions as _cf
import onnx2tf.onnx2tf as _o2t
_cf.download_test_image_data = lambda *a, **k: np.zeros((1, 224, 224, 3), dtype=np.float32)
if hasattr(_o2t, "download_test_image_data"):
_o2t.download_test_image_data = _cf.download_test_image_data
sm_dir = "/tmp/saved_model"
onnx2tf.convert(input_onnx_file_path=onnx_path, output_folder_path=sm_dir,
not_use_onnxsim=True, verbosity="error")
print("onnx2tf done", flush=True)
print("step: import tensorflow", flush=True)
import tensorflow as tf
print("step: tf imported", flush=True)
conv = tf.lite.TFLiteConverter.from_saved_model(sm_dir)
conv.optimizations = [tf.lite.Optimize.DEFAULT] # dynamic-range int8 weights
tfl = conv.convert()
out = "/tmp/model_full_224_dynamic_int8.tflite"
open(out, "wb").write(tfl)
print("tflite dynamic-int8:", os.path.getsize(out) // 1024, "KB", flush=True)
api.upload_file(path_or_fileobj=out, path_in_repo="esp32/model_full_224_dynamic_int8.tflite",
repo_id="Tonic/challenge5-acoustic-code", repo_type="model", token=tok,
commit_message=f"esp32: dynamic-int8 tflite of squeezenet1_1-s7 ({os.path.getsize(out)//1024} KB)")
api.upload_file(path_or_fileobj=onnx_path, path_in_repo="esp32/model_full_224.onnx",
repo_id="Tonic/challenge5-acoustic-code", repo_type="model", token=tok,
commit_message="esp32: onnx of squeezenet1_1-s7")
print("UPLOADED esp32/ artifacts", flush=True)
|