prosperious's picture
Add 1000 files (OCR_INDRA/.gitignore ...)
f9bc0c1 verified
Raw History Blame Contribute Delete
5.92 kB
"""Inspect rec.onnx graph structure and model.safetensors weights.
Answers:
- exact input/output shapes and names of the recognizer graph
- op sequence: conv attributes (strides, padding), activations, pooling
- whether BN is folded into convs
- parameter count from the safetensors and cross-check with ONNX initializers
"""
import json
from pathlib import Path
import numpy as np
import onnx
import safetensors.numpy
MODEL_DIR = Path("/mnt/c/Users/Dell/Desktop/OCR_INDRA/models/indra-ocr-v3.1.1")
def graph_summary(model):
g = model.graph
print("== INPUTS ==")
for i in g.input:
dims = [(d.dim_param or d.dim_value) for d in i.type.tensor_type.shape.dim]
print(f" {i.name}: dims={dims} dtype={i.type.tensor_type.elem_type}")
print("== OUTPUTS ==")
for o in g.output:
dims = [(d.dim_param or d.dim_value) for d in o.type.tensor_type.shape.dim]
print(f" {o.name}: dims={dims} dtype={o.type.tensor_type.elem_type}")
print(f"== {len(g.node)} nodes, {len(g.initializer)} initializers ==")
ops = {}
for n in g.node:
ops[n.op_type] = ops.get(n.op_type, 0) + 1
print(" op counts:", dict(sorted(ops.items(), key=lambda x: -x[1])))
imp = model.opset_import
for i in imp:
print(f" opset domain={i.domain!r} version={i.version}")
def shape_of(t):
dims = getattr(t, "dims", None) or []
return list(dims)
def initializer_map(model):
# large models may be external; these are small
return {t.name: onnx.numpy_helper.to_array(t) for t in model.graph.initializer}
def conv_attrs(model):
return {
n.name: {a.name: onnx.helper.get_attribute_value(a) for a in n.attribute}
for n in model.graph.node
if n.op_type in ("Conv",)
}
def tensors_of(model, prefix=""):
# ValueInfo for nursing intermediate shapes
return model.graph
def main():
print("### rec.onnx graph ###")
m = onnx.load(str(MODEL_DIR / "rec.onnx"))
graph_summary(m)
# print the full node sequence with attrs for the main ops
print("\n== node sequence (abbrev) ==")
for k, n in enumerate(m.graph.node):
attrs = {a.name: onnx.helper.get_attribute_value(a) for a in n.attribute}
if n.op_type == "Constant":
continue
ins = ", ".join(n.input[:2])
outs = ", ".join(n.output[:1])
if k < 200 and n.op_type in ("Conv", "ConvTranspose", "BatchNormalization",
"Relu", "LeakyRelu", "Sigmoid", "Multiply",
"Add", "ReduceMean", "Div", "GlobalAveragePool",
"AveragePool", "MaxPool", "Resize", "Squeeze",
"Unsqueeze", "Concat", "Transpose", "Reshape",
"Cast", "ArgMax", "ReduceMax", "ReduceArgMax",
"Gather", "Shape", "Less", "Greater", "Where"):
print(f" {k:3d} {n.op_type:22s} {outs} <- {ins} {dict(list(attrs.items())[:6])}")
elif k < 3 or n.op_type not in ("Cast", "Shape", "Gather", "Concat", "Unsqueeze",
"Squeeze", "Reshape", "Sub", "Mul", "Div",
"Slice", "Constant"):
pass
# count up to 60 first 20 lines then finalize
print("\n(total nodes:", len(m.graph.node), ")")
def main2():
print("\n### conv attributes (kernel/dilation/pad only) ###")
m = onnx.load(str(MODEL_DIR / "rec.onnx"))
for n in m.graph.node:
if n.op_type != "Conv":
continue
attrs = {a.name: onnx.helper.get_attribute_value(a) for a in n.attribute}
w = attrs.get("kernel_shape")
print(f" {n.name:20s} kernel={w} stride={attrs.get('strides')} "
f"dil={attrs.get('dilations')} pad_mode={attrs.get('auto_pad')} "
f"group={attrs.get('group', 1)} pads={attrs.get('pads')}")
def main3():
print("\n### safetensors ###")
with open(MODEL_DIR / "model.safetensors", "rb") as f:
header_size = int.from_bytes(f.read(8), "little")
header = json.loads(f.read(header_size))
tensors = {k: v for k, v in header.items() if k != "__metadata__"}
print(f"{len(tensors)} tensors")
total = 0
for name, info in list(tensors.items()):
n = int(np.prod(info["shape"])) if info["shape"] else 1
total += n
print(f" {name:44s} {str(info['shape']):24s} {info['dtype']:6s} {n}")
print(f"{total} total parameters")
def main4():
print("\n### crosscheck: rec.onnx initializer count/params ###")
m = onnx.load(str(MODEL_DIR / "rec.onnx"))
init = {t.name: onnx.numpy_helper.to_array(t) for t in m.graph.initializer}
tot = sum(int(np.prod(a.shape)) if a.ndim else 1 for a in init.values())
print(f"{len(init)} initializers, {tot} parameters")
# show the head and first block names:
sizes = sorted(
[(n, a.shape) for n, a in init.items() if a.ndim >= 1],
key=lambda x: -int(np.prod(x[1])) if len(x[1]) else 0,
)[:10]
for n, s in sizes:
print(f" big: {n:36s} {s}")
def main5():
print("\n### other models: graph I/O ###")
for f in ["det.onnx", "orient.onnx", "order.onnx", "script.onnx", "rec_cjk.onnx"]:
m = onnx.load(str(MODEL_DIR / f))
print(f"\n{f}:")
for i in m.graph.input:
dims = [(d.dim_param or d.dim_value) for d in i.type.tensor_type.shape.dim]
print(f" in {i.name}: dims={dims}")
for o in m.graph.output:
dims = [(d.dim_param or d.dim_value) for d in o.type.tensor_type.shape.dim]
print(f" out {o.name}: dims={dims}")
ops = {}
for n in m.graph.node:
ops[n.op_type] = ops.get(n.op_type, 0) + 1
print(" ops:", ops)
if __name__ == "__main__":
main()
main2()
main3()
main4()
main5()