File size: 6,716 Bytes
ce847d4 |
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 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 |
"""Unlock model_33 (LineLayout) — replace OneOCRFeatureExtract with Conv 1x1.
Model 33 architecture:
data[1,3,60,W] → CNN(conv1-conv5, 256ch) → OneOCRFeatureExtract(256→16ch)
→ Transpose → Reshape → MLP(64→32→2) → line_layout predictions
The config blob (16548 bytes) structure (big-endian float32):
W[256×16] = 4096 floats (stored as [in_ch, out_ch] or [out_ch, in_ch])
b[16] = 16 floats
metadata[25] = remaining floats
"""
import onnx
from onnx import numpy_helper, helper, TensorProto
import numpy as np
from pathlib import Path
import copy
import onnxruntime as ort
models_dir = Path("oneocr_extracted/onnx_models")
output_dir = Path("oneocr_extracted/onnx_models_unlocked")
output_dir.mkdir(exist_ok=True)
model_path = list(models_dir.glob("model_33_*"))[0]
model = onnx.load(str(model_path))
# Extract config blob
for init in model.graph.initializer:
if init.name == "feature/config":
blob = bytes(init.string_data[0] if init.string_data else init.raw_data)
break
be_arr = np.frombuffer(blob, dtype='>f4').copy()
print(f"Config blob: {len(be_arr)} big-endian float32 values")
# Extract W[256×16] and b[16]
# Pattern: first 4096 = W, next 16 = b, remaining = metadata
in_ch = 256
out_ch = 16
W_flat = be_arr[:in_ch * out_ch]
b = be_arr[in_ch * out_ch:in_ch * out_ch + out_ch]
metadata = be_arr[in_ch * out_ch + out_ch:]
print(f"W: {in_ch*out_ch} floats, range=[{W_flat.min():.4f}, {W_flat.max():.4f}]")
print(f"b: {out_ch} floats, range=[{b.min():.4f}, {b.max():.4f}]")
print(f"Metadata: {len(metadata)} floats")
print(f"Metadata: {metadata}")
# Check dimension markers in metadata
for i, v in enumerate(metadata):
if v > 10 or v < -10:
print(f" metadata[{i}] = {v}")
# Try both weight arrangements with Conv 1x1
# Conv weight shape in ONNX: [out_ch, in_ch, kH, kW]
for arrangement_name, W_conv in [
("W[out,in]=[16,256]", W_flat.reshape(out_ch, in_ch, 1, 1)),
("W[in,out]=[256,16]→T", W_flat.reshape(in_ch, out_ch).T.reshape(out_ch, in_ch, 1, 1)),
]:
print(f"\n--- Testing {arrangement_name} ---")
new_model = copy.deepcopy(model)
# Replace initializers
new_inits = [init for init in new_model.graph.initializer if init.name != "feature/config"]
new_inits.append(numpy_helper.from_array(W_conv.astype(np.float32), name="fe_conv_weight"))
new_inits.append(numpy_helper.from_array(b.astype(np.float32), name="fe_conv_bias"))
del new_model.graph.initializer[:]
new_model.graph.initializer.extend(new_inits)
# Replace OneOCRFeatureExtract with Conv
fe_node = None
for node in new_model.graph.node:
if node.op_type == "OneOCRFeatureExtract":
fe_node = node
break
fe_input = fe_node.input[0] # 'conv5_2_relu'
fe_output = fe_node.output[0] # 'oneocr_feature'
new_nodes = []
for node in new_model.graph.node:
if node.op_type == "OneOCRFeatureExtract":
conv_node = helper.make_node(
"Conv",
inputs=[fe_input, "fe_conv_weight", "fe_conv_bias"],
outputs=[fe_output],
kernel_shape=[1, 1],
strides=[1, 1],
pads=[0, 0, 0, 0],
)
new_nodes.append(conv_node)
else:
new_nodes.append(node)
del new_model.graph.node[:]
new_model.graph.node.extend(new_nodes)
# Clean up
new_inputs = [inp for inp in new_model.graph.input if inp.name != "feature/config"]
del new_model.graph.input[:]
new_model.graph.input.extend(new_inputs)
new_opsets = [op for op in new_model.opset_import if op.domain != "com.microsoft.oneocr"]
del new_model.opset_import[:]
new_model.opset_import.extend(new_opsets)
# Save and test
out_name = f"model_33_{arrangement_name.split('[')[0].strip()}.onnx"
out_path = output_dir / out_name
onnx.save(new_model, str(out_path))
try:
sess = ort.InferenceSession(str(out_path))
# Test with 60×100 image (width=100)
test_data = np.random.randn(1, 3, 60, 100).astype(np.float32) * 0.1
outputs = sess.run(None, {"data": test_data})
output_names = [o.name for o in sess.get_outputs()]
for name, out in zip(output_names, outputs):
print(f" {name}: shape={out.shape}, range=[{out.min():.4f}, {out.max():.4f}]")
# Also test with width=200
test_data2 = np.random.randn(1, 3, 60, 200).astype(np.float32) * 0.1
outputs2 = sess.run(None, {"data": test_data2})
for name, out in zip(output_names, outputs2):
print(f" (w=200) {name}: shape={out.shape}")
except Exception as e:
print(f" FAILED: {e}")
import traceback
traceback.print_exc()
# Save the best working version as the official unlocked model
print(f"\n--- Saving official unlocked model ---")
# Use W[in,out]=[256,16]→T arrangement (gives reasonable output ranges)
W_conv = W_flat.reshape(in_ch, out_ch).T.reshape(out_ch, in_ch, 1, 1)
new_model = copy.deepcopy(model)
new_inits = [init for init in new_model.graph.initializer if init.name != "feature/config"]
new_inits.append(numpy_helper.from_array(W_conv.astype(np.float32), name="fe_conv_weight"))
new_inits.append(numpy_helper.from_array(b.astype(np.float32), name="fe_conv_bias"))
del new_model.graph.initializer[:]
new_model.graph.initializer.extend(new_inits)
new_nodes = []
for node in new_model.graph.node:
if node.op_type == "OneOCRFeatureExtract":
conv_node = helper.make_node(
"Conv",
inputs=["conv5_2_relu", "fe_conv_weight", "fe_conv_bias"],
outputs=["oneocr_feature"],
kernel_shape=[1, 1],
)
new_nodes.append(conv_node)
else:
new_nodes.append(node)
del new_model.graph.node[:]
new_model.graph.node.extend(new_nodes)
new_inputs = [inp for inp in new_model.graph.input if inp.name != "feature/config"]
del new_model.graph.input[:]
new_model.graph.input.extend(new_inputs)
new_opsets = [op for op in new_model.opset_import if op.domain != "com.microsoft.oneocr"]
del new_model.opset_import[:]
new_model.opset_import.extend(new_opsets)
final_path = output_dir / "model_33_ir7_unknown_857KB_unlocked.onnx"
onnx.save(new_model, str(final_path))
print(f"Saved: {final_path}")
try:
sess = ort.InferenceSession(str(final_path))
test = np.random.randn(1, 3, 60, 100).astype(np.float32) * 0.1
outs = sess.run(None, {"data": test})
for name, out in zip([o.name for o in sess.get_outputs()], outs):
print(f" ✓ {name}: {out.shape}")
except Exception as e:
print(f" ✗ {e}")
|