netron/tools/onnx-script.py

from __future__ import unicode_literals

import onnx

import json
import io
import os
import re
import sys

from onnx import defs
from onnx.defs import OpSchema
from onnx.backend.test.case import collect_snippets

snippets = collect_snippets()

categories = {
    'Constant': 'Constant',

    'Conv': 'Layer',
    'ConvTranspose': 'Layer',
    'FC': 'Layer',
    'RNN': 'Layer',
    'LSTM': 'Layer',
    'GRU': 'Layer',
    'Gemm': 'Layer',

    'Dropout': 'Dropout',

    'Elu': 'Activation',
    'HardSigmoid': 'Activation',
    'LeakyRelu': 'Activation',
    'PRelu': 'Activation',
    'ThresholdedRelu': 'Activation',
    'Relu': 'Activation',
    'Selu': 'Activation',
    'Sigmoid': 'Activation',
    'Tanh': 'Activation',
    'LogSoftmax': 'Activation',
    'Softmax': 'Activation',
    'Softplus': 'Activation',
    'Softsign': 'Activation',

    'BatchNormalization': 'Normalization',
    'InstanceNormalization': 'Normalization',
    'LpNormalization': 'Normalization',
    'LRN': 'Normalization',

    'Flatten': 'Shape',
    'Reshape': 'Shape',
    'Tile': 'Shape',

    'Xor': 'Logic',
    'Not': 'Logic',
    'Or': 'Logic',
    'Less': 'Logic',
    'And': 'Logic',
    'Greater': 'Logic',
    'Equal': 'Logic',

    'AveragePool': 'Pool',
    'GlobalAveragePool': 'Pool',
    'GlobalLpPool': 'Pool',
    'GlobalMaxPool': 'Pool',
    'LpPool': 'Pool',
    'MaxPool': 'Pool',
    'MaxRoiPool': 'Pool',

    'Concat': 'Tensor',
    'Slice': 'Tensor',
    'Split': 'Tensor',
    'Pad': 'Tensor',

    'ImageScaler': 'Data',
    'Crop': 'Data',
    'Upsample': 'Data',

    'Transpose': 'Transform',
    'Gather': 'Transform',
    'Unsqueeze': 'Transform',
    'Squeeze': 'Transform',
}

attribute_type_table = {
    'undefined': None,
    'float': 'float32', 'int': 'int64', 'string': 'string', 'tensor': 'tensor', 'graph': 'graph',
    'floats': 'float32[]', 'ints': 'int64[]', 'strings': 'string[]', 'tensors': 'tensor[]', 'graphs': 'graph[]',
}

def generate_json_attr_type(type):
    assert isinstance(type, OpSchema.AttrType)
    s = str(type)
    s = s[s.rfind('.')+1:].lower()
    if s in attribute_type_table:
        return attribute_type_table[s]
    return None

def generate_json_attr_default_value(attr_value):
    if not str(attr_value):
        return None
    if attr_value.HasField('i'):
        return attr_value.i
    if attr_value.HasField('s'):
        return attr_value.s.decode('utf8')
    if attr_value.HasField('f'):
        return attr_value.f
    return None

def generate_json_support_level_name(support_level):
    assert isinstance(support_level, OpSchema.SupportType)
    s = str(support_level)
    return s[s.rfind('.')+1:].lower()

def generate_json_types(types):
    r = []
    for type in types:
        r.append(type)
    r = sorted(r)
    return r

def format_range(value):
    if value == 2147483647:
        return '∞'
    return str(value)

def format_description(description):
    def replace_line(match):
        link = match.group(1)
        url = match.group(2)
        if not url.startswith("http://") and not url.startswith("https://"):
            url = "https://github.com/onnx/onnx/blob/master/docs/" + url
        return "[" + link + "](" + url + ")";
    description = re.sub("\\[(.+)\\]\\(([^ ]+?)( \"(.+)\")?\\)", replace_line, description)
    return description

def generate_json(schemas, json_file):
    json_root = []
    for schema in schemas:
        json_schema = {}
        if schema.domain:
            json_schema['domain'] = schema.domain
        else:
            json_schema['domain'] = 'ai.onnx'
        json_schema['since_version'] = schema.since_version
        json_schema['support_level'] = generate_json_support_level_name(schema.support_level)
        if schema.doc:
            json_schema['description'] = format_description(schema.doc.lstrip())
        if schema.inputs:
            json_schema['inputs'] = []
            for input in schema.inputs:
                json_input = {}
                json_input['name'] = input.name
                json_input['description'] = format_description(input.description)
                json_input['type'] = input.typeStr
                if input.option == OpSchema.FormalParameterOption.Optional:
                    json_input['option'] = 'optional'
                elif input.option == OpSchema.FormalParameterOption.Variadic:
                    json_input['option'] = 'variadic'
                json_schema['inputs'].append(json_input)
        json_schema['min_input'] = schema.min_input
        json_schema['max_input'] = schema.max_input
        if schema.outputs:
            json_schema['outputs'] = []
            for output in schema.outputs:
                json_output = {}
                json_output['name'] = output.name
                json_output['description'] = format_description(output.description)
                json_output['type'] = output.typeStr
                if output.option == OpSchema.FormalParameterOption.Optional:
                    json_output['option'] = 'optional'
                elif output.option == OpSchema.FormalParameterOption.Variadic:
                    json_output['option'] = 'variadic'
                json_schema['outputs'].append(json_output)
        json_schema['min_output'] = schema.min_output
        json_schema['max_output'] = schema.max_output
        if schema.min_input != schema.max_input:
            json_schema['inputs_range'] = format_range(schema.min_input) + ' - ' + format_range(schema.max_input);
        if schema.min_output != schema.max_output:
            json_schema['outputs_range'] = format_range(schema.min_output) + ' - ' + format_range(schema.max_output);
        if schema.attributes:
            json_schema['attributes'] = []
            for _, attribute in sorted(schema.attributes.items()):
                json_attribute = {}
                json_attribute['name'] = attribute.name
                json_attribute['description'] = format_description(attribute.description)
                attribute_type = generate_json_attr_type(attribute.type)
                if attribute_type:
                    json_attribute['type'] = attribute_type
                elif 'type' in json_attribute:
                    del json_attribute['type']
                json_attribute['required'] = attribute.required
                default_value = generate_json_attr_default_value(attribute.default_value)
                if default_value:
                    json_attribute['default'] = default_value
                json_schema['attributes'].append(json_attribute)
        if schema.type_constraints:
            json_schema['type_constraints'] = []
            for type_constraint in schema.type_constraints:
                json_schema['type_constraints'].append({
                    'description': type_constraint.description,
                    'type_param_str': type_constraint.type_param_str,
                    'allowed_type_strs': type_constraint.allowed_type_strs
                })
        if schema.name in snippets:
            json_schema['examples'] = []
            for summary, code in sorted(snippets[schema.name]):
                json_schema['examples'].append({
                    'summary': summary,
                    'code': code
                })
        if schema.name in categories:
            json_schema['category'] = categories[schema.name]
        json_root.append({
            'name': schema.name,
            'schema': json_schema 
        })
    with io.open(json_file, 'w', newline='') as fout:
        json_root = json.dumps(json_root, sort_keys=True, indent=2)
        for line in json_root.splitlines():
            line = line.rstrip()
            if sys.version_info[0] < 3:
                line = unicode(line)
            fout.write(line)
            fout.write('\n')

def metadata():
    schemas = defs.get_all_schemas_with_history()
    schemas = sorted(schemas, key=lambda schema: schema.name)
    json_file = os.path.join(os.path.dirname(__file__), '../src/onnx-metadata.json')
    generate_json(schemas, json_file)

def convert():
    def pip_import(package):
        import importlib
        try:
            importlib.import_module(package)
        except:
            import subprocess
            subprocess.call([ 'pip', 'install', '--quiet', package ])
        finally:
            globals()[package] = importlib.import_module(package)
    file = sys.argv[2]
    base, extension = os.path.splitext(file)
    if extension == '.mlmodel':
        pip_import('coremltools')
        import onnxmltools
        coreml_model = coremltools.utils.load_spec(file)
        onnx_model = onnxmltools.convert.convert_coreml(coreml_model)
        onnxmltools.utils.save_model(onnx_model, base + '.onnx')
    elif extension == '.h5':
        pip_import('tensorflow')
        pip_import('keras')
        import onnxmltools
        keras_model = keras.models.load_model(file)
        onnx_model = onnxmltools.convert.convert_keras(keras_model)
        onnxmltools.utils.save_model(onnx_model, base + '.onnx')
    elif extension == '.pkl':
        pip_import('sklearn')
        import onnxmltools
        sklearn_model = sklearn.externals.joblib.load(file)
        onnx_model = onnxmltools.convert.convert_sklearn(sklearn_model)
        onnxmltools.utils.save_model(onnx_model, base + '.onnx')
    base, extension = os.path.splitext(file)
    if extension == '.onnx':
        import onnx
        from google.protobuf import text_format
        onnx_model = onnx.load(file)
        text = text_format.MessageToString(onnx_model)
        with open(base + '.pbtxt', 'w') as text_file:
            text_file.write(text)

def optimize():
    import onnx
    from onnx import optimizer
    file = sys.argv[2]
    base = os.path.splitext(file)
    onnx_model = onnx.load(file)
    passes = optimizer.get_available_passes()
    optimized_model = optimizer.optimize(onnx_model, passes)
    onnx.save(optimized_model, base + '.optimized.onnx')

def infer():
    import onnx
    import onnx.shape_inference
    from onnx import shape_inference
    file = sys.argv[2]
    base = os.path.splitext(file)[0]
    onnx_model = onnx.load(base + '.onnx')
    onnx_model = onnx.shape_inference.infer_shapes(onnx_model)
    onnx.save(onnx_model, base + '.shape.onnx')

if __name__ == '__main__':
    command_table = { 'metadata': metadata, 'convert': convert, 'optimize': optimize, 'infer': infer }
    command = sys.argv[1]
    command_table[command]()