使用 Core ML 转换器 为设备获取模型

2:58 - TensorFlow conversion using tfcoreml

# pip install tfcoreml
# pip install coremltools

import tfcoreml
mlmodel = tfcoreml.convert(tf_model, mlmodel_path="/tmp/model.mlmodel")

3:16 - New TensorFlow model conversion

# pip install coremltools

import coremltools as ct
mlmodel = ct.convert(tf_model)

3:57 - ONNX conversion to Core ML

# pip install onnx-coreml
# pip install coremltools

import onnx_coreml
onnx_model = torch.export(torch_model)
mlmodel = onnx_coreml.convert(onnx_model)

4:28 - New PyTorch model conversion

# pip install coremltools

import coremltools as ct 
mlmodel = ct.convert(torch_script_model)

4:52 - Unified conversion API

import coremltools as ct

model = ct.convert(
  source_model # TF1, TF2, or PyTorch model
)

6:42 - Demo 1: TF2 conversion

import coremltools as ct 
import tensorflow as tf

tf_model = tf.keras.applications.MobileNet()
mlmodel = ct.convert(tf_model)

7:41 - Demo 1: Pytorch conversion

import coremltools as ct 
import torch
import torchvision 

torch_model = torchvision.models.mobilenet_v2()

torch_model.eval()
example_input = torch.rand(1, 3, 256, 256)
traced_model = torch.jit.trace(torch_model, example_input)

mlmodel = ct.convert(traced_model,
                    inputs=[ct.TensorType(shape=example_input.shape)])

print(mlmodel)

spec = mlmodel.get_spec()
ct.utils.rename_feature(spec, "input.1", "myInputName")
ct.utils.rename_feature(spec, "1648", "myOutputName")
mlmodel = ct.models.MLModel(spec)

print(mlmodel)

10:37 - Demo 1 : TF1 conversion

import coremltools as ct 
import tensorflow as tf

mlmodel = ct.convert("mobilenet_frozen_graph.pb",
                    inputs=[ct.ImageType(bias=[-1,-1,-1], scale=1/127.0)],
                    classifier_config=ct.ClassifierConfig("labels.txt"))

mlmodel.short_description = 'An image classifier'
mlmodel.license = 'Apache 2.0'
mlmodel.author = "Original Paper: A. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, "\
                 "T. Weyand, M. Andreetto, H. Adam"

mlmodel.save("mobilenet.mlmodel")

13:33 - Demo 1 Recap: Using coremltools convert

import coremltools as ct
mlmodel = ct.convert("./tf1_inception_model.pb")
mlmodel = ct.convert("./tf2_inception_model.h5")
mlmodel = ct.convert(torch_model, inputs=[ct.TensorType(shape=example_input.shape)])

15:45 - Converting a Deep Speech model

import numpy as np
import IPython.display as ipd

import coremltools as ct

### Pretrained models and chekpoints are available on the repository: 
https://github.com/mozilla/DeepSpeech

!python DeepSpeech.py --export_dir /tmp --checkpoint_dir ./deepspeech-0.7.1-checkpoint --alphabet_config_path=alphabet.txt --scorer_path=kenlm.scorer >/dev/null 2>&1

ls /tmp/*.pb

tf_model = "/tmp/output_graph.pb"

from demo_utils import inspect_tf_outputs

inspect_tf_outputs(tf_model)

outputs = ["logits", "new_state_c", "new_state_h"]

mlmodel = ct.convert(tf_model, outputs=outputs)

audiofile = "./audio_sample_16bit_mono_16khz.wav"

ipd.Audio(audiofile) 

from demo_utils import preprocessing, postprocessing

mfccs = preprocessing(audiofile)

mfccs.shape

from demo_utils import inspect_inputs

inspect_inputs(mlmodel, tf_model)

start = 0 
step = 16

max_time_steps = mfccs.shape[1]

logits_sequence = []

input_dict = {}

input_dict["input_lengths"]  = np.array([step]).astype(np.float32)

input_dict["previous_state_c"] = np.zeros([1, 2048]).astype(np.float32) # Initializing cell state 
input_dict["previous_state_h"] = np.zeros([1, 2048]).astype(np.float32) # Initializing hidden state 


print("Transcription: \n")

while (start + step) < max_time_steps:
    input_dict["input_node"] = mfccs[:, start:(start + step), :, :]
    
    # Evaluation
    preds = mlmodel.predict(input_dict)
    
    
    start += step
    logits_sequence.append(preds["logits"])

    
    # Updating states
    input_dict["previous_state_c"] = preds["new_state_c"]
    input_dict["previous_state_h"] = preds["new_state_h"]
    
    
    # Decoding
    probs = np.concatenate(logits_sequence)
    transcription = postprocessing(probs)
    print(transcription[0][1], end="\r", flush=True)

!python DeepSpeech.py --n_steps -1 --export_dir /tmp --checkpoint_dir ./deepspeech-0.7.1-checkpoint --alphabet_config_path=alphabet.txt --scorer_path=kenlm.scorer >/dev/null 2>&1

mlmodel = ct.convert(tf_model, outputs=outputs)

inspect_inputs(mlmodel,tf_model)

input_dict = {}

input_dict["input_node"] = mfccs

input_dict["input_lengths"] = np.array([mfccs.shape[1]]).astype(np.float32)
input_dict["previous_state_c"] = np.zeros([1, 2048]).astype(np.float32) # Initializing cell state 
input_dict["previous_state_h"] = np.zeros([1, 2048]).astype(np.float32) # Initializing hidden state 

probs = mlmodel.predict(input_dict)["logits"]

transcription = postprocessing(probs)

print(transcription[0][1])

21:52 - Deep Speech Demo Recap: Convert with input type

import coremltools as ct

input = ct.TensorType(name="input_node", shape=(1, 16, 19, 26))
model = ct.convert(tf_model, outputs=outputs, inputs=[input])

26:26 - MIL Builder API sample

from coremltools.converters.mil import Builder as mb

@mb.program(input_specs=[mb.TensorSpec(shape=(1, 100, 100, 3))])
def prog(x):
    x = mb.relu(x=x)
    x = mb.transpose(x=x, perm=[0, 3, 1, 2])
    x = mb.reduce_mean(x=x, axes=[2, 3], keep_dims=False)
    x = mb.log(x=x)
    return x

28:20 - Converting with composite ops

import coremltools as ct

from transformers import TFT5Model

model = TFT5Model.from_pretrained('t5-small')

mlmodel = ct.convert(model)

# Einsum Notation

 $$ \Large "bnqd,bnkd \rightarrow bnqk" $$

$$ \large C(b, n, q, k) = \sum_d A(b, n, q, d) \times  B(b, n, k, d) $$

$$ \Large C = AB^{T}$$

from coremltools.converters.mil import Builder as mb

from coremltools.converters.mil import register_tf_op

@register_tf_op
def Einsum(context, node):

		assert node.attr['equation'] == 'bnqd,bnkd->bnqk'

    a = context[node.inputs[0]]
    b = context[node.inputs[1]]

    x = mb.matmul(x=a, y=b, transpose_x=False, transpose_y=True, name=node.name)

    context.add(node.name, x)

mlmodel = ct.convert(model)

print(mlmodel)

29:50 - Recap: Custom operation

@register_tf_op
def Einsum(context, node):
    assert node.attr['equation'] == 'bnqd,bnkd->bnqk'

    a = context[node.inputs[0]]
    b = context[node.inputs[1]]

    x = mb.matmul(x=a, y=b, transpose_x=False, transpose_y=True, name=node.name)
    
    context.add(node.name, x)

29:50 - Deep Speech demo utilities

import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow.python.ops import gen_audio_ops as contrib_audio


from deepspeech_training.util.text import Alphabet
from ds_ctcdecoder import ctc_beam_search_decoder, Scorer


## Preprocessing + Postprocessing functions are constructed using code in DeepSpeech repository: https://github.com/mozilla/DeepSpeech

audio_window_samples = 512
audio_step_samples = 320
n_input  = 26
audio_sample_rate = 16000
context = 9

lm_alpha = 0.931289039105002
lm_beta  = 1.1834137581510284
scorer_path = "./kenlm.scorer"

beam_width = 1024
cutoff_prob = 1.0
cutoff_top_n = 300

alphabet = Alphabet("./alphabet.txt")

scorer = Scorer(lm_alpha, lm_beta, scorer_path, alphabet)


def audiofile_to_features(wav_filename):
    
    samples = tf.io.read_file(wav_filename)
    
    decoded = contrib_audio.decode_wav(samples, desired_channels=1)
    
    spectrogram = contrib_audio.audio_spectrogram(decoded.audio,
                                                  window_size=audio_window_samples,
                                                  stride=audio_step_samples,
                                                  magnitude_squared=True)
        
    mfccs = contrib_audio.mfcc(spectrogram = spectrogram,
                               sample_rate = decoded.sample_rate,
                               dct_coefficient_count=n_input,
                               upper_frequency_limit=audio_sample_rate/2)
    
    mfccs = tf.reshape(mfccs, [-1, n_input])

    return mfccs, tf.shape(input=mfccs)[0]



def create_overlapping_windows(batch_x):
    
    batch_size = tf.shape(input=batch_x)[0]
    window_width = 2 * context + 1
    num_channels = n_input

    eye_filter = tf.constant(np.eye(window_width * num_channels)
                               .reshape(window_width, num_channels, window_width * num_channels), tf.float32) 
    
    # Create overlapping windows
    batch_x = tf.nn.conv1d(input=batch_x, filters=eye_filter, stride=1, padding='SAME')

    batch_x = tf.reshape(batch_x, [batch_size, -1, window_width, num_channels])

    return batch_x


sess = tf.Session(graph=tf.Graph())

with sess.graph.as_default() as g:
    path = tf.placeholder(tf.string)
    _features, _ = audiofile_to_features(path)
    _features = tf.expand_dims(_features, 0)
    _features = create_overlapping_windows(_features)

    

def preprocessing(input_file_path):
    return _features.eval(session=sess, feed_dict={path: input_file_path})




def postprocessing(logits):
    logits = np.squeeze(logits)

    decoded = ctc_beam_search_decoder(logits, alphabet, beam_width,
                                      scorer=scorer, cutoff_prob=cutoff_prob,
                                      cutoff_top_n=cutoff_top_n)

    return decoded



def inspect_tf_outputs(path):
    
    with open(path, 'rb') as f:
        serialized = f.read()
    gdef = tf.GraphDef()
    gdef.ParseFromString(serialized)

    with tf.Graph().as_default() as g:
        tf.import_graph_def(gdef, name="")

    output_nodes = []
    for op in g.get_operations():
    
        if op.type == "Const":
            continue
        
        if all([len(g.get_tensor_by_name(tensor.name).consumers()) == 0 for tensor in op.outputs]):
            
            output_nodes.append(op.name)

    return output_nodes


def inspect_inputs(mlmodel, tfmodel):
    
    names = []
    ranks = []
    shapes = []

    spec = mlmodel.get_spec()

    with open(tfmodel, 'rb') as f:
        serialized = f.read()
    gdef = tf.GraphDef()
    gdef.ParseFromString(serialized)

    with tf.Graph().as_default() as g:
        tf.import_graph_def(gdef, name="")

    for tensor in spec.description.input:
        name = tensor.name
        shape = tensor.type.multiArrayType.shape

        if tensor.type.multiArrayType.shapeRange:
            for dim, size in enumerate(tensor.type.multiArrayType.shapeRange.sizeRanges):
                if size.upperBound == -1:
                    shape[dim] = -1
                elif size.lowerBound < size.upperBound:
                    shape[dim] = -1
                elif size.lowerBound == size.upperBound:
                    assert shape[dim] == size.lowerBound
                else:
                    raise TypeError("Invalid shape range")

        coreml_shape = tuple(None if i == -1 else i for i in shape)

        tf_shape = tuple(g.get_tensor_by_name(name + ":0").shape.as_list())

        shapes.append({"Core ML shape": coreml_shape, "TF shape": tf_shape})
        names.append(name)
        ranks.append(len(coreml_shape))


    columns = [shapes[i] for i in np.argsort(ranks)[::-1]]
    indices = [names[i] for i in np.argsort(ranks)[::-1]]

    return pd.DataFrame(columns, index= indices)