Running a frozen model on iOS

[Medicmind]

These are step to run a keras model you've created from this repository on your smartphone under IOS. This example uses resnet_family model with the cats_vs_dogs dataset:

First train your model with your dataset using on of the cv attention models:
python train_script.py -m resnet_family.RegNetZD -d cats_vs_dogs -b 8 --pretrained imagenet -i 256

This may create a model named:
checkpoints/resnet_family.RegNetZD32_LAMB_cats_vs_dogs_batchsize_8_randaug_6_mixup_0.1_cutmix_1.0_RRC_0.08_lr512_0.008_wd_0.02_epoch_6_val_acc_0.9985.h5

You can then quantize it into a tflite file as follows:

import tensorflow as tf
from keras_cv_attention_models import regnet, resnet_family, nfnets, cotnet, beit
from keras_cv_attention_models.imagenet import data
from PIL import Image
import glob
import numpy as np
from keras_cv_attention_models import model_surgery

cotnet.set_global_tpu_test(True)
mm = resnet_family.RegNetZD32(num_classes=2,pretrained=None)
mm.load_weights('checkpoints/resnet_family.RegNetZD32_LAMB_cats_vs_dogs_batchsize_8_randaug_6_mixup_0.1_cutmix_1.0_RRC_0.08_lr512_0.008_wd_0.02_epoch_6_val_acc_0.9985.h5') 
mm = model_surgery.convert_groups_conv2d_2_split_conv2d(mm) #, drop_rate=(0, 0.2))
print("ll",mm.input.shape, mm.input.dtype)

# Default one
converter = tf.lite.TFLiteConverter.from_keras_model(mm)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
#converter.experimental_new_converter=True

converter.target_spec.supported_types = [tf.int8]
converter.target_spec.supported_ops = [
    #tf.lite.OpsSet.TFLITE_BUILTINS,
    tf.lite.OpsSet.SELECT_TF_OPS,
	tf.lite.OpsSet.TFLITE_BUILTINS_INT8
	
]
#converter.target_spec.supported_ops = [tf.lite.OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8,
#tf.lite.OpsSet.TFLITE_BUILTINS]
def representative_dataset_gen():
    global mm
    files=glob.glob('/home/ubuntu/efs/diabetic_macular_oedematrain/0.00__negative/*')
    a = []

    for fl in files:
        img = np.array(Image.open(fl).resize((mm.input.shape[1],mm.input.shape[2])),  dtype="float32") #dtype=mm.input.dtype) 
        img = img / 255.0

        a.append(img)
    a = np.array(a) #,  dtype=mm.input.dtype)
    print('total shape',a.shape) # a is np array of 160 3D images
    img = tf.data.Dataset.from_tensor_slices(a).batch(1)
    for i in img.take(100): #len(a)): #BATCH_SIZE):
        print('shape',i.shape)
        yield [i]

converter.representative_dataset = representative_dataset_gen

converter.inference_input_type = tf.uint8
converter.inference_output_type = tf.uint8

open("resnet_familyRegNetZD32cats_vs_dogssuint.tflite", "wb").write(converter.convert())

Then in OSX download

https://github.com/tensorflow/examples

Then modify the pod file under examples-master/lite/examples/image_classification/ios/ImageClassification so that it includes TensorFlowLiteSelectTfOps like below

# Uncomment the next line to define a global platform for your project
 platform :ios, '12.0'

target 'ImageClassification' do
  # Comment the next line if you're not using Swift and don't want to use dynamic frameworks
  use_frameworks!

  # Pods for ImageClassification
   pod 'TensorFlowLiteSwift', '2.6.0'

   pod 'TensorFlowLiteSelectTfOps', '2.6.0'
end

Then install this pod:
pod install

Then with the frozen model you created (eg resnet_familyRegNetZD32cats_vs_dogssuint.tflite ) replace replace the mobilenet model with resnet_family model in this dir
examples-master/lite/examples/image_classification/ios/ImageClassification/Model

Also change labels.txt to the classes in your tflite file eg cats and dogs

Also change the ModelDataHandler.swift model file to be the name of the resnet_familyRegNetZD32cats_vs_dogssuint.tflite and change the inputWidth and inputHeight parameters to be the size of images you trained on eg 256x256 instead 224x224

The app should then run