EvalScale2.py

import numpy as np
import os
import re
from keras.models import Model, load_model
from keras.preprocessing.image import load_img, img_to_array
from keras import applications
from keras.layers import Dropout, Conv2D, Reshape, Dense, Flatten, AvgPool2D, MaxPool2D
from keras.layers.normalization import BatchNormalization
from keras.applications.resnet50 import preprocess_input

import matplotlib.pyplot as plt

img_width, img_height   = 224*9, 224*9

# generates a crop from one image
def eval_generator(img_dir):

    file_names = [f for f in os.listdir(img_dir) if re.match(r'[0-9]+.*\.jpg', f)]
    file_names = sorted(file_names, key=lambda
        item: (int(item.partition('.')[0]) if item[0].isdigit() else float('inf'), item))

    for filename in file_names:
        img = load_img(os.path.join(img_dir, filename))
        width, height = img.size
        x, y = np.random.randint(width-img_width), np.random.randint(height-img_height)
        # special type image to np array
        dd = img_to_array(img)
        ss = dd[y:y+img_height, x:x+img_width]
        yy = np.expand_dims(ss, axis=0)
        yield preprocess_input(yy)


    yield np.array([[[]]])


def create_model(weights_path=None, inp_shape=None):
    # img_input = Input(shape=inp_shape)

    model = applications.ResNet50(weights='imagenet', include_top=False, input_shape=inp_shape)

    pretr_layer_outputs = {}
    for layer in model.layers:
        pretr_layer_outputs[layer.name] = layer.get_output_at(0)

    if model.name == 'resnet50':
        x = pretr_layer_outputs['activation_49']       # we need to skip the very last layer in case of ResNet
    else:
        x = model.output

    # model_pretrained.summary()
    # x = Conv2D(512, (1, 1), activation='elu', padding='valid', name='Kir_0')(x)
    # x = BatchNormalization()(x)
    # x = Conv2D(512, (7, 7), activation='elu', padding='valid', name='Kir_1')(x)
    # x = Dropout(0.25)(x)
    # x = Conv2D(256, (1, 1), activation='elu', padding='valid', name='Kir_2')(x)
    # x = BatchNormalization()(x)
    # x = Conv2D(1, (1, 1), activation='relu', padding='valid', name='Kir_3')(x)
    # # predictions = Reshape(target_shape=(1,))(x)

    x = MaxPool2D(pool_size=(7, 7), padding='same', name='Kir_Pool0')(x)
    x = BatchNormalization()(x)
    x = Conv2D(256, (1, 1), activation='elu', padding='valid', name='Kir_0')(x)
    x = MaxPool2D(pool_size=(3, 3), padding='same', name='Kir_Pool2')(x)
    # x = BatchNormalization()(x)
    # x = Conv2D(256, (1, 1), activation='elu', padding='valid', name='Kir_2')(x)
    # x = BatchNormalization()(x)
    x = Conv2D(1, (1, 1), activation='relu', padding='valid', name='Kir_3')(x)
    # predictions = Reshape(target_shape=(1,))(x)

    model = Model(model.input, x, name='kir_b')

    # model.summary()

    # load weights
    model.load_weights(weights_path)

    return model



model = create_model(weights_path="./cp/43-e054-vl0.62.hdf5", inp_shape=(img_width, img_height, 3) )

gen = eval_generator(img_dir="../Sealion/Train")
np.set_printoptions(formatter={'float': '{: 0.4f}'.format})

# while True:
for i in range(30):
    img_batch = next(gen)

    if img_batch.size == 0:
        break

    prediction = model.predict(img_batch)
    aaa = prediction.squeeze()
    print("%d: mean: %s  med: %s min: %s max: %s" % (i, np.mean(aaa, (0,1)),np.median(aaa, (0,1)),np.min(aaa, (0,1)),np.max(aaa, (0,1))))
    img = img_batch.squeeze()
    img = (img - np.min(img))/300.
    # plt.imshow(img); plt.show()