diff --git a/src/eynollah/training/models.py b/src/eynollah/training/models.py index 0dc78d2..406e937 100644 --- a/src/eynollah/training/models.py +++ b/src/eynollah/training/models.py @@ -154,19 +154,13 @@ def conv_block(input_tensor, kernel_size, filters, stage, block, strides=(2, 2)) x = Activation('relu')(x) return x - -def resnet50_unet_light(n_classes, input_height=224, input_width=224, task="segmentation", weight_decay=1e-6, pretraining=False): - assert input_height % 32 == 0 - assert input_width % 32 == 0 - - img_input = Input(shape=(input_height, input_width, 3)) - +def resnet50(inputs, weight_decay=1e-6, pretraining=False): if IMAGE_ORDERING == 'channels_last': bn_axis = 3 else: bn_axis = 1 - x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(img_input) + x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(inputs) x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2), kernel_regularizer=l2(weight_decay), name='conv1')(x) f1 = x @@ -200,7 +194,17 @@ def resnet50_unet_light(n_classes, input_height=224, input_width=224, task="segm f5 = x if pretraining: - model = Model(img_input, x).load_weights(RESNET50_WEIGHTS_PATH) + model = Model(inputs, x).load_weights(RESNET50_WEIGHTS_PATH) + + return f1, f2, f3, f4, f5 + +def resnet50_unet_light(n_classes, input_height=224, input_width=224, task="segmentation", weight_decay=1e-6, pretraining=False): + assert input_height % 32 == 0 + assert input_width % 32 == 0 + + img_input = Input(shape=(input_height, input_width, 3)) + + f1, f2, f3, f4, f5 = resnet50(img_input, weight_decay, pretraining) v512_2048 = Conv2D(512, (1, 1), padding='same', data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))(f5) v512_2048 = (BatchNormalization(axis=bn_axis))(v512_2048) @@ -262,46 +266,7 @@ def resnet50_unet(n_classes, input_height=224, input_width=224, task="segmentati img_input = Input(shape=(input_height, input_width, 3)) - if IMAGE_ORDERING == 'channels_last': - bn_axis = 3 - else: - bn_axis = 1 - - x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(img_input) - x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2), kernel_regularizer=l2(weight_decay), - name='conv1')(x) - f1 = x - - x = BatchNormalization(axis=bn_axis, name='bn_conv1')(x) - x = Activation('relu')(x) - x = MaxPooling2D((3, 3), data_format=IMAGE_ORDERING, strides=(2, 2))(x) - - x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1)) - x = identity_block(x, 3, [64, 64, 256], stage=2, block='b') - x = identity_block(x, 3, [64, 64, 256], stage=2, block='c') - f2 = one_side_pad(x) - - x = conv_block(x, 3, [128, 128, 512], stage=3, block='a') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='b') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='c') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='d') - f3 = x - - x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f') - f4 = x - - x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c') - f5 = x - - if pretraining: - Model(img_input, x).load_weights(RESNET50_WEIGHTS_PATH) + f1, f2, f3, f4, f5 = resnet50(img_input, weight_decay, pretraining) v1024_2048 = Conv2D(1024, (1, 1), padding='same', data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay))( f5) @@ -372,47 +337,7 @@ def vit_resnet50_unet(num_patches, transformer_mlp_head_units = [128, 64] inputs = Input(shape=(input_height, input_width, 3)) - if IMAGE_ORDERING == 'channels_last': - bn_axis = 3 - else: - bn_axis = 1 - - x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(inputs) - x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2),kernel_regularizer=l2(weight_decay), name='conv1')(x) - f1 = x - - x = BatchNormalization(axis=bn_axis, name='bn_conv1')(x) - x = Activation('relu')(x) - x = MaxPooling2D((3, 3), data_format=IMAGE_ORDERING, strides=(2, 2))(x) - - x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1)) - x = identity_block(x, 3, [64, 64, 256], stage=2, block='b') - x = identity_block(x, 3, [64, 64, 256], stage=2, block='c') - f2 = one_side_pad(x) - - x = conv_block(x, 3, [128, 128, 512], stage=3, block='a') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='b') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='c') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='d') - f3 = x - - x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f') - f4 = x - - x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c') - f5 = x - - if pretraining: - model = Model(inputs, x).load_weights(RESNET50_WEIGHTS_PATH) - - #num_patches = x.shape[1]*x.shape[2] + f1, f2, f3, f4, f5 = resnet50(inputs, weight_decay, pretraining) patches = Patches(transformer_patchsize_x, transformer_patchsize_y)(x) # Encode patches. @@ -540,42 +465,9 @@ def vit_resnet50_unet_transformer_before_cnn(num_patches, encoded_patches = Conv2D(3, (1, 1), padding='same', data_format=IMAGE_ORDERING, kernel_regularizer=l2(weight_decay), name='convinput')(encoded_patches) - x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(encoded_patches) - x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2),kernel_regularizer=l2(weight_decay), name='conv1')(x) - f1 = x + f1, f2, f3, f4, f5 = resnet50(encoded_patches, weight_decay, pretraining) - x = BatchNormalization(axis=bn_axis, name='bn_conv1')(x) - x = Activation('relu')(x) - x = MaxPooling2D((3, 3), data_format=IMAGE_ORDERING, strides=(2, 2))(x) - - x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1)) - x = identity_block(x, 3, [64, 64, 256], stage=2, block='b') - x = identity_block(x, 3, [64, 64, 256], stage=2, block='c') - f2 = one_side_pad(x) - - x = conv_block(x, 3, [128, 128, 512], stage=3, block='a') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='b') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='c') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='d') - f3 = x - - x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f') - f4 = x - - x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c') - f5 = x - - if pretraining: - model = Model(encoded_patches, x).load_weights(RESNET50_WEIGHTS_PATH) - - v1024_2048 = Conv2D( 1024 , (1, 1), padding='same', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay))(x) + v1024_2048 = Conv2D( 1024 , (1, 1), padding='same', data_format=IMAGE_ORDERING,kernel_regularizer=l2(weight_decay))(f5) v1024_2048 = (BatchNormalization(axis=bn_axis))(v1024_2048) v1024_2048 = Activation('relu')(v1024_2048) @@ -633,47 +525,7 @@ def resnet50_classifier(n_classes,input_height=224,input_width=224,weight_decay= img_input = Input(shape=(input_height,input_width , 3 )) - if IMAGE_ORDERING == 'channels_last': - bn_axis = 3 - else: - bn_axis = 1 - - x = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(img_input) - x = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2),kernel_regularizer=l2(weight_decay), name='conv1')(x) - f1 = x - - x = BatchNormalization(axis=bn_axis, name='bn_conv1')(x) - x = Activation('relu')(x) - x = MaxPooling2D((3, 3) , data_format=IMAGE_ORDERING , strides=(2, 2))(x) - - - x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1)) - x = identity_block(x, 3, [64, 64, 256], stage=2, block='b') - x = identity_block(x, 3, [64, 64, 256], stage=2, block='c') - f2 = one_side_pad(x ) - - - x = conv_block(x, 3, [128, 128, 512], stage=3, block='a') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='b') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='c') - x = identity_block(x, 3, [128, 128, 512], stage=3, block='d') - f3 = x - - x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e') - x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f') - f4 = x - - x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b') - x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c') - f5 = x - - if pretraining: - Model(img_input, x).load_weights(RESNET50_WEIGHTS_PATH) + _, _, _, _, x = resnet50(img_input, weight_decay, pretraining) x = AveragePooling2D((7, 7), name='avg_pool')(x) x = Flatten()(x) @@ -693,43 +545,10 @@ def machine_based_reading_order_model(n_classes,input_height=224,input_width=224 img_input = Input(shape=(input_height,input_width , 3 )) - if IMAGE_ORDERING == 'channels_last': - bn_axis = 3 - else: - bn_axis = 1 - - x1 = ZeroPadding2D((3, 3), data_format=IMAGE_ORDERING)(img_input) - x1 = Conv2D(64, (7, 7), data_format=IMAGE_ORDERING, strides=(2, 2),kernel_regularizer=l2(weight_decay), name='conv1')(x1) - - x1 = BatchNormalization(axis=bn_axis, name='bn_conv1')(x1) - x1 = Activation('relu')(x1) - x1 = MaxPooling2D((3, 3) , data_format=IMAGE_ORDERING , strides=(2, 2))(x1) + _, _, _, _, x = resnet50(img_input, weight_decay, pretraining) - x1 = conv_block(x1, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1)) - x1 = identity_block(x1, 3, [64, 64, 256], stage=2, block='b') - x1 = identity_block(x1, 3, [64, 64, 256], stage=2, block='c') - - x1 = conv_block(x1, 3, [128, 128, 512], stage=3, block='a') - x1 = identity_block(x1, 3, [128, 128, 512], stage=3, block='b') - x1 = identity_block(x1, 3, [128, 128, 512], stage=3, block='c') - x1 = identity_block(x1, 3, [128, 128, 512], stage=3, block='d') - - x1 = conv_block(x1, 3, [256, 256, 1024], stage=4, block='a') - x1 = identity_block(x1, 3, [256, 256, 1024], stage=4, block='b') - x1 = identity_block(x1, 3, [256, 256, 1024], stage=4, block='c') - x1 = identity_block(x1, 3, [256, 256, 1024], stage=4, block='d') - x1 = identity_block(x1, 3, [256, 256, 1024], stage=4, block='e') - x1 = identity_block(x1, 3, [256, 256, 1024], stage=4, block='f') - - x1 = conv_block(x1, 3, [512, 512, 2048], stage=5, block='a') - x1 = identity_block(x1, 3, [512, 512, 2048], stage=5, block='b') - x1 = identity_block(x1, 3, [512, 512, 2048], stage=5, block='c') - - if pretraining: - Model(img_input , x1).load_weights(RESNET50_WEIGHTS_PATH) - - x1 = AveragePooling2D((7, 7), name='avg_pool1')(x1) - flattened = Flatten()(x1) + x = AveragePooling2D((7, 7), name='avg_pool1')(x) + flattened = Flatten()(x) o = Dense(256, activation='relu', name='fc512')(flattened) o=Dropout(0.2)(o)