Extract feature using pytorch

1

I need to write a code using PyTorch to extract features like this one

def extract_features(image):
    feature_extractor = models[config.base_model_name]["model"](
        include_top=False, weights="imagenet"
    )

    print(feature_extractor.summary())
    if config.pooling != "avg":
        feature_extractor = tf.keras.Model(
            feature_extractor.input,
            tf.keras.layers.AveragePooling2D(int(config.pooling[0]))(
                feature_extractor.output
            ),
        )

I tried this one


model.head = nn.Identity()
model.eval()
def extract_features(directory):
    results = []
    preprocess = T.Compose([
        T.Resize(256, interpolation=3),
        T.CenterCrop(224),
        T.ToTensor(),
        T.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
    ])
    for name in listdir(directory):
        filename=directory + '/' + name
        image= load_img(filename)
        vector=model(preprocess(image))

but the averaging pool I don’t use, how can I apply it with PyTorch, or does the Keras function extract features from the layer before classification only?

2

Could you show which exact model are you using and which layers do you want to remove? According to classic implementation of the Style Transferring (which i believe is very similar to your problem), features are extracted from various depths of the model. But you might require another approach.

Please do, print(model) and describe your problem better.

3

I’m using the ResNet50 model and I need to remove the last layer which is for classification as I need to extract the features for the images before classifying them. but i found this code for extract features of medical images

def extract_features(image):
    feature_extractor = models[config.base_model_name]["model"](
        include_top=False, weights="imagenet"
    )

    print(feature_extractor.summary())
    if config.pooling != "avg":
        feature_extractor = tf.keras.Model(
            feature_extractor.input,
            tf.keras.layers.AveragePooling2D(int(config.pooling[0]))(
                feature_extractor.output
            ),
        )

so I compared what I’m using for the code I saw and found that difference, does that make changes in the features I extract or not