annotate utils.py @ 5:4f7e6612906b draft

"planemo upload for repository https://github.com/bgruening/galaxytools/tree/recommendation_training/tools/tool_recommendation_model commit 5eebc0cb44e71f581d548b7e842002705dd155eb"

import json
import random

import h5py
import numpy as np
import tensorflow as tf
from numpy.random import choice
from tensorflow.keras import backend


def read_file(file_path):
    """
    Read a file
    """
    with open(file_path, "r") as json_file:
        file_content = json.loads(json_file.read())
    return file_content


def format_tool_id(tool_link):
    """
    Extract tool id from tool link
    """
    tool_id_split = tool_link.split("/")
    tool_id = tool_id_split[-2] if len(tool_id_split) > 1 else tool_link
    return tool_id


def set_trained_model(dump_file, model_values):
    """
    Create an h5 file with the trained weights and associated dicts
    """
    hf_file = h5py.File(dump_file, "w")
    for key in model_values:
        value = model_values[key]
        if key == "model_weights":
            for idx, item in enumerate(value):
                w_key = "weight_" + str(idx)
                if w_key in hf_file:
                    hf_file.modify(w_key, item)
                else:
                    hf_file.create_dataset(w_key, data=item)
        else:
            if key in hf_file:
                hf_file.modify(key, json.dumps(value))
            else:
                hf_file.create_dataset(key, data=json.dumps(value))
    hf_file.close()


def weighted_loss(class_weights):
    """
    Create a weighted loss function. Penalise the misclassification
    of classes more with the higher usage
    """
    weight_values = list(class_weights.values())
    weight_values.extend(weight_values)

    def weighted_binary_crossentropy(y_true, y_pred):
        # add another dimension to compute dot product
        expanded_weights = tf.expand_dims(weight_values, axis=-1)
        bce = backend.binary_crossentropy(y_true, y_pred)
        return backend.dot(bce, expanded_weights)

    return weighted_binary_crossentropy


def balanced_sample_generator(
    train_data, train_labels, batch_size, l_tool_tr_samples, reverse_dictionary
):
    while True:
        dimension = train_data.shape[1]
        n_classes = train_labels.shape[1]
        tool_ids = list(l_tool_tr_samples.keys())
        random.shuffle(tool_ids)
        generator_batch_data = np.zeros([batch_size, dimension])
        generator_batch_labels = np.zeros([batch_size, n_classes])
        generated_tool_ids = choice(tool_ids, batch_size)
        for i in range(batch_size):
            random_toolid = generated_tool_ids[i]
            sample_indices = l_tool_tr_samples[str(random_toolid)]
            random_index = random.sample(range(0, len(sample_indices)), 1)[0]
            random_tr_index = sample_indices[random_index]
            generator_batch_data[i] = train_data[random_tr_index]
            generator_batch_labels[i] = train_labels[random_tr_index]
        yield generator_batch_data, generator_batch_labels


def compute_precision(
    model,
    x,
    y,
    reverse_data_dictionary,
    usage_scores,
    actual_classes_pos,
    topk,
    standard_conn,
    last_tool_id,
    lowest_tool_ids,
):
    """
    Compute absolute and compatible precision
    """
    pred_t_name = ""
    top_precision = 0.0
    mean_usage = 0.0
    usage_wt_score = list()
    pub_precision = 0.0
    lowest_pub_prec = 0.0
    lowest_norm_prec = 0.0
    pub_tools = list()
    actual_next_tool_names = list()
    test_sample = np.reshape(x, (1, len(x)))

    # predict next tools for a test path
    prediction = model.predict(test_sample, verbose=0)

    # divide the predicted vector into two halves - one for published and
    # another for normal workflows
    nw_dimension = prediction.shape[1]
    half_len = int(nw_dimension / 2)

    # predict tools
    prediction = np.reshape(prediction, (nw_dimension,))
    # get predictions of tools from published workflows
    standard_pred = prediction[:half_len]
    # get predictions of tools from normal workflows
    normal_pred = prediction[half_len:]

    standard_prediction_pos = np.argsort(standard_pred, axis=-1)
    standard_topk_prediction_pos = standard_prediction_pos[-topk]

    normal_prediction_pos = np.argsort(normal_pred, axis=-1)
    normal_topk_prediction_pos = normal_prediction_pos[-topk]

    # get true tools names
    for a_t_pos in actual_classes_pos:
        if a_t_pos > half_len:
            t_name = reverse_data_dictionary[int(a_t_pos - half_len)]
        else:
            t_name = reverse_data_dictionary[int(a_t_pos)]
        actual_next_tool_names.append(t_name)
    last_tool_name = reverse_data_dictionary[x[-1]]
    # compute scores for published recommendations
    if standard_topk_prediction_pos in reverse_data_dictionary:
        pred_t_name = reverse_data_dictionary[int(standard_topk_prediction_pos)]
        if last_tool_name in standard_conn:
            pub_tools = standard_conn[last_tool_name]
            if pred_t_name in pub_tools:
                pub_precision = 1.0
                # count precision only when there is actually true published tools
                if last_tool_id in lowest_tool_ids:
                    lowest_pub_prec = 1.0
                else:
                    lowest_pub_prec = np.nan
                if standard_topk_prediction_pos in usage_scores:
                    usage_wt_score.append(
                        np.log(usage_scores[standard_topk_prediction_pos] + 1.0)
                    )
        else:
            # count precision only when there is actually true published tools
            # else set to np.nan. Set to 0 only when there is wrong prediction
            pub_precision = np.nan
            lowest_pub_prec = np.nan
    # compute scores for normal recommendations
    if normal_topk_prediction_pos in reverse_data_dictionary:
        pred_t_name = reverse_data_dictionary[int(normal_topk_prediction_pos)]
        if pred_t_name in actual_next_tool_names:
            if normal_topk_prediction_pos in usage_scores:
                usage_wt_score.append(
                    np.log(usage_scores[normal_topk_prediction_pos] + 1.0)
                )
            top_precision = 1.0
            if last_tool_id in lowest_tool_ids:
                lowest_norm_prec = 1.0
            else:
                lowest_norm_prec = np.nan
    if len(usage_wt_score) > 0:
        mean_usage = np.mean(usage_wt_score)
    return mean_usage, top_precision, pub_precision, lowest_pub_prec, lowest_norm_prec


def get_lowest_tools(l_tool_freq, fraction=0.25):
    l_tool_freq = dict(sorted(l_tool_freq.items(), key=lambda kv: kv[1], reverse=True))
    tool_ids = list(l_tool_freq.keys())
    lowest_ids = tool_ids[-int(len(tool_ids) * fraction):]
    return lowest_ids


def verify_model(
    model,
    x,
    y,
    reverse_data_dictionary,
    usage_scores,
    standard_conn,
    lowest_tool_ids,
    topk_list=[1, 2, 3],
):
    """
    Verify the model on test data
    """
    print("Evaluating performance on test data...")
    print("Test data size: %d" % len(y))
    size = y.shape[0]
    precision = np.zeros([len(y), len(topk_list)])
    usage_weights = np.zeros([len(y), len(topk_list)])
    epo_pub_prec = np.zeros([len(y), len(topk_list)])
    epo_lowest_tools_pub_prec = list()
    epo_lowest_tools_norm_prec = list()
    lowest_counter = 0
    # loop over all the test samples and find prediction precision
    for i in range(size):
        lowest_pub_topk = list()
        lowest_norm_topk = list()
        actual_classes_pos = np.where(y[i] > 0)[0]
        test_sample = x[i, :]
        last_tool_id = str(int(test_sample[-1]))
        for index, abs_topk in enumerate(topk_list):
            (
                usg_wt_score,
                absolute_precision,
                pub_prec,
                lowest_p_prec,
                lowest_n_prec,
            ) = compute_precision(
                model,
                test_sample,
                y,
                reverse_data_dictionary,
                usage_scores,
                actual_classes_pos,
                abs_topk,
                standard_conn,
                last_tool_id,
                lowest_tool_ids,
            )
            precision[i][index] = absolute_precision
            usage_weights[i][index] = usg_wt_score
            epo_pub_prec[i][index] = pub_prec
            lowest_pub_topk.append(lowest_p_prec)
            lowest_norm_topk.append(lowest_n_prec)
        epo_lowest_tools_pub_prec.append(lowest_pub_topk)
        epo_lowest_tools_norm_prec.append(lowest_norm_topk)
        if last_tool_id in lowest_tool_ids:
            lowest_counter += 1
    mean_precision = np.mean(precision, axis=0)
    mean_usage = np.mean(usage_weights, axis=0)
    mean_pub_prec = np.nanmean(epo_pub_prec, axis=0)
    mean_lowest_pub_prec = np.nanmean(epo_lowest_tools_pub_prec, axis=0)
    mean_lowest_norm_prec = np.nanmean(epo_lowest_tools_norm_prec, axis=0)
    return (
        mean_usage,
        mean_precision,
        mean_pub_prec,
        mean_lowest_pub_prec,
        mean_lowest_norm_prec,
        lowest_counter,
    )


def save_model(
    results,
    data_dictionary,
    compatible_next_tools,
    trained_model_path,
    class_weights,
    standard_connections,
):
    # save files
    trained_model = results["model"]
    best_model_parameters = results["best_parameters"]
    model_config = trained_model.to_json()
    model_weights = trained_model.get_weights()
    model_values = {
        "data_dictionary": data_dictionary,
        "model_config": model_config,
        "best_parameters": best_model_parameters,
        "model_weights": model_weights,
        "compatible_tools": compatible_next_tools,
        "class_weights": class_weights,
        "standard_connections": standard_connections,
    }
    set_trained_model(trained_model_path, model_values)