PyPI 官网下载 | vqsr_cnn-0.0.174.tar.gz

import os import json # Keras Imports from keras import layers from keras import metrics import keras.backend as K from keras.optimizers import Adam from keras.models import Model, load_model from keras.layers.convolutional import Conv1D, Conv2D, MaxPooling1D, MaxPooling2D from keras.callbacks import ModelCheckpoint, EarlyStopping, TensorBoard, ReduceLROnPlateau from keras.layers import Input, Dense, Dropout, BatchNormalization, SpatialDropout2D, Activation, Flatten from . import plots from . import defines from . import arguments from . import tensor_maps def start_session_get_args_and_model(intra_ops, inter_ops, semantics_json, weights_hd5=None, tensor_type=None): K.clear_session() K.get_session().close() cfg = K.tf.ConfigProto(intra_op_parallelism_threads=intra_ops, inter_op_parallelism_threads=inter_ops) cfg.gpu_options.allow_growth = True K.set_session(K.tf.Session(config=cfg)) return args_and_model_from_semantics(semantics_json, weights_hd5, tensor_type) def args_and_model_from_semantics(semantics_json, weights_hd5=None, tensor_type=None): args = arguments.parse_args() if semantics_json is not None and os.path.exists(semantics_json): print('try to load from semantics json:', semantics_json) model = set_args_and_get_model_from_semantics(args, semantics_json, weights_hd5) else: print('try to load from weights HD5:', weights_hd5) model = load_model(weights_hd5, custom_objects=get_metric_dict(args.labels)) args.tensor_name = tensor_type return args, model def set_args_and_get_model_from_semantics(args, semantics_json, weights_hd5=None): '''Recreate a model from a json file specifying model semantics. Update the args namespace from the semantics file values. Assert that the serialized tensor map and the recreated one are the same. Arguments: args.tensor_name: String which indicates tensor map to use or None args.window_size: sites included in the tensor map args.read_limit: Maximum reads included in the tensor map args.annotations: List of annotations or None semantics_json: Semantics json file (created with serialize_model_semantics()) Returns: The Keras model ''' with open(semantics_json, 'r') as infile: semantics = json.load(infile) if 'input_tensor_map' in semantics: args.tensor_name = semantics['input_tensor_map_name'] args.window_size = semantics['window_size'] args.read_limit = semantics['read_limit'] tm = tensor_maps.get_tensor_channel_map_from_args(args) assert(len(tm) == len(semantics['input_tensor_map'])) for key in tm: assert(tm[key] == semantics['input_tensor_map'][key]) if 'input_annotations' in semantics: args.annotations = semantics['input_annotations'] args.annotation_set = semantics['input_annotation_set'] args.input_symbols = semantics['input_symbols'] args.labels = semantics['output_labels'] if 'channels_last' in semantics: args.channels_last = semantics['channels_last'] if args.channels_last: K.set_image_data_format('channels_last') else: K.set_image_data_format('channels_first') if weights_hd5 is None: weights_hd5 = os.path.join(os.path.dirname(semantics_json), semantics['architecture']) print('Updated arguments:', args, '\nWeight file from:', weights_hd5) model = load_model(weights_hd5, custom_objects=get_metric_dict(args.labels)) model.summary() return model # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ # ~~~~~~~ Models ~~~~~~~~~~~~~~~~~~ # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ def build_reference_annotation_model(args): '''Build Reference 1d CNN model for classifying variants with skip connected annotations. Convolutions followed by dense connection, concatenated with annotations. Dynamically sets input channels based on args via tensor_maps.get_tensor_channel_map_from_args(args) Uses the functional API. Prints out model summary. Arguments args.tensor_name: The name of the tensor mapping which data goes to which channels args.annotation_set: The variant annotation set, perhaps from a HaplotypeCaller VCF. args.labels: The output labels (e.g. SNP, NOT_SNP, INDEL, NOT_INDEL) Returns The keras model ''' if K.image_data_format() == 'channels_last': channel_axis = -1 else: channel_axis = 1 channel_map = tensor_maps.get_tensor_channel_map_from_args(args) reference = Input(shape=(args.window_size, len(channel_map)), name=args.tensor_name) conv_width = 12 conv_dropout = 0.1 fc_dropout = 0.2 x = Conv1D(filters=256, kernel_size=conv_width, activation="relu", kernel_initializer='he_normal')(reference) x = Conv1D(filters=256, kernel_size=conv_width, activation="relu", kernel_initializer='he_normal')(x) x = Dropout(conv_dropout)(x) x = Conv1D(filters=128, kernel_size=conv_width, activation="relu", kernel_initializer='he_normal')(x) x = Dropout(conv_dropout)(x) x = Flatten()(x) annotations = Input(shape=(len(args.annotations),), name=args.annotation_set) annos_normed = BatchNormalization(axis=channel_axis)(annotations) annos_normed_x = Dense(units=40, kernel_initializer='normal', activation='relu')(annos_normed) x = layers.concatenate([x, annos_normed_x], axis=channel_axis) x = Dense(units=40, kernel_initializer='normal', activation='relu')(x) x = Dropout(fc_dropout)(x) x = layers.concatenate([x, annos_normed], axis=channel_axis) prob_output = Dense(units=len(args.labels), kernel_initializer='glorot_normal', activation='softmax')(x) model = Model(inputs=[reference, annotations], outputs=[prob_output]) adamo = Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, clipnorm=1.) model.compile(optimizer=adamo, loss='categorical_crossentropy', metrics=get_metrics(args.labels)) model.summary() if os.path.exists(args.weights_hd5): model.load_weights(args.weights_hd5, by_name=True) print('Loaded model weights from:', args.weights_hd5) return model def build_read_tensor_2d_and_annotations_model(args): '''Build Read Tensor 2d CNN model with variant annotations mixed in for classifying variants. 2d Convolutions followed by dense connection mixed with annotation values. Dynamically sets input channels based on args via defines.total_input_channels_from_args(args) Uses the functional API. Supports theano or tensorflow channel ordering via K.image_data_format(). Prints out model summary. Arguments args.window_size: Length in base-pairs of sequence centered at the variant to use as input. args.labels: The output labels (e.g. SNP, NOT_SNP, INDEL, NOT_INDEL) Returns The keras model ''' in_channels = tensor_maps.total_input_channels_from_args(args) if K.image_data_format() == 'channels_last': in_shape = (args.read_limit, args.window_size, in_channels) concat_axis = -1 else: in_shape = (in_channels, args.read_limit, args.window_size) concat_axis = 1 read_tensor = Input(shape=in_shape, name=args.tensor_name) read_conv_width = 16 conv_dropout = 0.2 fc_dropout = 0.3 x = Conv2D(216, (read_conv_width, 1), padding='valid', activation="relu")(read_tensor) x = Conv2D(160, (1, read_conv_width), padding='valid', activation="relu")(x) x = Conv2D(128, (read_conv_width, 1), padding='valid', activation="relu")(x) x = MaxPooling2D((2,1))(x) x = Conv2D(96, (1, read_conv_width), padding='valid', activation="relu")(x) x = MaxPooling2D((2,1))(x) x = Dropout(conv_dropout)(x) x = Conv2D(64, (read_conv_width, 1), padding='valid', activation="relu")(x) x = MaxPooling2D((2,1))(x) x = Dropout(conv_dropout)(x) x = Flatten()(x) # Mix the variant annotations in