[pos_loader, bin_loader, chr_loader], [five_p_loader, three_p_loader, ref_loader, alt_loader, strand_loader], ] # set y label y_label = np.log( sample_df['non_syn_counts'].values / (panels.loc[panels['Panel'] == 'Agilent_kit']['cds'].values[0] / 1e6) + 1)[:, np.newaxis] y_strat = np.argmax(samples['histology'], axis=-1) losses = [Losses.QuantileLoss()] metrics = [Metrics.QuantileLoss()] encoders = [ InstanceModels.PassThrough(shape=(1, )), InstanceModels.VariantPositionBin(24, 100), InstanceModels.VariantSequence(6, 4, 2, [16, 16, 8, 8], fusion_dimension=32) ] all_weights = [ pickle.load( open( cwd / 'figures' / 'tmb' / 'tcga' / 'VICC_01_R2' / 'results' / 'run_naive.pkl', 'rb')), pickle.load( open( cwd / 'figures' / 'tmb' / 'tcga' / 'VICC_01_R2' / 'results' /
runs = 3 initial_weights = [] losses = [Losses.QuantileLoss()] metrics = [Metrics.QuantileLoss()] callbacks = [ tf.keras.callbacks.EarlyStopping(monitor='val_QL', min_delta=0.0001, patience=40, mode='min', restore_best_weights=True) ] ##for sequence for i in range(runs): pass_encoder = InstanceModels.PassThrough(shape=(1, )) # position_encoder = InstanceModels.VariantPositionBin(24, 100) # sequence_encoder = InstanceModels.VariantSequence(6, 4, 2, [16, 16, 8, 8], fusion_dimension=32) mil = RaggedModels.MIL(instance_encoders=[pass_encoder.model], output_dim=1, pooling='sum', mil_hidden=(64, 32, 16), output_type='quantiles', regularization=0) # mil = RaggedModels.MIL(instance_encoders=[position_encoder.model], output_dim=1, pooling='sum', mil_hidden=(64, 32, 16), output_type='quantiles', regularization=0) # mil = RaggedModels.MIL(instance_encoders=[sequence_encoder.model], output_dim=1, pooling='sum', mil_hidden=(64, 32, 16), output_type='quantiles', regularization=0) mil.model.compile(loss=losses, metrics=metrics, optimizer=tf.keras.optimizers.Adam(learning_rate=0.001)) initial_weights.append(mil.model.get_weights())
ds_test = tf.data.Dataset.from_tensor_slices((idx_test, y_label[idx_test])) ds_test = ds_test.batch(len(idx_test), drop_remainder=False) ds_test = ds_test.map(lambda x, y: ((tf.gather(tf.constant(D['seq_5p'], dtype=tf.int32), x), tf.gather(tf.constant(D['seq_3p'], dtype=tf.int32), x), tf.gather(tf.constant(D['seq_ref'], dtype=tf.int32), x), tf.gather(tf.constant(D['seq_alt'], dtype=tf.int32), x), tf.gather(tf.constant(D['strand_emb'], dtype=tf.float32), x), tf.gather(tf.constant(D['cds_emb'], dtype=tf.float32), x) ), y, )) sequence_encoder = InstanceModels.VariantSequence(6, 4, 2, [64, 64, 64, 64], fusion_dimension=128, use_frame=True) mil = RaggedModels.MIL(instance_encoders=[], sample_encoders=[sequence_encoder.model], output_dim=y_label.shape[-1], output_type='other', mil_hidden=[128, 128, 64, 32], mode='none') losses = [Losses.CrossEntropy()] mil.model.compile(loss=losses, metrics=[Metrics.Accuracy(), Metrics.CrossEntropy()], weighted_metrics=[Metrics.Accuracy(), Metrics.CrossEntropy()], optimizer=tf.keras.optimizers.Adam(learning_rate=0.001, )) callbacks = [tf.keras.callbacks.EarlyStopping(monitor='val_weighted_CE', min_delta=0.001, patience=10, mode='min', restore_best_weights=True)] mil.model.fit(ds_train, steps_per_epoch=50, validation_data=ds_valid, epochs=10000, callbacks=callbacks,
ds_test = tf.data.Dataset.from_tensor_slices((idx_test, y_label[idx_test])) ds_test = ds_test.batch(len(idx_test), drop_remainder=False) ds_test = ds_test.map(lambda x, y: ( ( tf.gather(tf.constant(D['seq_5p'], dtype=tf.int32), x), tf.gather(tf.constant(D['seq_3p'], dtype=tf.int32), x), tf.gather(tf.constant(D['seq_ref'], dtype=tf.int32), x), tf.gather(tf.constant(D['seq_alt'], dtype=tf.int32), x), tf.gather(tf.constant(D['strand_emb'], dtype=tf.float32), x), ), y, )) sequence_encoder = InstanceModels.VariantSequence(6, 4, 2, [16, 16, 8, 8], fusion_dimension=128) mil = RaggedModels.MIL(instance_encoders=[], sample_encoders=[sequence_encoder.model], output_dim=y_label.shape[-1], output_type='other', mil_hidden=[128, 128], mode='none') losses = [Losses.CrossEntropy()] mil.model.compile( loss=losses, metrics=[Metrics.Accuracy(), Metrics.CrossEntropy()], weighted_metrics=[Metrics.Accuracy(), Metrics.CrossEntropy()], optimizer=tf.keras.optimizers.Adam(learning_rate=0.001, ))
alt_loader(x, ragged_output=True), strand_loader(x, ragged_output=True), tf.gather(tf.constant(types), x)), y)) ds_test = tf.data.Dataset.from_tensor_slices((idx_test, y_label[idx_test])) ds_test = ds_test.batch(len(idx_test), drop_remainder=False) ds_test = ds_test.map(lambda x, y: ( (five_p_loader(x, ragged_output=True), three_p_loader( x, ragged_output=True), ref_loader(x, ragged_output=True), alt_loader(x, ragged_output=True), strand_loader(x, ragged_output=True), tf.gather(tf.constant(types), x)), y)) histories = [] evaluations = [] weights = [] for i in range(3): sequence_encoder = InstanceModels.VariantSequence(6, 4, 2, [16, 16, 8, 8]) sample_encoder = SampleModels.Type(shape=(), dim=len(np.unique(types))) # mil = RaggedModels.MIL(instance_encoders=[sequence_encoder.model], sample_encoders=[sample_encoder.model], sample_layers=[64, ], output_dim=1, pooling='both', output_type='other', pooled_layers=[32, ]) mil = RaggedModels.MIL(instance_encoders=[sequence_encoder.model], sample_encoders=[sample_encoder.model], fusion='before', output_dim=1, pooling='both', output_type='other', pooled_layers=[ 32, ]) losses = ['mse'] mil.model.compile(loss=losses, metrics=['mse'], optimizer=tf.keras.optimizers.Adam(
y, tf.gather(tf.constant(y_weights, dtype=tf.float32), x) )) ds_valid = tf.data.Dataset.from_tensor_slices((idx_valid, y_label[idx_valid])) ds_valid = ds_valid.batch(len(idx_valid), drop_remainder=False) ds_valid = ds_valid.map(lambda x, y: ((pos_loader(x, ragged_output=True), bin_loader(x, ragged_output=True), chr_loader(x, ragged_output=True), ), y, tf.gather(tf.constant(y_weights, dtype=tf.float32), x) )) while True: position_encoder = InstanceModels.VariantPositionBin(24, 100) mil = RaggedModels.MIL(instance_encoders=[position_encoder.model], output_dim=2, pooling='sum', mil_hidden=(64, 32, 16, 8), output_type='anlulogits') mil.model.compile(loss=losses, metrics=[Metrics.CrossEntropy(), Metrics.Accuracy()], weighted_metrics=[Metrics.CrossEntropy(), Metrics.Accuracy()], optimizer=tf.keras.optimizers.Adam(learning_rate=0.005, clipvalue=10000)) mil.model.fit(ds_train, steps_per_epoch=20, validation_data=ds_valid, epochs=10000, callbacks=callbacks) eval = mil.model.evaluate(ds_valid)
for group, event in zip(cancer_labels, y_label[:, 1]) ]) class_counts = dict(zip(*np.unique(y_strat, return_counts=True))) y_weights = np.array([1 / class_counts[_] for _ in y_strat]) y_weights /= np.sum(y_weights) weights = [] callbacks = [ tf.keras.callbacks.EarlyStopping(monitor='val_CE', min_delta=0.0001, patience=50, mode='min', restore_best_weights=True) ] losses = [Losses.CrossEntropy(from_logits=False)] sequence_encoder = InstanceModels.VariantSequence(20, 4, 2, [8, 8, 8, 8]) mil = RaggedModels.MIL(instance_encoders=[sequence_encoder.model], output_dim=2, pooling='sum', mil_hidden=(64, 64, 32, 16), output_type='classification_probability') mil.model.compile( loss=losses, metrics=[Metrics.CrossEntropy(from_logits=False), Metrics.Accuracy()], optimizer=tf.keras.optimizers.Adam(learning_rate=0.001, clipvalue=10000)) initial_weights = mil.model.get_weights() ##stratified K fold for test for idx_train, idx_test in StratifiedKFold(n_splits=9, random_state=0,
loaders = [ [ones_loader], ] # set y label y_label = np.log( sample_df['non_syn_counts'].values / (panels.loc[panels['Panel'] == 'Agilent_kit']['cds'].values[0] / 1e6) + 1)[:, np.newaxis] y_strat = np.argmax(samples['histology'], axis=-1) losses = [Losses.QuantileLoss()] metrics = [Metrics.QuantileLoss()] encoders = [ InstanceModels.PassThrough(shape=(1, )), ] all_weights = [ pickle.load( open( cwd / 'figures' / 'tmb' / 'tcga' / 'nonsyn_table' / 'DFCI_ONCO' / 'results' / 'run_naive.pkl', 'rb')) ] results = {} for encoder, loaders, weights, name in zip(encoders, loaders, all_weights, ['naive']): mil = RaggedModels.MIL(instance_encoders=[encoder.model],