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,
)
with open(cwd / 'figures' / 'controls' / 'instances' / 'sequence' / 'codons' / 'results' / 'weights_with_frame.pkl', 'wb') as f:
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,
shuffle=True).split(
y_strat, y_strat):
##due to the y_strat levels not being constant this idx_train/idx_valid split is not deterministic
idx_train, idx_valid = [
idx_train[idx] for idx in list(
StratifiedShuffleSplit(n_splits=1, test_size=300, random_state=0).
split(np.zeros_like(y_strat)[idx_train], y_strat[idx_train]))[0]
]
))
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, ))
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=200,
validation_data=ds_valid,
ds_train = iter(ds_train).get_next()
ds_valid_batch = iter(ds_valid).get_next()
losses = [Losses.CrossEntropy()]
while True:
position_encoder = InstanceModels.VariantPositionBin(24, 100)
mil = RaggedModels.MIL(instance_encoders=[],
sample_encoders=[position_encoder.model],
output_dim=y_label.shape[-1],
output_type='anlulogits',
mil_hidden=[32, 16],
mode='none')
mil.model.compile(loss=losses,
metrics=[Metrics.CrossEntropy(),
Metrics.Accuracy()],
optimizer=tf.keras.optimizers.Adam(
learning_rate=0.01, ))
mil.model.fit(x=ds_train[0],
y=ds_train[1],
batch_size=len(idx_train) // 4,
epochs=1000000,
validation_data=ds_valid_batch,
shuffle=True,
callbacks=callbacks)
eval = mil.model.evaluate(ds_valid)
if eval[1] < .0005:
break
