ensemble

app.py

@@ -48,6 +48,23 @@ class Sequence:
         return to_categorical(np.array([padded_ie[0], all_ohe]))[:1]
 
 
+def get_lstm_model():
+    x_input = Input(shape=(100,))
+    max_length = 100
+    emb = Embedding(21, 128, input_length=max_length)(x_input)
+    bi_rnn = CuDNNLSTM(64, kernel_regularizer=l2(0.01), recurrent_regularizer=l2(0.01), bias_regularizer=l2(0.01))(emb)
+    x = Dropout(0.3)(bi_rnn)
+
+    # softmax classifier
+    x_output = Dense(1000, activation='softmax')(x)
+
+    model1 = Model(inputs=x_input, outputs=x_output)
+    model1.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
+    weights = cached_download(hf_hub_url("jonathang/Protein_Family_Models", 'model1.h5'))
+    model1.load_weights(weights)
+    return model1
+
+
 def residual_block(data, filters, d_rate):
     """
     _data: input
@@ -97,19 +114,22 @@ def get_model():
     return model2
 
 
-model = get_model()
+cnn_model = get_model()
+lstm_model = get_lstm_model()
 mappings_path = cached_download(hf_hub_url("jonathang/Protein_Family_Models", 'prot_mappings.json'))
 with open(mappings_path) as f:
     prot_mappings = json.load(f)
 
 def greet(Amino_Acid_Sequence):
     processed_seq = Sequence.prepare(Amino_Acid_Sequence)
-    raw_prediction = model.predict(processed_seq)[0]
-    idx = raw_prediction.argmax()
+    cnn_raw_prediction = cnn_model.predict(processed_seq)[0]
+    lstm_raw_prediction = lstm_model.predict(processed_seq)[0]
+    joined_prediction = (cnn_raw_prediction + lstm_raw_prediction) / 2.0
+    idx = joined_prediction.argmax()
     fam_asc = prot_mappings['id2fam_asc'][str(idx)]
     fam_id = prot_mappings['fam_asc2fam_id'][fam_asc]
     gc.collect()
-    return f"Input is {Amino_Acid_Sequence}.\nProcessed input is:\n{processed_seq}\n\nModel makes prediction which maps to:\nFamily Accession={fam_asc} and ID={fam_id}\n\nRaw Prediction:\n{raw_prediction}\n"
+    return f"Input is {Amino_Acid_Sequence}.\nProcessed input is:\n{processed_seq}\n\nModel makes prediction which maps to:\nFamily Accession={fam_asc} and ID={fam_id}\n\nRaw Joined Prediction:\n{joined_prediction}\n"
 
 iface = gr.Interface(fn=greet, inputs="text", outputs="text")
 iface.launch()