SHOW:
|
|
- or go back to the newest paste.
| 1 | def model3(kernel_number = 200, kernel_shape = (window_height,3)): | |
| 2 | #stride = 1 | |
| 3 | #dim = 40 | |
| 4 | #window_height = 8 | |
| 5 | #splits = ((40-8)+1)/1 = 33 | |
| 6 | #next(test_generator()) | |
| 7 | #next(train_generator(batch_size)) | |
| 8 | ||
| 9 | #kernel_number = 200 | |
| 10 | list_of_input = [Input(shape = (window_height,total_frames_with_deltas,3)) for i in range(splits)] | |
| 11 | list_of_conv_output = [] | |
| 12 | list_of_max_out = [] | |
| 13 | for i in range(splits): | |
| 14 | if splits == 1: | |
| 15 | list_of_conv_output.append(Conv2D(filters = kernel_number , kernel_size = kernel_shape, activation = 'relu')(list_of_input[i])) | |
| 16 | list_of_max_out.append((MaxPooling2D(pool_size=((1,11)))(list_of_conv_output[i]))) | |
| 17 | else: | |
| 18 | list_of_conv_output.append(Conv2D(filters = 200 , kernel_size = (window_height,3) , activation = 'relu')(list_of_input[i])) | |
| 19 | list_of_max_out.append((MaxPooling2D(pool_size=((1,11)))(list_of_conv_output[i]))) | |
| 20 | ||
| 21 | merge = keras.layers.concatenate(list_of_max_out) | |
| 22 | print merge.shape | |
| 23 | reshape = Reshape((total_frames/total_frames,-1))(merge) | |
| 24 | ||
| 25 | dense1 = Dense(units = 1000, activation = 'relu', name = "dense_1")(reshape) | |
| 26 | dense2 = Dense(units = 1000, activation = 'relu', name = "dense_2")(dense1) | |
| 27 | dense3 = Dense(units = 145 , activation = 'softmax', name = "dense_3")(dense2) | |
| 28 | ||
| 29 | ||
| 30 | model = Model(inputs = list_of_input , outputs = dense3) | |
| 31 | model.compile(loss="categorical_crossentropy", optimizer="SGD" , metrics = [metrics.categorical_accuracy]) | |
| 32 | ||
| 33 | reduce_lr=ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3, verbose=1, mode='auto', epsilon=0.001, cooldown=0) | |
| 34 | stop = EarlyStopping(monitor='val_loss', min_delta=0, patience=5, verbose=1, mode='auto') | |
| 35 | log=csv_logger = CSVLogger('/home/keerthikan/kaldi-trunk/dnn/training_'+str(total_frames)+"_"+str(dim)+"_"+str(window_height)+"_"+str(batch_size)+".csv")
| |
| 36 | checkpoint = ModelCheckpoint(filepath="/media/keerthikan/E2302E68302E443F/Timit-dataset/timit/fbank/nn/"+str(total_frames)+"_"+str(dim)+"_"+str(window_height)+"_"+str(batch_size)+".hdf5",save_best_only=True) | |
| 37 | ||
| 38 | if len(sys.argv) == 7: | |
| 39 | model.load_weigts(weights) | |
| 40 | ||
| 41 | print model.summary() | |
| 42 | ||
| 43 | #raw_input("okay?")
| |
| 44 | #hist_current = model.fit_generator(train_generator(batch_size), | |
| 45 | # steps_per_epoch=10, | |
| 46 | # epochs = 100000, | |
| 47 | # verbose = 1, | |
| 48 | # validation_data = test_generator(), | |
| 49 | # validation_steps=1, | |
| 50 | # pickle_safe = True, | |
| 51 | # workers = 4, | |
| 52 | # callbacks = [log,checkpoint]) | |
| 53 | return model | |
| 54 | ||
| 55 | ||
| 56 | #model3() | |
| 57 | ||
| 58 | model = KerasClassifier(build_fn=model3,epochs = 10,verbose=1) | |
| 59 | kernel_number = [10,50,100,150,200,250] | |
| 60 | kernel_shape = [(window_height,3),(window_height,5),(window_height,8)] | |
| 61 | param_grid = dict(kernel_number = kernel_number , kernel_shape=kernel_shape) | |
| 62 | grid = GridSearchCV(estimator=model, param_grid=param_grid) | |
| 63 | train_input,train_output = next(train_generator(1)) | |
| 64 | - | print train_input.shape |
| 64 | + | |
| 65 | - | print train_output.shape |
| 65 | + | |
| 66 | ||
| 67 | print("Best: %f using %s" % (grid_result.best_score_, grid_result.best_params_))
| |
| 68 | means = grid_result.cv_results_['mean_test_score'] | |
| 69 | stds = grid_result.cv_results_['std_test_score'] | |
| 70 | params = grid_result.cv_results_['params'] | |
| 71 | for mean, stdev, param in zip(means, stds, params): | |
| 72 | print("%f (%f) with: %r" % (mean, stdev, param)) |
