unit='tanh', init_W=init_W, init_U=init_U, init_b=init_b) h4 = FullyConnectedLayer(name='h4', parent=['h1', 'h2', 'h3'], nout=res, unit='sigmoid', init_W=init_W, init_b=init_b) cost = MSELayer(name='cost', parent=['h4', 'y']) nodes = [h1, h2, h3, h4, cost] rnn = Net(inputs=inputs, inputs_dim=inputs_dim, nodes=nodes) cost = unpack(rnn.build_recurrent_graph(output_args=[cost])) cost = cost.mean() cost.name = 'cost' model.graphs = [rnn] optimizer = Adam( lr=0.001 ) extension = [ GradientClipping(batch_size=batch_size), EpochCount(100), Monitoring(freq=100, ddout=[cost]), Picklize(freq=200, path=save_path)
# Global average pooling missing h4 = FullyConnectedLayer(name='h4', parent=['c2'], nout=10, unit='softmax', init_W=init_W, init_b=init_b) cost = MulCrossEntropyLayer(name='cost', parent=['y', 'h4']) # You will fill in a list of nodes and fed them to the model constructor nodes = [c1, c2, h1, h2, h3, h4, cost] # Your model will build the Theano computational graph cnn = Net(inputs=inputs, inputs_dim=inputs_dim, nodes=nodes) cnn.build_graph() # You can access any output of a node by doing model.nodes[$node_name].out cost = cnn.nodes['cost'].out err = error(predict(cnn.nodes['h4'].out), predict(y)) cost.name = 'cost' err.name = 'error_rate' model.graphs = [cnn] # Define your optimizer: Momentum (Nesterov), RMSProp, Adam optimizer = Adam( #lr=0.00005 lr=0.0005 )
unit='tanh', init_W=init_W, init_U=init_U, init_b=init_b) h4 = FullyConnectedLayer(name='h4', parent=['h1', 'h2', 'h3'], nout=res, unit='sigmoid', init_W=init_W, init_b=init_b) cost = MSELayer(name='cost', parent=['h4', 'y']) nodes = [h1, h2, h3, h4, cost] rnn = Net(inputs=inputs, inputs_dim=inputs_dim, nodes=nodes) cost = unpack(rnn.build_recurrent_graph(output_args=[cost])) cost = cost.mean() cost.name = 'cost' model.graphs = [rnn] optimizer = Adam(lr=0.001) extension = [ GradientClipping(batch_size=batch_size), EpochCount(100), Monitoring(freq=100, ddout=[cost]), Picklize(freq=200, path=save_path) ] mainloop = Training(name='toy_bb_gflstm',
# Global average pooling missing h4 = FullyConnectedLayer(name='h4', parent=['c2'], nout=10, unit='softmax', init_W=init_W, init_b=init_b) cost = MulCrossEntropyLayer(name='cost', parent=['y', 'h4']) # You will fill in a list of nodes and fed them to the model constructor nodes = [c1, c2, h1, h2, h3, h4, cost] # Your model will build the Theano computational graph cnn = Net(inputs=inputs, inputs_dim=inputs_dim, nodes=nodes) cnn.build_graph() # You can access any output of a node by doing model.nodes[$node_name].out cost = cnn.nodes['cost'].out err = error(predict(cnn.nodes['h4'].out), predict(y)) cost.name = 'cost' err.name = 'error_rate' model.graphs = [cnn] # Define your optimizer: Momentum (Nesterov), RMSProp, Adam optimizer = Adam( #lr=0.00005 lr=0.0005) extension = [
init_b=init_b) h2 = FullyConnectedLayer(name='h2', parent=['h1'], nout=10, unit='softmax', init_W=init_W, init_b=init_b) cost = MulCrossEntropyLayer(name='cost', parent=['onehot', 'h2']) # You will fill in a list of nodes and fed them to the model constructor nodes = [onehot, h1, h2, cost] # Your model will build the Theano computational graph mlp = Net(inputs=inputs, inputs_dim=inputs_dim, nodes=nodes) mlp.build_graph() # You can access any output of a node by doing model.nodes[$node_name].out cost = mlp.nodes['cost'].out err = error(predict(mlp.nodes['h2'].out), predict(mlp.nodes['onehot'].out)) cost.name = 'cost' err.name = 'error_rate' model.graphs = [mlp] # Define your optimizer: Momentum (Nesterov), RMSProp, Adam optimizer = RMSProp( lr=0.001 ) extension = [
batch_size=batch_size, nout=50, unit='tanh', init_W=init_W, init_U=init_U, init_b=init_b) h4 = FullyConnectedLayer(name='h4', parent=['h1', 'h2', 'h3'], nout=nlabel, unit='sigmoid', init_W=init_W, init_b=init_b) nodes = [h1, h2, h3, h4] rnn = Net(inputs=inputs, inputs_dim=inputs_dim, nodes=nodes) y_hat = rnn.build_recurrent_graph(output_args=[h4])[0] masked_y = y[mask.nonzero()] masked_y_hat = y_hat[mask.nonzero()] cost = NllBin(masked_y, masked_y_hat).sum() nll = NllBin(masked_y, masked_y_hat).mean() cost.name = 'cost' nll.name = 'nll' model.graphs = [rnn] optimizer = RMSProp( lr=0.0001, mom=0.95 ) extension = [