chem_tensorflow.py

#!/usr/bin/env/python

import json
import os
import pickle
import random
import time
from typing import List, Any, Sequence

import numpy as np
import tensorflow as tf

from utils import MLP, ThreadedIterator, SMALL_NUMBER


class ChemModel(object):
    @classmethod
    def default_params(cls):
        return {
            'num_epochs': 3000,
            'patience': 25,
            'learning_rate': 0.001,
            'clamp_gradient_norm': 1.0,
            'out_layer_dropout_keep_prob': 1.0,

            'hidden_size': 100,
            'num_timesteps': 4,
            'use_graph': True,

            'tie_fwd_bkwd': True,
            'task_ids': [0],

            'random_seed': 0,

            'train_file': 'molecules_train.json',
            'valid_file': 'molecules_valid.json'
        }

    def __init__(self, args):
        self.args = args

        # Collect argument things:
        data_dir = ''
        if '--data_dir' in args and args['--data_dir'] is not None:
            data_dir = args['--data_dir']
        self.data_dir = data_dir

        self.run_id = "_".join([time.strftime("%Y-%m-%d-%H-%M-%S"), str(os.getpid())])
        log_dir = args.get('--log_dir') or '.'
        os.makedirs(log_dir, exist_ok=True)
        self.log_file = os.path.join(log_dir, "%s_log.json" % self.run_id)
        self.best_model_file = os.path.join(log_dir, "%s_model_best.pickle" % self.run_id)
        tb_log_dir = os.path.join(log_dir, "tb", self.run_id)
        os.makedirs(tb_log_dir, exist_ok=True)

        # Collect parameters:
        params = self.default_params()
        config_file = args.get('--config-file')
        if config_file is not None:
            with open(config_file, 'r') as f:
                params.update(json.load(f))
        config = args.get('--config')
        if config is not None:
            params.update(json.loads(config))
        self.params = params
        with open(os.path.join(log_dir, "%s_params.json" % self.run_id), "w") as f:
            json.dump(params, f)
        print("Run %s starting with following parameters:\n%s" % (self.run_id, json.dumps(self.params)))
        random.seed(params['random_seed'])
        np.random.seed(params['random_seed'])

        # Load data:
        self.max_num_vertices = 0
        self.num_edge_types = 0
        self.annotation_size = 0
        self.train_data = self.load_data(params['train_file'], is_training_data=True)
        self.valid_data = self.load_data(params['valid_file'], is_training_data=False)

        # Build the actual model
        config = tf.ConfigProto()
        config.gpu_options.allow_growth = True
        self.graph = tf.Graph()
        self.sess = tf.Session(graph=self.graph, config=config)
        with self.graph.as_default():
            tf.set_random_seed(params['random_seed'])
            self.placeholders = {}
            self.weights = {}
            self.ops = {}
            self.make_model()
            self.make_train_step()
            self.make_summaries()

            # Restore/initialize variables:
            restore_file = args.get('--restore')
            if restore_file is not None:
                self.train_step_id, self.valid_step_id = self.restore_progress(restore_file)
            else:
                self.initialize_model()
                self.train_step_id = 0
                self.valid_step_id = 0
            self.train_writer = tf.summary.FileWriter(os.path.join(tb_log_dir, 'train'), graph=self.graph)
            self.valid_writer = tf.summary.FileWriter(os.path.join(tb_log_dir, 'validation'), graph=self.graph)

    def load_data(self, file_name, is_training_data: bool):
        full_path = os.path.join(self.data_dir, file_name)

        print("Loading data from %s" % full_path)
        with open(full_path, 'r') as f:
            data = json.load(f)

        restrict = self.args.get("--restrict_data")
        if restrict is not None and restrict > 0:
            data = data[:restrict]

        # Get some common data out:
        num_fwd_edge_types = 0
        for g in data:
            self.max_num_vertices = max(self.max_num_vertices, max([v for e in g['graph'] for v in [e[0], e[2]]]))
            num_fwd_edge_types = max(num_fwd_edge_types, max([e[1] for e in g['graph']]))
        self.num_edge_types = max(self.num_edge_types, num_fwd_edge_types * (1 if self.params['tie_fwd_bkwd'] else 2))
        self.annotation_size = max(self.annotation_size, len(data[0]["node_features"][0]))

        return self.process_raw_graphs(data, is_training_data)

    @staticmethod
    def graph_string_to_array(graph_string: str) -> List[List[int]]:
        return [[int(v) for v in s.split(' ')]
                for s in graph_string.split('\n')]

    def process_raw_graphs(self, raw_data: Sequence[Any], is_training_data: bool) -> Any:
        raise Exception("Models have to implement process_raw_graphs!")

    def make_model(self):
        self.placeholders['target_values'] = tf.placeholder(tf.float32, [len(self.params['task_ids']), None],
                                                            name='target_values')
        self.placeholders['target_mask'] = tf.placeholder(tf.float32, [len(self.params['task_ids']), None],
                                                          name='target_mask')
        self.placeholders['num_graphs'] = tf.placeholder(tf.int32, [], name='num_graphs')
        self.placeholders['out_layer_dropout_keep_prob'] = tf.placeholder(tf.float32, [], name='out_layer_dropout_keep_prob')

        with tf.variable_scope("graph_model"):
            self.prepare_specific_graph_model()
            # This does the actual graph work:
            if self.params['use_graph']:
                self.ops['final_node_representations'] = self.compute_final_node_representations()
            else:
                self.ops['final_node_representations'] = tf.zeros_like(self.placeholders['initial_node_representation'])

        self.ops['losses'] = []
        for (internal_id, task_id) in enumerate(self.params['task_ids']):
            with tf.variable_scope("out_layer_task%i" % task_id):
                with tf.variable_scope("regression_gate"):
                    self.weights['regression_gate_task%i' % task_id] = MLP(2 * self.params['hidden_size'], 1, [],
                                                                           self.placeholders['out_layer_dropout_keep_prob'])
                with tf.variable_scope("regression"):
                    self.weights['regression_transform_task%i' % task_id] = MLP(self.params['hidden_size'], 1, [],
                                                                                self.placeholders['out_layer_dropout_keep_prob'])
                computed_values = self.gated_regression(self.ops['final_node_representations'],
                                                        self.weights['regression_gate_task%i' % task_id],
                                                        self.weights['regression_transform_task%i' % task_id])
                diff = computed_values - self.placeholders['target_values'][internal_id, :]
                task_target_mask = self.placeholders['target_mask'][internal_id, :]
                task_target_num = tf.reduce_sum(task_target_mask) + SMALL_NUMBER
                diff = diff * task_target_mask  # Mask out unused values
                self.ops['accuracy_task%i' % task_id] = tf.reduce_sum(tf.abs(diff)) / task_target_num
                task_loss = tf.reduce_sum(0.5 * tf.square(diff)) / task_target_num
                # Normalise loss to account for fewer task-specific examples in batch:
                task_loss = task_loss * (1.0 / (self.params['task_sample_ratios'].get(task_id) or 1.0))
                self.ops['losses'].append(task_loss)
        self.ops['loss'] = tf.reduce_sum(self.ops['losses'])

    def make_train_step(self):
        trainable_vars = self.sess.graph.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
        if self.args.get('--freeze-graph-model'):
            graph_vars = set(self.sess.graph.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="graph_model"))
            filtered_vars = []
            for var in trainable_vars:
                if var not in graph_vars:
                    filtered_vars.append(var)
                else:
                    print("Freezing weights of variable %s." % var.name)
            trainable_vars = filtered_vars
        optimizer = tf.train.AdamOptimizer(self.params['learning_rate'])
        grads_and_vars = optimizer.compute_gradients(self.ops['loss'], var_list=trainable_vars)
        clipped_grads = []
        for grad, var in grads_and_vars:
            if grad is not None:
                clipped_grads.append((tf.clip_by_norm(grad, self.params['clamp_gradient_norm']), var))
            else:
                clipped_grads.append((grad, var))
        self.ops['train_step'] = optimizer.apply_gradients(clipped_grads)
        # Initialize newly-introduced variables:
        self.sess.run(tf.local_variables_initializer())

    def make_summaries(self):
        with tf.name_scope('summary'):
            tf.summary.scalar('loss', self.ops['loss'])
            for task_id in self.params['task_ids']:
                tf.summary.scalar('accuracy%i' % task_id, self.ops['accuracy_task%i' % task_id])
        self.ops['summary'] = tf.summary.merge_all()

    def gated_regression(self, last_h, regression_gate, regression_transform):
        raise Exception("Models have to implement gated_regression!")

    def prepare_specific_graph_model(self) -> None:
        raise Exception("Models have to implement prepare_specific_graph_model!")

    def compute_final_node_representations(self) -> tf.Tensor:
        raise Exception("Models have to implement compute_final_node_representations!")

    def make_minibatch_iterator(self, data: Any, is_training: bool):
        raise Exception("Models have to implement make_minibatch_iterator!")

    def run_epoch(self, epoch_name: str, data, is_training: bool, start_step: int = 0):
        chemical_accuracies = np.array([0.066513725, 0.012235489, 0.071939046, 0.033730778, 0.033486113, 0.004278493,
                                        0.001330901, 0.004165489, 0.004128926, 0.00409976, 0.004527465, 0.012292586,
                                        0.037467458])

        loss = 0
        accuracies = []
        accuracy_ops = [self.ops['accuracy_task%i' % task_id] for task_id in self.params['task_ids']]
        start_time = time.time()
        processed_graphs = 0
        steps = 0
        batch_iterator = ThreadedIterator(self.make_minibatch_iterator(data, is_training), max_queue_size=5)
        for step, batch_data in enumerate(batch_iterator):
            num_graphs = batch_data[self.placeholders['num_graphs']]
            processed_graphs += num_graphs
            if is_training:
                batch_data[self.placeholders['out_layer_dropout_keep_prob']] = self.params['out_layer_dropout_keep_prob']
                fetch_list = [self.ops['loss'], accuracy_ops, self.ops['summary'], self.ops['train_step']]
            else:
                batch_data[self.placeholders['out_layer_dropout_keep_prob']] = 1.0
                fetch_list = [self.ops['loss'], accuracy_ops, self.ops['summary']]
            result = self.sess.run(fetch_list, feed_dict=batch_data)
            (batch_loss, batch_accuracies, batch_summary) = (result[0], result[1], result[2])
            writer = self.train_writer if is_training else self.valid_writer
            writer.add_summary(batch_summary, start_step + step)
            loss += batch_loss * num_graphs
            accuracies.append(np.array(batch_accuracies) * num_graphs)

            print("Running %s, batch %i (has %i graphs). Loss so far: %.4f" % (epoch_name,
                                                                               step,
                                                                               num_graphs,
                                                                               loss / processed_graphs),
                  end='\r')
            steps += 1

        accuracies = np.sum(accuracies, axis=0) / processed_graphs
        loss = loss / processed_graphs
        error_ratios = accuracies / chemical_accuracies[self.params["task_ids"]]
        instance_per_sec = processed_graphs / (time.time() - start_time)
        return loss, accuracies, error_ratios, instance_per_sec, steps

    def train(self):
        log_to_save = []
        total_time_start = time.time()
        with self.graph.as_default():
            if self.args.get('--restore') is not None:
                _, valid_accs, _, _, steps = self.run_epoch("Resumed (validation)", self.valid_data, False)
                best_val_acc = np.sum(valid_accs)
                best_val_acc_epoch = 0
                print("\r\x1b[KResumed operation, initial cum. val. acc: %.5f" % best_val_acc)
            else:
                (best_val_acc, best_val_acc_epoch) = (float("+inf"), 0)
            for epoch in range(1, self.params['num_epochs'] + 1):
                print("== Epoch %i" % epoch)
                train_loss, train_accs, train_errs, train_speed, train_steps = self.run_epoch("epoch %i (training)" % epoch,
                                                                                              self.train_data, True, self.train_step_id)
                self.train_step_id += train_steps
                accs_str = " ".join(["%i:%.5f" % (id, acc) for (id, acc) in zip(self.params['task_ids'], train_accs)])
                errs_str = " ".join(["%i:%.5f" % (id, err) for (id, err) in zip(self.params['task_ids'], train_errs)])
                print("\r\x1b[K Train: loss: %.5f | acc: %s | error_ratio: %s | instances/sec: %.2f" % (train_loss,
                                                                                                        accs_str,
                                                                                                        errs_str,
                                                                                                        train_speed))
                valid_loss, valid_accs, valid_errs, valid_speed, valid_steps = self.run_epoch("epoch %i (validation)" % epoch,
                                                                                              self.valid_data, False, self.valid_step_id)
                self.valid_step_id += valid_steps
                accs_str = " ".join(["%i:%.5f" % (id, acc) for (id, acc) in zip(self.params['task_ids'], valid_accs)])
                errs_str = " ".join(["%i:%.5f" % (id, err) for (id, err) in zip(self.params['task_ids'], valid_errs)])
                print("\r\x1b[K Valid: loss: %.5f | acc: %s | error_ratio: %s | instances/sec: %.2f" % (valid_loss,
                                                                                                        accs_str,
                                                                                                        errs_str,
                                                                                                        valid_speed))

                epoch_time = time.time() - total_time_start
                log_entry = {
                    'epoch': epoch,
                    'time': epoch_time,
                    'train_results': (train_loss, train_accs.tolist(), train_errs.tolist(), train_speed),
                    'valid_results': (valid_loss, valid_accs.tolist(), valid_errs.tolist(), valid_speed),
                }
                log_to_save.append(log_entry)
                with open(self.log_file, 'w') as f:
                    json.dump(log_to_save, f, indent=4)

                val_acc = np.sum(valid_accs)  # type: float
                if val_acc < best_val_acc:
                    self.save_progress(self.best_model_file, self.train_step_id, self.valid_step_id)
                    print("  (Best epoch so far, cum. val. acc decreased to %.5f from %.5f. Saving to '%s')" % (
                        val_acc, best_val_acc, self.best_model_file))
                    best_val_acc = val_acc
                    best_val_acc_epoch = epoch
                elif epoch - best_val_acc_epoch >= self.params['patience']:
                    print("Stopping training after %i epochs without improvement on validation accuracy." % self.params['patience'])
                    break

    def save_progress(self, model_path: str, train_step: int, valid_step: int) -> None:
        weights_to_save = {}
        for variable in self.sess.graph.get_collection(tf.GraphKeys.GLOBAL_VARIABLES):
            assert variable.name not in weights_to_save
            weights_to_save[variable.name] = self.sess.run(variable)

        data_to_save = {
            "params": self.params,
            "weights": weights_to_save,
            "train_step": train_step,
            "valid_step": valid_step,
        }

        with open(model_path, 'wb') as out_file:
            pickle.dump(data_to_save, out_file, pickle.HIGHEST_PROTOCOL)

    def initialize_model(self) -> None:
        init_op = tf.group(tf.global_variables_initializer(),
                           tf.local_variables_initializer())
        self.sess.run(init_op)

    def restore_progress(self, model_path: str) -> (int, int):
        print("Restoring weights from file %s." % model_path)
        with open(model_path, 'rb') as in_file:
            data_to_load = pickle.load(in_file)

        # Assert that we got the same model configuration
        assert len(self.params) == len(data_to_load['params'])
        for (par, par_value) in self.params.items():
            # Fine to have different task_ids:
            if par not in ['task_ids', 'num_epochs']:
                assert par_value == data_to_load['params'][par]

        variables_to_initialize = []
        with tf.name_scope("restore"):
            restore_ops = []
            used_vars = set()
            for variable in self.sess.graph.get_collection(tf.GraphKeys.GLOBAL_VARIABLES):
                used_vars.add(variable.name)
                if variable.name in data_to_load['weights']:
                    restore_ops.append(variable.assign(data_to_load['weights'][variable.name]))
                else:
                    print('Freshly initializing %s since no saved value was found.' % variable.name)
                    variables_to_initialize.append(variable)
            for var_name in data_to_load['weights']:
                if var_name not in used_vars:
                    print('Saved weights for %s not used by model.' % var_name)
            restore_ops.append(tf.variables_initializer(variables_to_initialize))
            self.sess.run(restore_ops)

        return data_to_load['train_step'], data_to_load['valid_step']