diff --git a/vdpwi/__main__.py b/vdpwi/__main__.py
new file mode 100644
index 0000000..05a56be
--- /dev/null
+++ b/vdpwi/__main__.py
@@ -0,0 +1,139 @@
+from collections import namedtuple
+
+from tqdm import tqdm
+import numpy as np
+import scipy.stats as stats
+import torch
+import torch.optim as optim
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils as utils
+
+from utils.log import LogWriter
+import data
+import model as mod
+
+Context = namedtuple("Context", "model, train_loader, dev_loader, test_loader, optimizer, criterion, params, log_writer")
+EvaluateResult = namedtuple("EvaluateResult", "pearsonr, spearmanr")
+
+def create_context(config):
+    def collate_fn(batch):
+        emb1 = []
+        emb2 = []
+        labels = []
+        cmp_labels = []
+        pad_cube = []
+        max_len1 = 0; max_len2 = 0
+
+        for s1, s2, l, cl in batch:
+            emb1.append(s1)
+            emb2.append(s2)
+            max_len1 = max(max_len1, len(s1))
+            max_len2 = max(max_len2, len(s2))
+            labels.append(l)
+            cmp_labels.append(cl)
+
+        for s1, s2 in zip(emb1, emb2):
+            pad1 = (max_len1 - len(s1))
+            pad2 = (max_len2 - len(s2))
+            pad_mask = np.ones((max_len1, max_len2))
+            pad_mask[:len(s1), :len(s2)] = 0
+            pad_cube.append(pad_mask)
+            s1.extend([embedding.weight.size(0) - 1] * pad1)
+            s2.extend([embedding.weight.size(0) - 1] * pad2)
+
+        pad_cube = np.array(pad_cube)
+        emb1 = torch.LongTensor(emb1)
+        emb2 = torch.LongTensor(emb2)
+        labels = torch.Tensor(labels)
+        emb1 = torch.autograd.Variable(emb1, requires_grad=False)
+        emb2 = torch.autograd.Variable(emb2, requires_grad=False)
+        labels = torch.autograd.Variable(labels, requires_grad=False)
+        pad_cube = torch.autograd.Variable(torch.from_numpy(pad_cube).float(), requires_grad=False)
+        if not config.cpu:
+            emb1 = emb1.cuda()
+            emb2 = emb2.cuda()
+            labels = labels.cuda()
+            pad_cube = pad_cube.cuda()
+        return emb1, emb2, labels, pad_cube, cmp_labels
+
+    embedding, (train_set, dev_set, test_set) = data.load_dataset(config.dataset)
+    model = mod.VDPWIModel(embedding, config)
+    if config.restore:
+        model.load(config.input_file)
+    if not config.cpu:
+        model = model.cuda()
+
+    train_loader = utils.data.DataLoader(train_set, shuffle=True, batch_size=config.mbatch_size, collate_fn=collate_fn)
+    dev_loader = utils.data.DataLoader(dev_set, batch_size=1, collate_fn=collate_fn)
+    test_loader = utils.data.DataLoader(test_set, batch_size=1, collate_fn=collate_fn)
+
+    params = list(filter(lambda x: x.requires_grad, model.parameters()))
+    if config.optimizer == "adam":
+        optimizer = optim.Adam(params, lr=config.lr, weight_decay=config.weight_decay)
+    elif config.optimizer == "sgd":
+        optimizer = optim.SGD(params, lr=config.lr, momentum=config.momentum, weight_decay=config.weight_decay)
+    elif config.optimizer == "rmsprop":
+        optimizer = optim.RMSprop(params, lr=config.lr, alpha=config.decay, momentum=config.momentum, weight_decay=config.weight_decay)
+    criterion = nn.KLDivLoss()
+    log_writer = LogWriter()
+    return Context(model, train_loader, dev_loader, test_loader, optimizer, criterion, params, log_writer)
+
+def test(config):
+    context = create_context(config)
+    result = evaluate(context, context.test_loader)
+    print("Final test result: {}".format(result))
+
+def evaluate(context, data_loader):
+    model = context.model
+    model.eval()
+    predictions = []
+    true_labels = []
+    for sent1, sent2, _, pad_cube, truth in data_loader:
+        scores = model(sent1, sent2, pad_cube)
+        scores = F.softmax(scores).cpu().data.numpy()[0]
+        prediction = np.dot(np.arange(1, len(scores) + 1), scores)
+        predictions.append(prediction); true_labels.append(truth[0][0])
+    
+    pearsonr = stats.pearsonr(predictions, true_labels)[0]
+    spearmanr = stats.spearmanr(predictions, true_labels)[0]
+    context.log_writer.log_dev_metrics(pearsonr, spearmanr)
+    return EvaluateResult(pearsonr, spearmanr)
+
+def train(config):
+    context = create_context(config)
+    context.log_writer.log_hyperparams()
+    best_dev_pr = 0
+    for epoch_no in range(config.n_epochs):
+        print("Epoch number: {}".format(epoch_no + 1))
+        loader_wrapper = tqdm(context.train_loader, total=len(context.train_loader), desc="Loss")
+        context.model.train()
+        loss = 0
+        for sent1, sent2, label_pmf, pad_cube, _ in loader_wrapper:
+            context.optimizer.zero_grad()
+            scores = F.log_softmax(context.model(sent1, sent2, pad_cube))
+
+            loss = context.criterion(scores, label_pmf)
+            loss.backward()
+            nn.utils.clip_grad_norm(context.params, config.clip_norm)
+            context.optimizer.step()
+
+            loss = loss.cpu().data[0]
+            loader_wrapper.set_description("Loss: {:<8}".format(round(loss, 5)))
+            context.log_writer.log_train_loss(loss)
+        result = evaluate(context, context.dev_loader)
+        print("Dev result: {}".format(result))
+        if best_dev_pr < result.pearsonr:
+            best_dev_pr = result.pearsonr
+            print("Saving best model...")
+            context.model.save(config.output_file)
+
+def main():
+    config = data.Configs.base_config()
+    if config.mode == "train":
+        train(config)
+    elif config.mode == "test":
+        test(config)
+
+if __name__ == "__main__":
+    main()
\ No newline at end of file
diff --git a/vdpwi/data.py b/vdpwi/data.py
new file mode 100644
index 0000000..04cd560
--- /dev/null
+++ b/vdpwi/data.py
@@ -0,0 +1,104 @@
+import argparse
+import os
+
+import torch
+import torch.nn as nn
+import torch.utils.data as data
+
+class Configs(object):
+    @staticmethod
+    def base_config():
+        parser = argparse.ArgumentParser()
+        parser.add_argument("--classifier", type=str, default="vdpwi", choices=["vdpwi", "resnet"])
+        parser.add_argument("--clip_norm", type=float, default=50)
+        parser.add_argument("--cpu", action="store_true", default=False)
+        parser.add_argument("--dataset", type=str, default="sick", choices=["sick"])
+        parser.add_argument("--decay", type=float, default=0.95)
+        parser.add_argument("--input_file", type=str, default="local_saves/model.pt")
+        parser.add_argument("--lr", type=float, default=5E-4)
+        parser.add_argument("--mbatch_size", type=int, default=16)
+        parser.add_argument("--mode", type=str, default="train", choices=["train", "test"])
+        parser.add_argument("--momentum", type=float, default=0.1)
+        parser.add_argument("--n_epochs", type=int, default=35)
+        parser.add_argument("--n_labels", type=int, default=5)
+        parser.add_argument("--optimizer", type=str, default="rmsprop", choices=["adam", "sgd", "rmsprop"])
+        parser.add_argument("--output_file", type=str, default="local_saves/model.pt")
+        parser.add_argument("--res_fmaps", type=int, default=32)
+        parser.add_argument("--res_layers", type=int, default=16)
+        parser.add_argument("--restore", action="store_true", default=False)
+        parser.add_argument("--rnn_hidden_dim", type=int, default=250)
+        parser.add_argument("--weight_decay", type=float, default=1E-5)
+        parser.add_argument("--wordvecs_file", type=str, default="local_data/glove/glove.840B.300d.txt")
+        return parser.parse_known_args()[0]
+
+    @staticmethod
+    def sick_config():
+        parser = argparse.ArgumentParser()
+        parser.add_argument("--n_labels", type=int, default=5)
+        parser.add_argument("--sick_cache", type=str, default="local_data/sick/.vec-cache")
+        parser.add_argument("--sick_data", type=str, default="local_data/sick")
+        return parser.parse_known_args()[0]
+
+class LabeledEmbeddedDataset(data.Dataset):
+    def __init__(self, sentence_indices1, sentence_indices2, labels, compare_labels=None):
+        assert len(sentence_indices1) == len(labels) == len(sentence_indices2)
+        self.sentence_indices1 = sentence_indices1
+        self.sentence_indices2 = sentence_indices2
+        self.labels = labels
+        self.compare_labels = compare_labels
+
+    def __getitem__(self, idx):
+        cmp_lbl = None if self.compare_labels is None else self.compare_labels[idx]
+        return self.sentence_indices1[idx], self.sentence_indices2[idx], self.labels[idx], cmp_lbl
+
+    def __len__(self):
+        return len(self.labels)
+
+def load_sick():
+    config = Configs.sick_config()
+    def fetch_indices(name):
+        sentence_indices = []
+        filename = os.path.join(config.sick_data, dataset, name)
+        with open(filename) as f:
+            for line in f:
+                indices = [embed_ids.get(word, -1) for word in line.strip().split()]
+                indices = list(filter(lambda x: x >= 0, indices))
+                sentence_indices.append(indices)
+        return sentence_indices
+
+    def read_labels(filename):
+        labels = []
+        with open(filename) as f:
+            for line in f:
+                labels.append([float(val) for val in line.split()])
+        return labels
+
+    sets = []
+    embeddings = []
+    embed_ids = {}
+    with open(os.path.join(config.sick_cache)) as f:
+        for i, line in enumerate(f):
+            word, vec = line.split(" ", 1)
+            vec = list(map(float, vec.strip().split()))
+            embed_ids[word] = i
+            embeddings.append(vec)
+    padding_idx = len(embeddings)
+    embeddings.append([0.0] * 300)
+
+    for dataset in ("train", "dev", "test"):
+        sparse_filename = os.path.join(config.sick_data, dataset, "sim_sparse.txt")
+        truth_filename = os.path.join(config.sick_data, dataset, "sim.txt")
+        sparse_labels = read_labels(sparse_filename)
+        cmp_labels = read_labels(truth_filename)
+        indices1 = fetch_indices("a.toks")
+        indices2 = fetch_indices("b.toks")
+        sets.append(LabeledEmbeddedDataset(indices1, indices2, sparse_labels, cmp_labels))
+    embedding = nn.Embedding(len(embeddings), 300)
+    embedding.weight.data.copy_(torch.Tensor(embeddings))
+    embedding.weight.requires_grad = False
+    return embedding, sets
+
+def load_dataset(dataset):
+    return _loaders[dataset]()
+
+_loaders = dict(sick=load_sick)
diff --git a/vdpwi/model.py b/vdpwi/model.py
new file mode 100644
index 0000000..c31bcbc
--- /dev/null
+++ b/vdpwi/model.py
@@ -0,0 +1,168 @@
+from torch.autograd import Variable
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision.models as models
+import numpy as np
+
+class SerializableModule(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def save(self, filename):
+        torch.save(self.state_dict(), filename)
+
+    def load(self, filename):
+        self.load_state_dict(torch.load(filename, map_location=lambda storage, loc: storage))
+
+def hard_pad2d(x, pad):
+    def pad_side(idx):
+        pad_len = max(pad - x.size(idx), 0)
+        return [0, pad_len]
+    padding = pad_side(3)
+    padding.extend(pad_side(2))
+    x = F.pad(x, padding)
+    return x[:, :, :pad, :pad]
+
+class ResNet(SerializableModule):
+    def __init__(self, config):
+        super().__init__()
+        n_layers = config.res_layers
+        n_maps = config.res_fmaps
+        n_labels = config.n_labels
+        self.conv0 = nn.Conv2d(12, n_maps, (3, 3), padding=1)
+        self.convs = [nn.Conv2d(n_maps, n_maps, (3, 3), padding=1) for _ in range(n_layers)]
+        self.output = nn.Linear(n_maps, n_labels)
+        self.input_len = None
+        for i, conv in enumerate(self.convs):
+            self.add_module("conv{}".format(i + 1), conv)
+
+    def forward(self, x):
+        x = F.relu(self.conv0(x))
+        old_x = x
+        for i, conv in enumerate(self.convs):
+            x = F.relu(conv(x))
+            if i % 2 == 1:
+                x += old_x
+                old_x = x
+        x = torch.mean(x.view(x.size(0), x.size(1), -1), 2)
+        return self.output(x)
+
+class VDPWIConvNet(SerializableModule):
+    def __init__(self, config):
+        super().__init__()
+        def make_conv(n_in, n_out):
+            conv = nn.Conv2d(n_in, n_out, 3, padding=1)
+            conv.bias.data.zero_()
+            nn.init.xavier_normal(conv.weight)
+            return conv
+        self.conv1 = make_conv(12, 128)
+        self.conv2 = make_conv(128, 164)
+        self.conv3 = make_conv(164, 192)
+        self.conv4 = make_conv(192, 192)
+        self.conv5 = make_conv(192, 128)
+        self.maxpool2 = nn.MaxPool2d(2, ceil_mode=True)
+        self.dnn = nn.Linear(128, 128)
+        self.output = nn.Linear(128, config.n_labels)
+        self.input_len = 32
+
+    def forward(self, x):
+        x = hard_pad2d(x, self.input_len)
+        pool_final = nn.MaxPool2d(2, ceil_mode=True) if x.size(2) == 32 else nn.MaxPool2d(3, 1, ceil_mode=True)
+        x = self.maxpool2(F.relu(self.conv1(x)))
+        x = self.maxpool2(F.relu(self.conv2(x)))
+        x = self.maxpool2(F.relu(self.conv3(x)))
+        x = self.maxpool2(F.relu(self.conv4(x)))
+        x = pool_final(F.relu(self.conv5(x)))
+        x = F.relu(self.dnn(x.view(x.size(0), -1)))
+        return self.output(x)
+
+class VDPWIModel(SerializableModule):
+    def __init__(self, embedding, config):
+        super().__init__()
+        self.hidden_dim = config.rnn_hidden_dim
+        self.rnn = nn.LSTM(300, self.hidden_dim, 1, batch_first=True)
+        self.embedding = embedding
+        self.use_cuda = not config.cpu
+        if config.classifier == "vdpwi":
+            self.classifier_net = VDPWIConvNet(config)
+        elif config.classifier == "resnet":
+            self.classifier_net = ResNet(config)
+
+    def compute_sim_cube(self, seq1, seq2):
+        def compute_sim(prism1, prism2):
+            prism1_len = prism1.norm(dim=3)
+            prism2_len = prism2.norm(dim=3)
+
+            dot_prod = torch.matmul(prism1.unsqueeze(3), prism2.unsqueeze(4))
+            dot_prod = dot_prod.squeeze(3).squeeze(3)
+            cos_dist = dot_prod / (prism1_len * prism2_len + 1E-8)
+            l2_dist = -((prism1 - prism2).norm(dim=3))
+            return torch.stack([dot_prod, cos_dist, l2_dist], 1)
+
+        def compute_prism(seq1, seq2):
+            prism1 = seq1.repeat(seq2.size(1), 1, 1, 1)
+            prism2 = seq2.repeat(seq1.size(1), 1, 1, 1)
+            prism1 = prism1.permute(1, 2, 0, 3).contiguous()
+            prism2 = prism2.permute(1, 0, 2, 3).contiguous()
+            return compute_sim(prism1, prism2)
+
+        sim_cube = Variable(torch.Tensor(seq1.size(0), 12, seq1.size(1), seq2.size(1)))
+        if self.use_cuda:
+            sim_cube = sim_cube.cuda()
+        seq1_f = seq1[:, :, :self.hidden_dim]
+        seq1_b = seq1[:, :, self.hidden_dim:]
+        seq2_f = seq2[:, :, :self.hidden_dim]
+        seq2_b = seq2[:, :, self.hidden_dim:]
+        sim_cube[:, 0:3] = compute_prism(seq1, seq2)
+        sim_cube[:, 3:6] = compute_prism(seq1_f, seq2_f)
+        sim_cube[:, 6:9] = compute_prism(seq1_b, seq2_b)
+        sim_cube[:, 9:12] = compute_prism(seq1_f + seq1_b, seq2_f + seq2_b)
+        return sim_cube
+
+    def compute_focus_cube(self, sim_cube, pad_cube):
+        neg_magic = -10000
+        pad_cube = pad_cube.repeat(12, 1, 1, 1)
+        pad_cube = pad_cube.permute(1, 0, 2, 3).contiguous()
+        sim_cube = neg_magic * pad_cube + sim_cube
+        mask = Variable(torch.Tensor(*sim_cube.size()))
+        if self.use_cuda:
+            mask = mask.cuda()
+        mask[:, :, :, :] = 0.1
+
+        def build_mask(index):
+            max_mask = sim_cube[:, index].clone()
+            for _ in range(min(sim_cube.size(2), sim_cube.size(3))):
+                values, indices = torch.max(max_mask.view(sim_cube.size(0), -1), 1)
+                row_indices = indices / sim_cube.size(3)
+                col_indices = indices % sim_cube.size(3)
+                row_indices = row_indices.unsqueeze(1)
+                col_indices = col_indices.unsqueeze(1).unsqueeze(1)
+                for i, (row_i, col_i, val) in enumerate(zip(row_indices, col_indices, values)):
+                    if val < neg_magic / 2:
+                        continue
+                    mask[i, :, row_i, col_i] = 1
+                    max_mask[i, row_i, :] = neg_magic
+                    max_mask[i, :, col_i] = neg_magic
+        build_mask(9)
+        build_mask(10)
+        focus_cube = mask * sim_cube * (1 - pad_cube)
+        return focus_cube
+
+    def forward(self, x1, x2, pad_cube):
+        x1 = self.embedding(x1)
+        x2 = self.embedding(x2)
+        seq1f, _ = self.rnn(x1)
+        seq2f, _ = self.rnn(x2)
+        seq1b, _ = self.rnn(torch.cat(x1.split(1, 1)[::-1], 1))
+        seq2b, _ = self.rnn(torch.cat(x2.split(1, 1)[::-1], 1))
+        seq1 = torch.cat([seq1f, seq1b], 2)
+        seq2 = torch.cat([seq2f, seq2b], 2)
+        sim_cube = self.compute_sim_cube(seq1, seq2)
+        truncate = self.classifier_net.input_len
+        if truncate is not None:
+            sim_cube = sim_cube[:, :, :truncate, :truncate].contiguous()
+            pad_cube = pad_cube[:, :truncate, :truncate].contiguous()
+        focus_cube = self.compute_focus_cube(sim_cube, pad_cube)
+        logits = self.classifier_net(focus_cube)
+        return logits
diff --git a/vdpwi/train_default.sh b/vdpwi/train_default.sh
new file mode 100755
index 0000000..230a237
--- /dev/null
+++ b/vdpwi/train_default.sh
@@ -0,0 +1,2 @@
+#!/bin/sh
+python . --clip_norm 50 --decay 0.95 --lr 1E-4 --mbatch_size 1 --momentum 0 --optimizer rmsprop --weight_decay 0
diff --git a/vdpwi/utils/__init__.py b/vdpwi/utils/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/vdpwi/utils/log.py b/vdpwi/utils/log.py
new file mode 100644
index 0000000..82f5eef
--- /dev/null
+++ b/vdpwi/utils/log.py
@@ -0,0 +1,26 @@
+import datetime
+import sys
+
+from tensorboardX import SummaryWriter
+
+class LogWriter(object):
+    def __init__(self, run_name_fmt="run_{}"):
+        self.writer = SummaryWriter()
+        self.run_name = run_name_fmt.format(datetime.datetime.now().strftime("%Y-%m-%d-%H:%M:%S"))
+        self.train_idx = 0
+        self.dev_idx = 0
+
+    def log_hyperparams(self):
+        self.writer.add_text("{}/hyperparams".format(self.run_name), " ".join(sys.argv))
+
+    def log_train_loss(self, loss):
+        self.writer.add_scalar("{}/train_loss".format(self.run_name), loss, self.train_idx)
+        self.train_idx += 1
+
+    def log_dev_metrics(self, pearsonr, spearmanr):
+        results = dict(pearsonr=pearsonr, spearmanr=spearmanr)
+        self.writer.add_scalars("{}/dev_metrics".format(self.run_name), results, self.dev_idx)
+        self.dev_idx += 1
+
+    def next(self):
+        self.i += 1
diff --git a/vdpwi/utils/preprocess.py b/vdpwi/utils/preprocess.py
new file mode 100644
index 0000000..c0b4e77
--- /dev/null
+++ b/vdpwi/utils/preprocess.py
@@ -0,0 +1,78 @@
+import argparse
+import os
+
+from scipy.special import erf
+from scipy.stats import truncnorm
+import numpy as np
+
+import data
+
+def build_vector_cache(glove_filename, vec_cache_filename, vocab):
+    print("Building vector cache...")
+    with open(glove_filename) as f, open(vec_cache_filename, "w") as f2:
+        for line in f:
+            tok, vec = line.split(" ", 1)
+            if tok in vocab:
+                vocab.remove(tok)
+                f2.write("{} {}".format(tok, vec))
+
+def discrete_tnorm(a, b, tgt_loc, sigma=1, n_steps=100):
+    def phi(zeta):
+        return 1 / (np.sqrt(2 * np.pi)) * np.exp(-0.5 * zeta**2)
+    def Phi(x):
+        return 0.5 * (1 + erf(x / np.sqrt(2)))
+    def tgt_loc_update(x):
+        y1 = phi((a - x) / sigma)
+        y2 = phi((b - x) / sigma)
+        x1 = Phi((b - x) / sigma)
+        x2 = Phi((a - x) / sigma)
+        denom = x1 - x2 + 1E-4
+        return y1 / denom - y2 / denom
+
+    x = tgt_loc
+    direction = np.sign(tgt_loc - (b - a))
+    for _ in range(n_steps):
+        x = tgt_loc - sigma * tgt_loc_update(x)
+    tn = truncnorm((a - x) / sigma, (b - x) / sigma, loc=x, scale=sigma)
+    rrange = np.arange(a, b + 1)
+    pmf = tn.pdf(rrange)
+    pmf /= np.sum(pmf)
+    return pmf
+
+def discrete_lerp(a, b, ground_truth):
+    pmf = np.zeros(b - a + 1)
+    c = int(np.ceil(ground_truth + 1E-8))
+    f = int(np.floor(ground_truth))
+    pmf[min(c - a, b - a)] = ground_truth - f
+    pmf[f - a] = c - ground_truth
+    return pmf
+
+def smoothed_labels(truth, n_labels):
+    return discrete_lerp(1, n_labels, truth)
+
+def preprocess(filename, output_name="sim_sparse.txt"):
+    print("Preprocessing {}...".format(filename))
+    with open(filename) as f:
+        values = [float(l.strip()) for l in f.readlines()]
+    values = [" ".join([str(l) for l in smoothed_labels(v, 5)]) for v in values]
+    with open(os.path.join(os.path.dirname(filename), output_name), "w") as f:
+        f.write("\n".join(values))
+
+def add_vocab(tok_filename, vocab):
+    with open(tok_filename) as f:
+        for line in f:
+            vocab.update(line.strip().split())
+
+def main():
+    base_conf = data.Configs.base_config()
+    sick_conf = data.Configs.sick_config()
+    sick_folder = sick_conf.sick_data
+    vocab = set()
+    for name in ("train", "dev", "test"):
+        preprocess(os.path.join(sick_folder, name, "sim.txt"))
+        add_vocab(os.path.join(sick_folder, name, "a.toks"), vocab)
+        add_vocab(os.path.join(sick_folder, name, "b.toks"), vocab)
+    build_vector_cache(base_conf.wordvecs_file, sick_conf.sick_cache, vocab)
+
+if __name__ == "__main__":
+    main()
diff --git a/vdpwi/utils/tune.py b/vdpwi/utils/tune.py
new file mode 100644
index 0000000..bd4df6c
--- /dev/null
+++ b/vdpwi/utils/tune.py
@@ -0,0 +1,37 @@
+import os
+import random
+
+class RandomParamIterator(object):
+    def __init__(self, param_sets):
+        self.param_sets = param_sets
+
+    def random_param_set(self):
+        param_set = {}
+        for param_key, param_values in self.param_sets.items():
+            param_set[param_key] = random.choice(param_values)
+        return param_set
+
+class Tuner(object):
+    def __init__(self, *iterators, limit=500):
+        self.iterators = iterators
+        self.limit = limit
+
+    def start(self):
+        for i in range(self.limit):
+            iterator = random.choice(self.iterators)
+            params = iterator.random_param_set()
+            print(params)
+            arg_str = " ".join("--{}={}".format(k, v) for k, v in params.items())
+            os.system("python . {} --output_file local_saves/model{}.pt".format(arg_str, i))
+
+def main():
+    vgg_param_sets = dict(classifer=["vdpwi"], decay=[0.99, 0.95], lr=[5E-4, 1E-4], mbatch_size=[8, 16], 
+        optimizer=["adam", "rmsprop"], weight_decay=[0, 1E-5, 5E-4], momentum=[0, 0.15, 0.05])
+    res_param_sets = dict(classifier=["resnet"], clip_norm=[5, 7, 9], decay=[0.9, 0.95], lr=[5E-3, 1E-3, 5E-4],
+        mbatch_size=[8, 16, 32], rnn_hidden_dim=[150, 250, 300], res_fmaps=[16, 24, 32], res_layers=[4, 8, 16, 24])
+    vgg_iterator = RandomParamIterator(vgg_param_sets)
+    res_iterator = RandomParamIterator(res_param_sets)
+    Tuner(vgg_iterator).start()
+
+if __name__ == "__main__":
+    main()
\ No newline at end of file