(models, features, params)
| 149 | |
| 150 | |
| 151 | def create_sampling_graph(models, features, params): |
| 152 | if not isinstance(models, (list, tuple)): |
| 153 | raise ValueError("'models' must be a list or tuple") |
| 154 | |
| 155 | features = copy.copy(features) |
| 156 | model_fns = [model.get_inference_func() for model in models] |
| 157 | |
| 158 | num_samples = params.num_samples |
| 159 | |
| 160 | # Compute initial state if necessary |
| 161 | states = [] |
| 162 | funcs = [] |
| 163 | |
| 164 | for model_fn in model_fns: |
| 165 | if callable(model_fn): |
| 166 | # For non-incremental decoding |
| 167 | states.append({}) |
| 168 | funcs.append(model_fn) |
| 169 | else: |
| 170 | # For incremental decoding where model_fn is a tuple: |
| 171 | # (encoding_fn, decoding_fn) |
| 172 | states.append(model_fn[0](features)) |
| 173 | funcs.append(model_fn[1]) |
| 174 | |
| 175 | batch_size = tf.shape(features["source"])[0] |
| 176 | pad_id = params.mapping["target"][params.pad] |
| 177 | bos_id = params.mapping["target"][params.bos] |
| 178 | eos_id = params.mapping["target"][params.eos] |
| 179 | |
| 180 | # Expand the inputs |
| 181 | features["source"] = utils.tile_batch(features["source"], num_samples) |
| 182 | features["source_length"] = utils.tile_batch(features["source_length"], |
| 183 | num_samples) |
| 184 | |
| 185 | min_length = tf.to_float(features["source_length"]) |
| 186 | max_length = tf.to_float(features["source_length"]) |
| 187 | |
| 188 | if params.min_length_ratio: |
| 189 | min_length = min_length * params.min_length_ratio |
| 190 | |
| 191 | if params.max_length_ratio: |
| 192 | max_length = max_length * params.max_length_ratio |
| 193 | |
| 194 | if params.min_sample_length: |
| 195 | min_length = min_length - params.min_sample_length |
| 196 | |
| 197 | if params.max_sample_length: |
| 198 | max_length = max_length + params.max_sample_length |
| 199 | |
| 200 | min_length = tf.to_int32(min_length) |
| 201 | max_length = tf.to_int32(max_length) |
| 202 | |
| 203 | decoding_fn = _get_inference_fn(funcs, features) |
| 204 | states = nest.map_structure(lambda x: utils.tile_batch(x, num_samples), |
| 205 | states) |
| 206 | |
| 207 | seqs, scores = random_sample(decoding_fn, states, batch_size * num_samples, |
| 208 | min_length, max_length, pad_id, bos_id, |
nothing calls this directly
no test coverage detected