|
| 1 | +# Copyright (c) Meta Platforms, Inc. and affiliates. |
| 2 | +# All rights reserved. |
| 3 | +# |
| 4 | +# This source code is licensed under the BSD-style license found in the |
| 5 | +# LICENSE file in the root directory of this source tree. |
| 6 | + |
| 7 | +import pytest |
| 8 | + |
| 9 | +import torch |
| 10 | +from tests.test_utils import assert_expected, init_weights_with_constant, set_rng_seed |
| 11 | +from torch import Tensor |
| 12 | +from torchmultimodal.modules.layers.transformer import ( |
| 13 | + TransformerEncoder, |
| 14 | + TransformerEncoderLayer, |
| 15 | +) |
| 16 | +from torchmultimodal.modules.layers.transformer import TransformerOutput |
| 17 | + |
| 18 | + |
| 19 | +@pytest.fixture(autouse=True) |
| 20 | +def random(): |
| 21 | + set_rng_seed(4) |
| 22 | + |
| 23 | + |
| 24 | +class TestTransformerEncoderLayer: |
| 25 | + @pytest.fixture |
| 26 | + def get_encoder_layer(self): |
| 27 | + def create_layer(norm_first): |
| 28 | + model = TransformerEncoderLayer( |
| 29 | + d_model=2, |
| 30 | + n_head=1, |
| 31 | + dim_feedforward=2, |
| 32 | + norm_first=norm_first, |
| 33 | + ) |
| 34 | + init_weights_with_constant(model) |
| 35 | + model.eval() |
| 36 | + return model |
| 37 | + |
| 38 | + return create_layer |
| 39 | + |
| 40 | + @pytest.fixture |
| 41 | + def inputs(self): |
| 42 | + return Tensor([[[1, 2], [4, 2], [1, 1]]]) |
| 43 | + |
| 44 | + @pytest.mark.parametrize( |
| 45 | + "norm_first, expected_output", |
| 46 | + [ |
| 47 | + (True, Tensor([[[15.0, 16.0], [18.0, 16.0], [15.0, 15.0]]])), |
| 48 | + (False, Tensor([[[0.0, 2.0], [2.0, 0.0], [1.0, 1.0]]])), |
| 49 | + ], |
| 50 | + ) |
| 51 | + def test_forward(self, norm_first, expected_output, inputs, get_encoder_layer): |
| 52 | + model = get_encoder_layer(norm_first) |
| 53 | + actual = model(inputs) |
| 54 | + assert_expected(actual, expected_output, rtol=0, atol=1e-4) |
| 55 | + |
| 56 | + @pytest.mark.parametrize( |
| 57 | + "norm_first", |
| 58 | + [(True,), (False,)], |
| 59 | + ) |
| 60 | + def test_scripting(self, norm_first, inputs, get_encoder_layer): |
| 61 | + model = get_encoder_layer(norm_first) |
| 62 | + scripted_model = torch.jit.script(model) |
| 63 | + assert_expected(scripted_model(inputs), model(inputs), rtol=0, atol=1e-4) |
| 64 | + |
| 65 | + |
| 66 | +class TestTransformerEncoder: |
| 67 | + @pytest.fixture |
| 68 | + def get_encoder(self): |
| 69 | + def create_encoder(norm_first, final_layer_norm_eps=None): |
| 70 | + model = TransformerEncoder( |
| 71 | + n_layer=2, |
| 72 | + d_model=2, |
| 73 | + n_head=1, |
| 74 | + dim_feedforward=2, |
| 75 | + norm_first=norm_first, |
| 76 | + final_layer_norm_eps=final_layer_norm_eps, |
| 77 | + ) |
| 78 | + init_weights_with_constant(model) |
| 79 | + model.eval() |
| 80 | + return model |
| 81 | + |
| 82 | + return create_encoder |
| 83 | + |
| 84 | + @pytest.fixture |
| 85 | + def inputs(self): |
| 86 | + return Tensor([[[2, 3], [1, 2]]]) |
| 87 | + |
| 88 | + @pytest.mark.parametrize( |
| 89 | + "norm_first, return_hidden_states, expected_output", |
| 90 | + [ |
| 91 | + ( |
| 92 | + True, |
| 93 | + False, |
| 94 | + TransformerOutput( |
| 95 | + last_hidden_state=Tensor([[[30.0, 31.0], [29.0, 30.0]]]) |
| 96 | + ), |
| 97 | + ), |
| 98 | + ( |
| 99 | + False, |
| 100 | + False, |
| 101 | + TransformerOutput(last_hidden_state=Tensor([[[0.0, 2.0], [0.0, 2.0]]])), |
| 102 | + ), |
| 103 | + ( |
| 104 | + True, |
| 105 | + True, |
| 106 | + TransformerOutput( |
| 107 | + last_hidden_state=Tensor([[[30.0, 31.0], [29.0, 30.0]]]), |
| 108 | + hidden_states=[ |
| 109 | + Tensor([[[16.0, 17.0], [15.0, 16.0]]]), |
| 110 | + Tensor([[[30.0, 31.0], [29.0, 30.0]]]), |
| 111 | + ], |
| 112 | + ), |
| 113 | + ), |
| 114 | + ( |
| 115 | + False, |
| 116 | + True, |
| 117 | + TransformerOutput( |
| 118 | + last_hidden_state=Tensor([[[0.0, 2.0], [0.0, 2.0]]]), |
| 119 | + hidden_states=[ |
| 120 | + Tensor([[[0.0, 2.0], [0.0, 2.0]]]), |
| 121 | + Tensor([[[0.0, 2.0], [0.0, 2.0]]]), |
| 122 | + ], |
| 123 | + ), |
| 124 | + ), |
| 125 | + ], |
| 126 | + ) |
| 127 | + def test_forward( |
| 128 | + self, norm_first, return_hidden_states, expected_output, inputs, get_encoder |
| 129 | + ): |
| 130 | + model = get_encoder(norm_first) |
| 131 | + actual = model(inputs, return_hidden_states=return_hidden_states) |
| 132 | + if expected_output.hidden_states is None: |
| 133 | + assert actual.hidden_states is None |
| 134 | + else: |
| 135 | + assert_expected(actual.hidden_states[0], inputs) |
| 136 | + for state_1, state_2 in zip( |
| 137 | + expected_output.hidden_states, actual.hidden_states[1:] |
| 138 | + ): |
| 139 | + assert_expected(state_1, state_2) |
| 140 | + |
| 141 | + assert actual.attentions == expected_output.attentions |
| 142 | + assert_expected( |
| 143 | + actual.last_hidden_state, |
| 144 | + expected_output.last_hidden_state, |
| 145 | + rtol=0, |
| 146 | + atol=1e-4, |
| 147 | + ) |
| 148 | + |
| 149 | + @pytest.mark.parametrize( |
| 150 | + "norm_first, expected_output", |
| 151 | + [ |
| 152 | + ( |
| 153 | + True, |
| 154 | + TransformerOutput( |
| 155 | + last_hidden_state=Tensor([[[1.9073e-05, 2.0], [2.2888e-05, 2.0]]]), |
| 156 | + hidden_states=[ |
| 157 | + Tensor([[[16.0, 17.0], [15.0, 16.0]]]), |
| 158 | + Tensor([[[30.0, 31.0], [29.0, 30.0]]]), |
| 159 | + ], |
| 160 | + ), |
| 161 | + ), |
| 162 | + ( |
| 163 | + False, |
| 164 | + TransformerOutput( |
| 165 | + last_hidden_state=Tensor([[[5.0068e-06, 2.0], [5.0068e-06, 2.0]]]), |
| 166 | + hidden_states=[ |
| 167 | + Tensor([[[0.0, 2.0], [0.0, 2.0]]]), |
| 168 | + Tensor([[[0.0, 2.0], [0.0, 2.0]]]), |
| 169 | + ], |
| 170 | + ), |
| 171 | + ), |
| 172 | + ], |
| 173 | + ) |
| 174 | + def test_forward_with_final_ln( |
| 175 | + self, norm_first, expected_output, inputs, get_encoder |
| 176 | + ): |
| 177 | + model = get_encoder(norm_first=norm_first, final_layer_norm_eps=1e-5) |
| 178 | + actual = model(inputs, return_hidden_states=True) |
| 179 | + assert_expected( |
| 180 | + expected_output.last_hidden_state, |
| 181 | + actual.last_hidden_state, |
| 182 | + rtol=0, |
| 183 | + atol=1e-4, |
| 184 | + ) |
| 185 | + for state_1, state_2 in zip( |
| 186 | + expected_output.hidden_states, actual.hidden_states[1:] |
| 187 | + ): |
| 188 | + assert_expected(state_1, state_2) |
| 189 | + |
| 190 | + @pytest.mark.parametrize( |
| 191 | + "norm_first", |
| 192 | + [(True,), (False,)], |
| 193 | + ) |
| 194 | + def test_scripting(self, norm_first, inputs, get_encoder): |
| 195 | + model = get_encoder(norm_first) |
| 196 | + scripted_model = torch.jit.script(model) |
| 197 | + assert_expected(scripted_model(inputs), model(inputs), rtol=0, atol=1e-4) |
0 commit comments