Libraries update for Stable Diffusion (#335)

mlperf_logging/benchmark_meta.py

@@ -10,6 +10,7 @@
         'minigo': 10,
         'resnet': 5,
         'ssd': 5,
+        'stable_diffusion': 10,
         'transformer': 10,
         'ncf': 10,
         'rnnt': 10,

mlperf_logging/compliance_checker/README.md

@@ -23,7 +23,7 @@ As log examples use [NVIDIA's training logs](https://github.com/mlperf/training_
 ### Existing config files for training submissions
 
     3.1.0/common.yaml          - currently the default config file, checks common fields complience and equeues benchmark-specific config file
-    3.1.0/closed_common.yaml   - the common rules file for closed submissions. These rules apply to all benchmarks 
+    3.1.0/closed_common.yaml   - the common rules file for closed submissions. These rules apply to all benchmarks
     3.1.0/open_common.yaml     - the common rules file for open submissions. These rules apply to all benchmarks
     3.1.0/closed_resnet.yaml   - Per-benchmark rules, closed submissions.
     3.1.0/closed_ssd.yaml
@@ -33,6 +33,7 @@ As log examples use [NVIDIA's training logs](https://github.com/mlperf/training_
     3.1.0/closed_bert.yaml
     3.1.0/closed_dlrm_dcnv2.yaml
     3.1.0/closed_gpt3.yaml
+    3.1.0/closed_stable_diffusion.yaml
     3.1.0/open_resnet.yaml   - Per-benchmark rules, closed submissions.
     3.1.0/open_ssd.yaml
     3.1.0/open_maskrcnn.yaml
@@ -41,6 +42,7 @@ As log examples use [NVIDIA's training logs](https://github.com/mlperf/training_
     3.1.0/open_bert.yaml
     3.1.0/open_dlrm_dcnv2.yaml
     3.1.0/open_gpt3.yaml
+    3.1.0/open_stable_diffusion.yaml
 
 ### Existing config files for HPC submissions
 
@@ -64,15 +66,15 @@ Compliance checking is done following below algorithm.
    2. If present, evaluate `CHECK` section, and raise an exception if the result is false
 7. Print all warning messages
 
-Possible side effects of yaml sections execution can be [printing output](#other-operations), or [enqueueing 
+Possible side effects of yaml sections execution can be [printing output](#other-operations), or [enqueueing
 additional yaml files to be verified](#enqueuing-additional-config-files).
 
 ### Config file syntax
 Rules to be checked are provided in yaml (config) file. A config file contains the following records:
 
 #### `BEGIN` record
 Defines `CODE` to be executed before any other rules defined in the current file. This record is optional
-and there can be up to a single `BEGIN` record per config file. 
+and there can be up to a single `BEGIN` record per config file.
 
 Example:
 
@@ -87,6 +89,7 @@ The following fields are optional:
 - `REQ` - specifies the requirement regarding occurrence. Possible values :
     - `EXACTLY_ONE` - current key has to appear exactly once
     - `AT_LEAST_ONE` - current key has to appear at least once
+    - `AT_LEAST(n)` - current key has to appear at least n times
     - `AT_LEAST_ONE_OR(alternatives)` - current key or one of the alternative has to appear at least once;
             alternatives is a comma separated list of keys
 - `PRE` - code to be executed before performing checks
@@ -112,14 +115,14 @@ The following fields are optional:
 
 #### Global and local state access
 
-During processing of the records there is a global state `s` maintained, accessible from 
+During processing of the records there is a global state `s` maintained, accessible from
 code provided in yaml. In addition, rules can access the information fields (values) `v`
 of the record, as well as timestamp and the original line string as part of the record `ll`.
 
-Global state `s` can be used to enforce any cross keys rules, by updating the global state 
+Global state `s` can be used to enforce any cross keys rules, by updating the global state
 in `POST` (or `PRE`) of one `KEY` and using that information for `CHECK` of another `KEY`.
-For each config file, `s` starts as an empty dictionary, so in order to track global state 
-it would require adding an entry to `s`. 
+For each config file, `s` starts as an empty dictionary, so in order to track global state
+it would require adding an entry to `s`.
 
 Example:
 
@@ -152,7 +155,7 @@ Config files in the queue are processed independently, meaning that they do not
 
 Each config file may define it's `BEGIN` and `END` records, as well as any other `KEY` rules.
 
-Example: 
+Example:
 
     - KEY:
         NAME:  submission_benchmark
@@ -164,7 +167,7 @@ Example:
 #### Other operations
 
 `CODE`, `REQ`, and `POST` fields are executed using python's `exec` function. `CHECK` is performed
-using `eval` call. As such, any legal python code would be suitable for use. 
+using `eval` call. As such, any legal python code would be suitable for use.
 
 For instance, can define rules that would print out information as shown in the [example above](#global-and-local-state-access).
 

mlperf_logging/compliance_checker/mlp_compliance.py

@@ -147,6 +147,11 @@ def parse_alternatives(self, string):
         alternatives = in_pharentises.split(',')
         return [s.strip() for s in alternatives]
 
+    def parse_at_least(self, string):
+        n_string = string[len('AT_LEAST(') : -1]
+        n = int(n_string)
+        return n
+
     def configured_checks(self, loglines, config_file):
         with open(config_file) as f:
             checks = yaml.load(f, Loader=yaml.BaseLoader)
@@ -164,7 +169,7 @@ def configured_checks(self, loglines, config_file):
         begin_blocks = [x for x in checks if list(x)[0]=='BEGIN']
         assert(len(begin_blocks)<=1) # up to one begin block
         if len(begin_blocks)==1:
-            exec(begin_blocks[0]['BEGIN']['CODE'].strip(), state)
+            exec(begin_blocks[0]['BEGIN']['CODE'].strip(), state, locals())
 
         key_records = {}
         for k in checks:
@@ -231,6 +236,12 @@ def configured_checks(self, loglines, config_file):
                     self.put_message(f"Required AT_LEAST_ONE occurrence of '{k}' but found {len(reported_values[k])}",
                                      key=k)
 
+            if v['REQ'].startswith('AT_LEAST'):
+                n = self.parse_at_least(v['REQ'])
+                if len(reported_values[k])<n:
+                    self.put_message(f"Required AT_LEAST({n}) occurrence of '{k}' but found {len(reported_values[k])}",
+                                     key=k)
+
             if v['REQ'].startswith('AT_LEAST_ONE_OR'):
                 alternatives.add(tuple({k, *self.parse_alternatives(v['REQ'])}))
 

mlperf_logging/compliance_checker/mlp_parser/ruleset_310.py

@@ -8,17 +8,18 @@
 import json
 import re
 import sys
+from dataclasses import dataclass
 
 from io import open
 
-LogLine = collections.namedtuple('LogLine', [
-    'full_string', # the complete line as a string
-    'timestamp',   # seconds as a float, e.g. 1234.567
-    'key',         # the string key
-    'value',       # the parsed value associated with the tag, or None if no value
-    'lineno',      # the line number in the file
-    ])
-
+@dataclass
+class LogLine:
+    """Class for keeping track of an item in inventory."""
+    full_string: str
+    timestamp: float
+    key: str
+    value: str
+    lineno: int
 
 TOKEN = ':::MLLOG '
 

mlperf_logging/compliance_checker/training_3.1.0/closed_common.yaml

@@ -2,7 +2,7 @@
 - KEY:
     NAME:  submission_benchmark
     REQ:   EXACTLY_ONE
-    CHECK: " v['value'] in ['resnet', 'ssd', 'maskrcnn', 'gpt3', 'dlrm_dcnv2', 'bert', 'rnnt', 'unet3d'] "
+    CHECK: " v['value'] in ['resnet', 'ssd', 'stable_diffusion', 'maskrcnn', 'gpt3', 'dlrm_dcnv2', 'bert', 'rnnt', 'unet3d'] "
     POST:  " enqueue_config('training_3.1.0/closed_{}.yaml'.format(v['value'])) "
 
 - KEY:

mlperf_logging/compliance_checker/training_3.1.0/closed_stable_diffusion.yaml

@@ -0,0 +1,75 @@
+# Stable diffusion uses two metrics, FID and CLIP.
+# These metrics can be calculated offline, using different scripts
+# and logged seperatly. Therefore, we create a virtual key
+# called aggregated_eval_accuracy, which aggregates
+# both metrics into a single log line
+
+- BEGIN:
+    CODE: |
+        from dataclasses import replace
+        agg_eval_lines = {}
+        for line in loglines:
+            if line.key == "eval_accuracy":
+                step_num = line.value['metadata']['step_num']
+                if step_num not in agg_eval_lines:
+                    new_line = replace(line) # Make a copy
+                    new_line.key = "aggregated_eval_accuracy"
+                    new_line.full_string = "" # Not needed
+                    new_line.lineno = -1      # Not needed
+                    new_line.value = {'value': {'step_num': step_num}, 'metadata':{}}
+                    agg_eval_lines[step_num] = new_line
+
+                agg_eval_lines[step_num].timestamp = max(line.timestamp, agg_eval_lines[step_num].timestamp)
+                agg_eval_lines[step_num].value['value'][line.value['metadata']['metric']] = line.value['value']
+        loglines.extend(agg_eval_lines.values())
+
+
+- KEY:
+    NAME:  global_batch_size
+    REQ:   AT_LEAST_ONE
+    CHECK: " v['value'] >= 0 "
+
+- KEY:
+    NAME:  opt_name
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 'adamw' "
+
+- KEY:
+    NAME:  opt_adamw_beta_1
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0.9 "
+
+- KEY:
+    NAME:  opt_adamw_beta_2
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0.999 "
+
+- KEY:
+    NAME:  opt_adamw_epsilon
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 1e-08 "
+
+- KEY:
+    NAME:  opt_adamw_weight_decay
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] == 0.01 "
+
+- KEY:
+    NAME:  opt_base_learning_rate
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] >= 0.0 "
+
+- KEY:
+    NAME:  opt_learning_rate_warmup_steps
+    REQ:   EXACTLY_ONE
+    CHECK: " v['value'] >= 0 "
+
+- KEY:
+    NAME: aggregated_eval_accuracy
+    REQ: AT_LEAST(2)
+    CHECK:
+        - "'FID' in v['value']"
+        - "'CLIP' in v['value']"
+        - "'step_num' in v['value']"
+    ATLEAST_ONE_CHECK: "(0.0 <= v['value']['FID'] <= 90.0) and (0.15 <= v['value']['CLIP'] <= 1.0)"
+

mlperf_logging/compliance_checker/training_3.1.0/open_common.yaml

@@ -2,6 +2,6 @@
 - KEY:
     NAME:  submission_benchmark
     REQ:   EXACTLY_ONE
-    CHECK: " v['value'] in ['resnet', 'ssd', 'maskrcnn', 'gpt3', 'dlrm_dcnv2', 'bert', 'rnnt', 'unet3d'] "
+    CHECK: " v['value'] in ['resnet', 'ssd', 'stable_diffusion', 'maskrcnn', 'gpt3', 'dlrm_dcnv2', 'bert', 'rnnt', 'unet3d'] "
     POST:  " enqueue_config('training_3.1.0/open_{}.yaml'.format(v['value'])) "
 

mlperf_logging/compliance_checker/training_3.1.0/open_stable_diffusion.yaml

@@ -0,0 +1,34 @@
+# Stable diffusion uses two metrics, FID and CLIP.
+# These metrics can be calculated offline, using different scripts
+# and logged seperatly. Therefore, we create a virtual key
+# called aggregated_eval_accuracy, which aggregates
+# both metrics into a single log line
+
+- BEGIN:
+    CODE: |
+        from dataclasses import replace
+        agg_eval_lines = {}
+        for line in loglines:
+            if line.key == "eval_accuracy":
+                step_num = line.value['metadata']['step_num']
+                if step_num not in agg_eval_lines:
+                    new_line = replace(line) # Make a copy
+                    new_line.key = "aggregated_eval_accuracy"
+                    new_line.full_string = "" # Not needed
+                    new_line.lineno = -1      # Not needed
+                    new_line.value = {'value': {'step_num': step_num}, 'metadata':{}}
+                    agg_eval_lines[step_num] = new_line
+
+                agg_eval_lines[step_num].timestamp = max(line.timestamp, agg_eval_lines[step_num].timestamp)
+                agg_eval_lines[step_num].value['value'][line.value['metadata']['metric']] = line.value['value']
+        loglines.extend(agg_eval_lines.values())
+
+- KEY:
+    NAME: aggregated_eval_accuracy
+    REQ: AT_LEAST(2)
+    CHECK:
+        - "'FID' in v['value']"
+        - "'CLIP' in v['value']"
+        - "'step_num' in v['value']"
+    ATLEAST_ONE_CHECK: "v['value']['FID'] >= 0.0 and v['value']['CLIP'] <= 1.0"
+

mlperf_logging/rcp_checker/rcp_checker.py

@@ -3,6 +3,7 @@
 '''
 
 import argparse
+from collections import defaultdict
 import glob
 import json
 import logging
@@ -26,6 +27,7 @@
         'ssd' : 5,
         'unet3d' : 40,
         'rnnt': 10,
+        'stable_diffusion': 10,
     },
     "hpc": {
         'cosmoflow': 10,
@@ -48,6 +50,10 @@ def read_submission_file(result_file, use_train_samples):
     bs = -1
     benchmark = None
 
+    # FID and CLIP metrics for stable diffusion are logged asynchronously
+    # and indepently from each others. We track the eval results
+    # so we can get the first eval step that passes the convergence criteria
+    stable_diffusion_eval_results = defaultdict(dict)
     with open(result_file, 'r', encoding='latin-1') as f:
         # TODO: use mlperf_logging.compliance_checker.mlp_parser instead
         file_contents = f.readlines()
@@ -63,23 +69,39 @@ def read_submission_file(result_file, use_train_samples):
                     benchmark = json.loads(str)["value"]
                     if benchmark != "bert" and use_train_samples:
                         use_train_samples = False
-                if not use_train_samples and ("eval_error" in str or "eval_accuracy" in str):
+
+                if benchmark == "stable_diffusion" and ("eval_error" in str or "eval_accuracy" in str):
+                    eval_accuracy_str = str
+                    eval_step = json.loads(eval_accuracy_str)["metadata"]["step_num"]
+                    eval_metric = json.loads(eval_accuracy_str)["metadata"]["metric"]
+                    eval_score = json.loads(eval_accuracy_str)["value"]
+                    stable_diffusion_eval_results[eval_step][eval_metric] = eval_score
+                elif not use_train_samples and ("eval_error" in str or "eval_accuracy" in str):
                     eval_accuracy_str = str
                     conv_epoch = json.loads(eval_accuracy_str)["metadata"]["epoch_num"]
                     conv_epoch = round(conv_epoch, 3)
-                if use_train_samples and "train_samples" in str:
+                elif use_train_samples and "train_samples" in str:
                     eval_accuracy_str = str
                     conv_epoch = json.loads(eval_accuracy_str)["value"]
+
                 if "run_stop" in str:
-                    # Epochs to converge is the the last epochs value on
-                    # eval_accuracy line before run_stop
                     conv_result = json.loads(str)["metadata"]["status"]
                     if conv_result == "success":
-                        subm_epochs = conv_epoch
                         not_converged = 0
+                        # Epochs to converge is the the last epochs value on
+                        # eval_accuracy line before run_stop. Except for Stable Diffusion
+                        # where we use the first eval step that passes the convergence criteria
+                        if benchmark == "stable_diffusion":
+                            passing_epochs = []
+                            for eval_step, eval_result in stable_diffusion_eval_results.items():
+                                # TODO: we shouldn't hardcode the convergence criteria here !
+                                if eval_result["FID"] <= 90.0 and eval_result["CLIP"] >= 0.15:
+                                    passing_epochs.append(eval_step)
+                            conv_epoch = min(passing_epochs)
+                        subm_epochs = conv_epoch
                     else:
-                        subm_epochs = 1e9
                         not_converged = 1
+                        subm_epochs = 1e9
 
     if not_converged:
         logging.warning(' Run incomplete or did not converge. Marking as infinite.')